We started our journey nearly 20 years ago, coming together as a team of veterinary software developers with a clear vision to create a PMS that can support a modern, dynamic veterinary practise. Our cloud-based practise management software offers revolutionary features to help deliver excellent and with customer service, appointments, billing, and more. The system allows for easy setup and customization to track patient records, scheduling, billing and inventory for practises that ranging from small single vet practises to large clinics.
The beauty of the vet ID system is that it is moulded and designed for each individual practise, and we make sure that you know how to get the most out of your investment. As part of NBS Group, we are working on integration solutions that will add value to your practise and help you stay competitive in the constantly evolving veterinary market. Hello, and thank you for joining my talks on unleashing AI and radiology.
These are what are called bite-size talks. These are 10 to 20-minute talks on AI technology and how to integrate AI technology into your practise. The goal of these talks is really to provide some guidelines or foundation for people who are looking at AI products.
And how to evaluate those products. My name is Seth Wallach. I'm a veterinary radiologist.
I'm the CEO and founder of Vitology and Vitology Innovations. Vitology is a teleradiy service, and Vitology AI or Vitology Innovations is the AI arm of that service. Photology is partnering with VetIT to bring our product to the UK, the UK market, and we're hoping that you'll try out the Photology product and provide us some feedback.
So, as I mentioned, these are 10 to 20-minute talks. There are 5 talks. The titles of the talks are, the first talk is Time Savings with Radiy AI.
The 2nd talk is how to conceptualise AI when you have a biological background rather than an IT background. The 3rd talk is evaluating how to evaluate an AI product, what criteria do you use. The 4th and 5th talks are about AI and evaluating thoracic and abdominal canine images, what to look for, where the AI may do really well, where the AI may actually incorrectly interpret things.
So let's jump right into this time savings with Ray AI because ultimately, bringing AI, any AI product in your practise should be saving you time. Again, as I mentioned, the founder and CEO of Vitology Innovations, Vitology does follow GDPR and I recommend that no matter what product you look at, they comply with GDPR regulations. When you're evaluating a radiology AI product, there are really 4 places that AI will save time, and that's obtaining radiographs, sending radiographs, evaluating radiographs, and reporting on radiographs, and we'll talk about each one of these separately.
So the first one is obtaining radiographs, and AI is a great tool to actually be able to save time and help your technicians with obtaining radiographs. Proper radiographs are a key step in an accurate AI interpretation, and AI is able to actually evaluate radiographs for positioning and technique. And early on we realised that hospitals, veterinarians, clinicians.
They were looking for a company to actually handle the backend oversight on those images. So we found that assistance with the actual radiographic technique and helping their techs out was absolutely massive, and it saved them a lot of time during the day. So that the veterinarians no longer had to be going back and checking in on radiographic technique, but instead we're actually doing the work and we were handling the back end, looking.
At images, making sure that there weren't any issues or talking to the techs about doing rechecks on the images that were incorrectly positioned or incorrectly exposed. So with photoology, we make sure there's an image quality feedback that comes through with the AI. We also have people who are on the back end that if there's a question, they can always reach out to Vitology and ask a question about how to position a certain radiograph, how to, what technique to use.
Antech, I, I, spoke with them and they also offer image quality feedback, which I, again, really recommend that there's a, that whatever product you're looking at has image quality feedback. So I applaud Antech for also seeing the importance of image quality feedback. The other companies, I hear back from them about this, but when you talk with them, make sure that they have that type of image quality feedback and assistance on the back end so that you can actually save time and not have to go into the actual radiograph room and approved radiographs.
Let the back end service handle that for you. Image uploading. This one simply automatic image sending is what you want to use with one of these AI services because as I said, AI is about saving time.
And if you have to go and look at the radiographs and then decide that you want to submit those radiographs to have an AI evaluation, you're already doing extra steps. So with an AI product you want to have take the have the techs take the images, have the text. You know, if they have questions, talk to the AI company about the images themselves.
Then when the study is done, you want that study to automatically be sent to the cloud and then have a report back so that when you're ready to look at the images, you have a report ready. The report or the, the measurements on the image really should be there about 5 to 10 minutes after the study is complete. And so when you're actually ready to look at the radiographs, you should have all your AI information right there.
Any other way that you do that is not saving you time. So again, automatic image sending, I think is a must with whatever service you're using. Human in the loop.
So, with the AI evaluation, and with the AI evaluation only, when you get a report back or you get a a measurement on the image back, you are that human in the loop. You're using that AI report to help you evaluate the radiographs, but you ultimately are looking at those radiographs, making some decisions, and then using AI to confirm what you're seeing. If you're unsure of what's on the radiographs.
Don't use the AI as much to confirm or to, to work for what the AI says because the AI can get things wrong. That's where you wanna actually have a service that also offers teleconsulting, telleradiologists, so that if you're unsure about what's going on in the images. Then you want to send those in to the specialists, the experts, the radiologists, so that they can actually provide that for you.
So instead of you actually looking at the report and then going to find a colleague to ask about the radiographs or even going to a book to look up something, if you're unsure what's going on in the radiographs and the AI report doesn't actually confirm what you think is going on there. Do the, the time saving step and send those off to have them reviewed, and then you can go see clinic cases, you can go see cases, work on clients, or work on diagnostics in the back end. The companies that actually offer both an AI interpretation and a radiologist overread are Antec, Radimal, Signalpeet, and Vitology.
So, again, to sum up, when you're looking at a product, when you're looking at an AI product, you really want to look at time-saving steps across all four areas. Is that company, is that AI product helping your technical staff obtain radiographs? Automatic image sending, you want to make sure it has that.
Evaluating the images, you wanna make sure that that product is going to have a result on those images, whether that's a written report or measurements on the images when you're ready to look at the images and so if you're ready to look at the images within 5 to 10 minutes after the study is done, make sure that that product is able to deliver those results within 5 to 10 minutes after the study is done, which again part of that is automatic image sending. And then the last part is on the reporting. Use the time that you have to look at the images, look at the report, and then if you're unsure about what's going on on the radiographs, send that in to the radiologist, and then you can do the other things while waiting for the radiologist's report to come back to the office.
All right, so, Let's talk about real quick. The companies That are offering AI So there's Antec, that offers rapidweed, and that's a small animal product. Equacyte is a lot of equine product.
Pecoxia raimal and Signalpet are small animal products. Excalibur from SK Telecom, that is a measurement tool, small animal product. Metron Mind is a measurement tool for a small and large animal, and Vitology, we're a small animal product and we're a pro style report.
And a prosty report is a radiologist-like report. So, if you have any questions about any of these AI products or questions about radiology or, or what, what these different products offer, you can always reach out to us at [email protected].
You can check out our website at vitology.net, online, the ACVR, the American College of Veterinary Radiology. We have some, some resources.
The American College of Radiology, which is the human radiology group, has a lot of resources on AI. And in the European, the European College of Veterinary Diagnostic Imaging and the European radiology websites have information on AI and the Royal College of Veterinary Surgeons has an AI roundtable report, which is more of a, a general AI discussion, but it does have some information on AI. So What should AI do for you?
Efficiency. It should streamline your workflow. If you're not getting efficiency and streamline workflow with an AI product, it's not worth paying for.
That is the most important thing about AI. It needs to improve your workflow and save you time, so you have more time with your two and four-legged family members because happy pets equal a happy life. My second talk is how to conceptualise AI when you have a biology training or medical background.
Again, I'm the founder and CEO of Vitology Innovations. Vitology does follow GDPR, but make sure that whatever product you're evaluating also follows GDPR regulations. So when you're talking to people or you when you're listening to talks about AI they may be talking about database and prompts.
Well, how do we actually think about database and and prompts in the world of biology with our medical training? Well, database and prompts are really like DNA and RNA primers. The DNA or the database stores the information.
The prompts, and I don't know how many of you have used AI before, but you can put prompts into a system and it will actually return a result. The prompt is like an RNA primer that you would type in certain things. It would go to the DNA that, that, that's your RNA primer.
It would go to the DNA and it would. Pull back a string of information and you can take each one of those strings of information that it brings back and put those together to build what you're looking for or what you ask the system to build so whatever the system builds is dependent on what's in the database or what's in the DNA so. Let me, I have an image here on the next one, next slide.
The DNA, then the RNA primer, and so that RNA primer again is the prompt, and then it will actually, the RNA will build a certain amount reading off of the DNA to a specific length and then that will actually be brought back and shown to you as text or an image. So that's what's going on between prompts and the database. It, it's as simple as that.
The DNA is storing the information, and the prompts are actually then going back and retrieving information out of the system. And then the magic happens where the, they're collating all that information together and creating what you're asking for in the prompts. So let's take an example of that.
Let's say that you want to have a picture of my dog Barney, which is very nice to actually volunteer to be in this talk. But let's say you wanted to have a picture of a standing yellow lab who is looking at a camera. What will happen is with those prompts, it'll go to the to the to the database and it will look for with those prompts as the primers, it will look for specific groups of numbers, which is the translation to the system of a standing dog in a field and also a dog looking in the camera.
So the first part of that is the standing dog is the yellow lab are the 1st 4 sets of numbers. That's what the, how the system actually understands what a The numerical values associated with a, a standing yellow lab. The last two numbers, the 0.78 and 0.45, are a dog looking at a camera.
And so, what I'll end up doing is it'll bring, bring to the back from the system, the 1st 4 sets of numbers, which is the standing dog in a field, and then it'll add in the part about the dog looking at the camera and then it will correlate that into that image that you've requested. So again, it's the prompts of the primers, the DNA is the database, it goes and finds specific things, it puts those together, and that's the output that you receive. So, how do we get those numbers?
Where are those numbers coming from? Because how does a number equate to some certain type of finding? Well, that's called classifiers.
And what classifiers are, and this is the beauty of what Hinton's group did, they ended up saying, OK, if you take a bunch of positives or something that you say is positive for something, and something that is negative for something and put them into the system. This system will be able to tell which things are positive and which things are negative. They'll decide what the difference is as to why you would call something positive and something negative.
And from there, it stores that information so that any future thing that you put in and ask it to evaluate, it will tell you whether it's more likely, more similar to the positives. Or the negatives that you put in there. And what it's doing with the numbers is it's actually giving you a confidence score.
0.5 means that it doesn't really know whether it's closer to the positives, or closer to the negatives. As you get closer to 1, it means that it's more, more and more confident that it's very similar to all the positive examples you put in for whatever the examples were that you put in.
And in bronchial pattern, that would be a, a grouping of. Radiographs with bronchial disease that were put in as positive, and the negatives will be radiographs without bronchial disease. So, in this case, the bronchial pattern classifier came up above 0.5 at 00.65, saying that, yeah, I'm a little more confident that this image of this bronchial pattern is more like the cases, examples of positives than it is like the negatives, the cases without bronchial pattern.
And we do that for each of these different findings. So for the interstitial pattern. The 0.54 means the system's not really sure if there is or isn't an interstitial pattern.
Now that can be because they're truly, it's not an example that it's seen before. That would be the most likely scenario where something's sitting at 0.5 is that.
The system doesn't really, can't really put it with the positives. It can't really put it with the negatives, so it gives you that result that's right in the middle of, 0.5 or 00.54.
But that doesn't mean that result is actually not worthwhile because it's part of that fingerprint. Then left sided cardiomegaly is as you can see on this image, there's a little bit of a left atrial bulge, and so the system said, yeah, there's a little bit, this matches more with the positives than the negatives, but it's not like a, a huge example or a hugely big example of left atrial enlargement. So I'm more confident that it is a positive case than it is a negative case.
And so the left sided cardiomegaly comes up at 0.63. And the right-sided cardiomegaly, the system is saying, I am very confident that this image matches all the examples of right-sided heart enlargement that I was trained with, and therefore I feel very confident that this is a case that is positive for that right-sided cardiomegaly.
Now the right side. Cardiomegaly, each of these numerics or the names on these classifiers are what we assign to it. The system doesn't know.
It's just saying whether the image has the characteristics that are more like the positive examples or negative examples that were fed to it for a specific thing. And the last one is pleural fluid in the system here is saying that it's, it doesn't, it matches more with the cases that didn't, that were negative for pleural fluid than the ones that were positive for pleural fluid, so the number is actually closer to 0. With that result of those numbers and that kind of fingerprint of numbers, all, you know, put together, just like we showed in the example of Barney.
It the system then takes that fingerprint and creates the wording to go with it, which is cardiomegaly predominantly right sided it's with a bronchial pattern and a minimal interstitial pattern but no pleural effusion, and that is kind of how a fingerprint or all those results is translated into an AI report. The conclusions on that again from the database because we would have both the results and the actual reports or what was said with each finding and the conclusion, the conclusion on that will be right sided cardiomegaly, no evidence of failure, and rule out core pulmonale because of the bronchial pattern. So we, we do that with all the different examples that we have.
We build classifiers. We create those classifiers of positives and negatives. We put the images through and then we look at those results and look at the whole grouping of results, and that tells the story of what is in the actual radiograph.
So, after I ended up listening to this talk, I noticed, something that I'd said incorrectly, and I wanna just clarify. On this, on this slide, I was talking about the interstitial pattern. The 0.54 indicates that it's a minimal positive.
0.5 result, as I had mentioned, is either the amount of positive and negative, or the amount of the image that matches the positive examples and the image. And matches the negative examples is right at 50%.
More likely, a 0.5 result is when the image actually is something the system hasn't seen. A 0.54 result indicates that the images slightly, it matches the positive images or the positive, training images more than the negative training images.
So that's why the findings state that there's a minimal interstitial pattern. I hope that clears that up. Thank you.
The other thing about AI that's really important to understand and it's where I think the future of AI is going, is it can be personalised. So a system will have a fingerprint of results, and that fingerprint of results can be interpreted or translated into different things for the way a radiologist would say things. So, radiologist number one.
Maybe talking about the right sided cardiomegaly and bronchial pattern a certain way versus radiologist number 2 may talk about it in a different way, but both of those are being generated off of those initial AI results. The initial AI results don't change. It's the, interpretation that actually can be adjusted for and personalised for each radiologist, so.
In the, in an example of that is, again, where I think that AI is going to get to, which is personalised generative AI for everybody, radiologist number one would say right-sided cardiomegaly, no evidence of failure, rule out core pulmonale. But radiologist number #2 may say it a little bit differently, moderate to severe right-sided cardiomegaly, with, without pleural effusion, concurrent bronchitis. But all of those are generated off of that original, fingerprint of results.
So why is AI so powerful and what, what can it do? Well, anything that can actually be translated into numerical values. And stored into the database can then be used for AI.
So whether that's Medical records, whether that's radiographs, whether that's, you know, cytology, ECGs, tracings, stethoscope, sounds, all of those can actually be converted into numerical values. A fingerprint of all of those values will correlate to a specific thing, and then if you put in a prompt or if you have an example of one of those things and put it into the system, it will go into the system and pull back the information as to what that is most like. And that's why AI is being kind of developed into everything.
It's not just in medical, it's in everything, cause everything that exists can pretty much be turned into a numerical string of characters. So we're seeing advances in AI and radiology, AI and cytology and histopathology, ECGs, stethoscopes, blood work interpretation. AI scribe medical records are very big right now because they save people time, and that's what AI needs to, is supposed to do.
And you're gonna keep seeing all of these different things with AI because they're taking examples of things, putting it into the database with and, with numerical values, and then being able to then query that database and bring back a result with a new example. And then the other part of that, as I mentioned, is personalization, is that those numerical results are the numerical results. But you can actually translate those numerical results into something that's personalised for each person as to how they like to describe things, or write about things, or interpret things, so that our, our personas aren't lost by using AI.
So in the case of an AI scribe, what happens is they take medical records, and they feed them into the system, and the system determines which words are associated with other words. And so for a case like vomiting, The other words that would be associated with it might be whether there's, what the abdomen feels like. Is the abdomen tense?
Is the abdomen soft, or is there salivary staining around the muzzle. And it learns the, the association between these, so that when somebody's in a room, having a discussion with an owner about a pet and doing a physical exam. The system will key in on the vomiting and we'll be looking for those specific things, and when those specific things, it finds those, it will actually bring them back and put them into the medical record.
And that's how the medical records of these AI scribes are building the medical records and including the important stuff because the AI system on the back end actually kind of figured out which things are, are related based on how often they show up in there all the examples on the back end. So, that's why, why you'll, all these AI scribes are using backend data and being able to generate your records, which is, I think also, besides radiology, I think the AI scribes are really changing the veterinary market and saving people, saving veterinarians, veterinary surgeons, a tremendous amount of time. So again, if you have any questions, you can always reach out to us at [email protected].
You can check out our website at vitology.net or if you're interested in AI products, there's a guide to AI tools for veterinary medicine, and they update that every year. It has a list of all the different products that are using AI, and I expect that list will continue to grow, and it will, it talks about both radiology products and non-radiy products, but.
As you can tell that anything that is able to be put in the database and enough examples of it exist, we will end up using AI to actually help improve in our workflow and improve our workday. So, To recap on this. AI needs to be about efficiency.
It needs to streamline your workflow, whether that's a radiology product or a subscribed product. If the AI is not saving you time, it's not worth paying for. You have, you need to get the extra time to spend time with your family, because happy pets equal a happy life.
The next talk is on the evaluation of AI products and what to look for. Now this is kind of a, a more personal thing because the, we have, you know, the common things that people are looking for, but there may be something unique to your clinical setting that you're gonna wanna understand before you actually start engaging with the AI vendors in radiology so that they can meet those those needs, right? So again, I am the founder and CEO of Vitology Innovations.
Vitology does follow GDPR and whatever product that you actually end up, utilising, make sure that they also utilise, or follow GDPR. So, in the evaluation of a product, One of the biggest things that, time and again that I tell people is always, always, always base your decision on current features of a system and not what the vendor promises, because sometimes what the vendor promises. They get built and sometimes what they promised doesn't get built.
So whatever you're looking for, the current features need to meet your needs or need to meet the most number of your needs in those current features. If the future things that they say they're going to build actually come to fruition, great. But based it on the current features.
And don't count on the future features because I've seen it time and again where things are promised on the features and they don't get built. All right, the other thing before, Before you start evaluating these different products, is sit down and think about what your 3 or 4 top radiology needs are and what the product differentiators are. You want to look at the market and see what the product differentiators are and what your clinic needs that's gonna improve your workflow.
The 3 or 4 things, the 4 things that really, we're asked about or that are very important to veterinarians who, talk to us about our product is how accurate is the product? And is there peer-reviewed publications on the AI performance and validation. Time savings, that's the whole thing about AI.
It should be saving you time, and if it's not saving you time, it's not worth paying for. What kind of tools do you want AI to provide? Do you want it to do a report?
Do you want it to look at the images and provide a pro-style radiologist report, or do you want it to look at the images and provide measurements? And then cost structure. The two typical cost structures are a subscription model or send as needed.
And going back to that first lecture, which is, I always recommend an automatic upload, so that you're not looking for the report or having to spend time to actually submit the images. I personally think a a subscription model is a better model. Because a subscription model allows you to send everything into the platform and so that report will be available at the time when you're looking at the images.
So which radiology groups or products have publications? Antec, Poxia, Radol, Signalpeet, and Vitology all have publications that you can find online and read about their products, and you're really looking for peer-reviewed publications, things that a third-party evaluated in a clinical setting. So for vitology, our first product, our first publication was done with the AMC in New York, and that was on left heart failure in dogs.
And that was, our results were fantastic in terms of, left heart failure identification. It was a sensitivity of about 93% and, and a specificity of 96%. And so it was a highly, highly accurate, study result.
They used about 500 cases in that study. And the reason it was so accurate is because we essentially were the first to take that aggregate of information of all those results and build the fingerprint to actually say what the disease was or what was going on. Before that, it was individual findings.
And so that. Fingerprint of information is why we got such really good results on the pulmonary edoema. When we actually look at these other studies, these are all individual classifier results.
And individual classifier results, in my experience, do not perform as well as a fingerprint of AI results altogether, telling the whole story. So, The, we did a study with Tufts on pleural effusion, and that was, in the 80% sensitivity specificity. We did another study with, Tufts on pulmonary nodules.
We did a study with Purdue on eurothelial neoplasia, identification of urothelial and neoplasia in dogs, and what's interesting on that is the VD projection was actually more accurate than the lateral projection for looking for urothelial neoplasia, and I'm still not sure why because the spine. Overlies the majority of the findings of the bladder and looking for any type of mineralization or any types of changes, but the AI system, when all the images were put through, it found something in those positives that wasn't in the negatives. That is why it put, you know, whatever it saw in that result, that's why it put everything into the positives and.
There's some difference between a positive and a negative VD image for animals with urothelial neoplasia. That the system picked up, but that, you know, as radiologists we really don't see and that's the amazing thing about AI is during that evaluation it can find things and the relationships between things that are previously unknown. Even to specialists.
And then the last study that we did was with Oregon State University, and that's on right-sided cardiomegaly. And in that instance, we actually compared that to echo data rather than the, as the gold standard rather than thoracic radiographs. And I'll, those results are all published, and so you can find those online.
As I mentioned before, having technical assistance on the back end by whichever AI company, radiology AI company you use is, it's so important. It's, it's a massive improvement of efficiency for clinic workflow and for getting your technicians the assistance that they need for obtaining properly positioned, properly exposed radiographs. Because that is the foundation of quality AI results.
And so at Vitology, we made it very important that we provide image quality feedback as well as human and AI assistance for technicians as needed. Make sure that whatever company you're, you're looking at provides that back end. And as I said before, automatic image upload.
I think it's a must. I think you wanna have a cost structure that actually automatically sends the images because then the reports automatically get generated. And the report is available to you at the time when you want to look at the radiographs.
Otherwise you'll find yourself looking at the images and then you'll have to do a manual send. You'll have to go and do something else for a little bit, and then you come back and look at the images when the report arrives. It's just extra steps, and that's not what AI is about.
AI is about improving workflow, not creating a disjointed workflow. So who provides what in terms of reporting? So the pro-style reports are provided by Antec Signal Patent Vitology.
Raimal provides a report, but it has colour sliders on it. The radiology tools where the measurements are done on the images is by MetronMind, Pecoxia, and Excalibur. And I'm gonna show you some examples on the next few slides.
So these are examples of the pro-style report. The far left is the vitology report. We put our conclusions at the top, and we have a vertebral hard score for the thoracic images.
And then everything is in, it's written out as a radiologist would do, but this is done by AI and this is using our backend database to build those reports. AIS, they do the same thing in terms of pro-style report, so does Signal Pet, and as I mentioned on Radimal, which is on the far right. They have more of these colour sliders that talk about each of the classifier, and whether it's positive or negative.
So, it's not creating the synthesis on the report, it's just putting the information there, and for you to look at which things the AI found positive, which things the AI found negative, and for you to synthesise what that all means. In terms of the image measuring tools that are done by AI on the far left is MetronMind, and that is doing all the measurements and providing all the measurement numbers and providing some statistical analysis on that. The top right is Pcoxia, which is AI doing a vertebral hard score and providing that.
And then at the bottom is Excalibur. It's a little more colourful. Also, an AI tool that does automatic measurements and provides a vertebral hard score, among other things, but that is pretty much doing AI measurements on the images, and then you, you would talk to the owner based on having the, the measurements versus Having a report and talking to the owner about the report.
And this is a personal preference. Do you wanna actually talk to them about a report? Do you wanna talk to them about measurements?
So in terms of what the different providers do in terms of the species, all, almost all of the companies provide small animal. Equisite is the only exception to that. They do an equine product only.
The rest of the companies do a small animal product. Metron Mine does small animal measurements and equine measurements on the radiographs. That, and that's all AI generated information.
The other part, as I mentioned before about AI is that AI is able to do personalization, and this is something that early on we also feel very strongly about is that AI should not turn everything into a one size fits all type of approach. So we actually create personalised reports on the back end for our radiologists using the AI. Fingerprint, but turning the results into how they would read an image and how they would interpret an image.
And so what that does is by providing that personalised prelim, the radiologist spends less time having to think about writing the report, more time focused on the images, and more time just kind of reviewing the report, making sure that it says what is on those images is correct. And what that translates into is faster turnaround times, so reports are getting delivered back to the veterinary clinics faster than they were before. So, the, the, the key point of that is when you're talking to these companies and about their AI and their AI systems, and you want to utilise their backend teleconsulting product, make sure that they're, they're using some type of personalization on the pre-lien for the reporting because that is gonna improve turnaround times.
The other thing is an automatic review versus a send-in. Now, this is a little bit of a difference of opinion on whether it's, it should be automatically flagged for review or not. I am of the belief that veterinarians looking at images, I don't, each veterinarian has a different level of experience and comfort level with radiographs.
So for, For certain vets, they may look at a case of a heart failure and they want confirmation on that, versus for another vet, they may look at a case of heart failure and say, oh yeah, that's heart failure, and they start treating, right? So, we've always believed in having it be request a review where you would actually decide if you wanted to send it in. The AI report would automatically be there, but from there, you would look at the images, look at your report, and decide if you want to send it in.
Antech does something which I think is a great service, but it's a great service for those that need, want that kind of service, which is they will. Flag certain cases, and particularly emergency cases for an overread by a consultant, so you'll receive, receive the AI result and automatically receive the, radiologist interpretation for certain emergency conditions. So you have to decide, do you want to have certain things automatically reviewed or not.
And that's one of those differentiators between the companies. So, Antech is the only one that provides that automatic radiologist overread. Antechradal Signal and vitology all provide the ability to send the images in for an evaluation by a consultant based on a a request, and so you'd have to actually select to send them in because there is a charge associated with that.
Then the cost structure, right? So, as I mentioned before, you want to have a manual, an automatic send. Excuse me, an automatic send rather than a manual send, because you want everything to actually get an AI interpretation.
But there are companies that will offer a per case charge for their business model. When if you're looking at one of those per case companies, find out whether or not you can automatically have everything sent to them. But This is kind of why I believe that the, the subscription model is a better model in that you wanna be able to just, for efficiency, you wanna be able to have the images automatically sent in and have that report back to you, so you can have that report when you're ready to look at the radiographs.
No matter which company you use, any type of radiologist overread will be an extra charge, whether there's a per case for the AI or a subscription model. If you end up sending those in, or if it's an automatic read that's flagged, such as with Ante, there will be an additional charge for that because essentially they're having a radiologist spend the time to look at all the images. And provide a report.
So again, if you have any questions about AI products for radiology or AI in general, you can always reach out to us at [email protected]. You can see our, check our website out.
It's vitology.net or you can look at the guide to AI tools for veterinary medicine, and that gets updated every year. I think they just released their 2025 list of AI products in in veterinary medicine.
All right, I, as I do with all these lectures, let's sum it up with, AI needs to be gaining you efficiency and streamlining your workflow, and if you're not seeing an improvement in efficiency and not seeing an improvement in your workflow, the AI product is not worth paying for. It needs to save you time. It needs to allow you to have more time with your family because happy pets equal a happy life.
All right. Well, let's, talk 4. Let's talk about radiology AI use cases for the canine thorax.
I'm Seth Wallach. I'm a veterinary radiologist, CEO, founder of Vitology Innovations. Vitology does follow GDPR, but whatever company you're looking at for AI, make sure they also follow GDPR.
Now, I'll say that some of the slides in this talk are single images, but studies should always be obtained with orthogonal radiographs. Because orthogonal radiographs provide better AI and human radiology results in terms of interpretation. So talking about radiographs, accurate AI results, accurate radiology interpretations always begin with proper radiographs.
Early on, we, we realised that as a company, and so we always provide back-end support to all of our customers, our subscribers, to provide radiology feedback on the images as well as human assistance. So, let's talk about why radiographic images or proper positioning technique are so important. And these are the things they talk about in school.
But as we talk about them, I'll kind of give you some examples, cause that will actually help kind of reinforce why positioning is so important, whether it's a, a radiologist or a human looking at the actual images, or AI. So, things that you look for on the images, things that our vitology angels look for on the images to assist your technical staff or technique, rotation. We want to make sure that thoracic radiographs are completely at inspiration, full inspiration.
You want to make sure that there's inclusion of all of the thoracic structures from cranial to caudal lung tips, as well as no overlying soft tissue structures that could obscure anatomy. So, On the image that I just showed you, there's a little bit of rotation of the cranial thorax. Well, that's important because left cranial lung lobe disease can be obscured by this rotation, and that's a place where pneumonia, a lot of times will set up in a dog.
So you want to make sure that that's a straight vertical so that you'll be able to see that area. The other thing is that you'll see here caudally, the caudal lung tips are included, which is, that's, we really need to actually include that area. And I'll show you on another image why that's so important.
The technique on these images is very nice. It's a well done technique, and there's no overlying structures. Let's look at this lateral projection now.
Now this lateral projection, the technique is fine, but the lungs are hyperinflated, and probably the most important thing is that the caudal lung tips are not included. Caudal lung tips are really important to include because that is the most common site for metastatic neoplasia. So there could be a nodule in that lung tip that's actually not on this image, Whether there are humans looking at that or AI is looking at that.
If it's not on the image, it's not going to be detected. So a better inflated image, and actually the full image is an example here. You can see from cranial to caudal, there's no overlying structures.
The technique is well done. Everything is included, and the toughest thing that is still a, just moving the slide down here, the toughest thing that techs always struggle with is rotation. And to look for rotation, a lot of people talk about looking at the rib heads, but essentially, I recommend looking at the costochondral junctions.
You want to see that the costochondral junctions are actually superimposed, and a lot of times what happens is when the animal's on their side, that sternum will actually rotate down towards the, the table. So if you put a radiolucent foam wedge under the sternum, you'll elevate that part of the sternum, and you'll get a more consistent lateral projection. So if you're having trouble with your techs being able to obtain that, try getting a foam wedge, 1520 degree wedge.
You can get a few different, wedges in there and see which one actually helps with, with, elevating the sternum. All right, so now let's talk about thoracic AI use cases, and AI really shines in thoracic emergencies. And one of the first things that we actually worked with AI to develop was a heart failure model and to be able to identify heart failure on radiographs and provide an AI report back to veterinarians within 5 minutes of obtaining the radiographs.
The other area that veterinarians really wanted to know about were alveolar patterns or pneumonias. Some of the other conditions that veterinarians struggle with are interstitial pattern, bronchial patterns, some cardiomegaly and some pleural fluid stuff. So we're gonna talk a little bit about those, and we're gonna finish up with owner discussion or how to approach the owner discussion with the AI results.
So as I mentioned, left heart failure, that was a big thing that we started with because we wanted to help veterinarians get results about animals that were in heart failure within 5 minutes, and the caseload was such that the reports, the turnaround times for stat radiographs were getting longer and longer. So after we developed our heart failure assessment, we actually did a study at AMC. It was with 500 dogs in their emergency department, and our sensitivity and specificity were excellent.
So we're really happy about that. We had a 93% sensitivity and a 96% specificity of detecting left heart failure in dogs in a clinical setting. The, The other part of that that actually was even so that the the results from that study we were very happy with, but that was at one facility.
When we received an email from a vet in Florida at a totally different location who was using the product. They said that they had a dog that came in that they were thinking about euthanizing, and they weren't sure what was going on, and they sent those radiographs off to vitology, and the result came back as left heart failure. And they started the dog on Lasix because the dog hadn't been responding to really any other medication.
And within days, the dog improved, and the veterinarian said that they were so thankful that the product was there. And that was really the most important part is that we're out there saving lives. And when we got that result, that actually was as good, if not better, than actually this study here that showed these great results because I, we knew then that we were actually making a difference for veterinarians in practise.
So what is the system looking at for left heart failure? What are radiologists looking at for left heart failure? Because if the study comes back as saying that there's a high likelihood of left heart failure, it's going to be one of two things.
It's going to either be that there's truly is left heart failure, or there's severe cardiomegaly and a severe pulmonary infiltrate, but not due to heart failure. But more commonly, as you saw than not, it's going to be due to left heart failure. So, When the system's looking at the radiographs or the radiologist is looking at the radiographs, what it's looking for is cardiomegaly, particularly left sided.
This is the left atrium, and it's evaluating this left atrial margin here to see if that's actually well defined or if that's poorly visualised. That becomes poorly visualised when there's a perihilar interstitial pattern. And so in this case you can see that this left atrial margin is actually obscured in that area.
The other thing we look at is the vessels, whether the vessels are distended and whether or not we can actually see the vessel margins very clearly if there's an overlying perihilar edoema or interstitial edoema, the vessel margins become fuzzy. So this is the area the perihilar infiltrate on the lateral projection, and on the VD projection, the perihilar infiltrate will reside in both of these caudal lung lobes. If you're worried about right heart failure, you'd be looking for pleural effusion, which we'll talk about in a little bit.
On the VED projection, again, it's looking in that inner perihilar region, but whenever I have a case of cardiomegaly, I always look to see if there's a change in opacity going from cranial to caudal. And in this case, what you'll see is that cranially, There's actually, it's less opaque, and then caudally, as we get to this right caudal lung lobe, it becomes more opaque. So when I, when there's a case of cardiomegaly and a caudal lung lobe infiltrate, first thing you should be thinking about is left heart failure and cardiogenic pulmonary edoema.
How about in cases of acute heart failure? So if it's a chronic mitral valve with a cordary rupture, a lot of times there will be cardiomegaly. But a lot of the heart is obscured because of the severity of the heart failure.
I can tell you that cases of endocarditis, particularly cases of severe acute endocarditis, where there's acute heart failure, those can be very challenging studies because the heart hasn't had time to enlarge. So what the radiograph will just look like a complete whiteout, and when you're, and you don't see the tracheal elevation to indicate that there is any cardiomegaly. For those types of cases, it, it could be a lot of different things, and I recommend sending those off to have them reviewed, or if you have a, ultrasound probe available, take a look at that left atrial size.
So, if you have a complete whiteout of the chest, but a normal sized heart, no tracheal elevation, you want to start thinking about is this, an acute endocarditis, particularly if the animal has a fever. Other things that will actually create this appearance will be an animal under anaesthesia. So I never recommend taking radiographs if the animal is under full anaesthesia, general anaesthesia, because these types of radiographs is what you'll see.
People will send those in thinking. If there's a severe lung pathology, when the animal's actually fully recovered from anaesthesia, the lungs look fine. So, if you end up taking radiographs on an anaesthetized animal, know that a lot of times you'll see this type of severe, pulmonary pattern.
And so now talking about pulmonary patterns, and that example before that was just severe alveolar pattern. So, alveolar patterns come in two flavours. The, one flavour is without volume loss, and what we term consolidation.
And that has to do with the alveoli being filled with blood, pus, or water, sometimes neoplasia, so they retain their actual volume on the alveoli. With volume losses atelectasis, where the alveoli actually collapsed down, bringing the airways next to each other, creating parallel airways. How does metology do on the alveolar pattern?
Well, we look at it with or without volume loss, and the alveolar pattern detection that has a sensitivity of about 80%, same for the specificity, and as the alveolar pattern becomes more severe, the accurate detection of an alveolar pattern improves. But what are you looking for when we have an alveolar pattern and you, the report says that there's an alveolar pattern on the radiograph, and it's important that as a clinician that you differentiate between consolidation, which is where there's volume retention because something's in the alveoli, versus adalectasis, which has to do with just a collapse of an airway. So Or collapse of a lung and the airways being next to each other.
So on the top right is an example, on this, on the left is the example, and on the right side is a zoomed in on that area. And so this is an example of consolidation where the alveoli are filled with material. So as the bronchus comes down, you can see that the bronchus actually is, they're separated here, and that is because of fluid or some type of material in the alveoli between those airways.
In a dog, eventually distributed alveolar infiltrate is the most common differential for that is a pneumonia. And the bottom right image actually is a case of adalectasis where you can see that the airways are actually right next to each other, and that's because there's volume loss in the alveoli, which allows the airways to sit there right next to each other. So that is more typical of where there's a bronchial plug or chronic lung disease that's resulted in collapse of that lung.
So there's nothing in the alveoli on the image on the bottom. On the top, there's going to be most likely purulent material, but blood or even edoema, you know, depending on location, there's material inside the alveoli that are keeping the bronchi apart. The other way we think about airway collapse or atelectasis is the weeping willow tree.
I always, And so this is an example of the weeping willow tree where all of the branches hang down and they're next to each other and so that's very similar to how we think about the airways next to each other when there's actual volume loss in the alveoli or lung collapse. How about on the VD projection? Well, I know this is a talk about dogs, but I included a cat thorax because cats commonly have lung lobe collapse, so they're great examples of lung lobe collapse because of feline asthma and bronchial plugging.
So one of the left-hand side image is a cat that has consolidation, areas where there's volume retention, and we know this because the heart is staying centre in the middle of the thorax. Adelectasis will result in a shift of the heart to one side or the other. Now the difference between a dog and a cat.
With an alveolar infiltrate is, as I mentioned, dogs get pneumonia a lot of times. It's a very common thing for dogs to get a pneumonia. Cats don't get pneumonia as much, so whenever there's an alveolar infiltrate, which we know that this is, and consolidation because we know that the heart isn't shifted.
In a cat, pneumonia is one consideration, but it's a lot of times just as equally going to be not atypical cardiogenic edoema, neoplasia, haemorrhage, so all of those differentials are equally common in a cat versus in a dog, when you see eventually distributed alveolar pattern, you really think about a pneumonia. How does it look different with atelectasis? Well, in this example, you can see that the heart is shifted to the left side of the, the cat's hoax, and that's because there's a left lung collapse.
And in this example, there's a right lung collapse resulting in the cardiac shift to the right side. So whenever you see an alveolar infiltrate, you want to look for that mediastinal shift to help you determine is there consolidation or no volume loss, or is there aectasis, which is with volume loss. Other thoracic conditions that the AI interprets and, you know, radiologists or clinicians need to assess for are the interstitial pattern, the bronchial pattern, cardiomegaly, or pseudocardiomegaly.
And pleural effusion. The canine unstructured interstitial pattern, how does metology do on the interstitial pattern? Well, the sensitivity, as you can see, is a little above 50%.
And that's actually because there's a difference in opinion on the radiologists as to what classifies as an interstitial pattern. Some radiologists will call it interstitial pattern on most or nearly all cases. Other radiologists will be a little bit more limited as to what they call an interstitial pattern.
And I can tell you that interstitial patterns are probably the thing that plagued veterinarian. Varians the most. And the reason I think that happens is because of the variances in the radiologists as to what they determine to be interstitial pattern.
So you may receive a report from a radiologist that says there's an interstitial pattern and a similar appearing radiograph, send that off to another radiologist, and they'll say that's not an interstitial pattern. And that's where a lot of that confusion comes from. That, that's my opinion on that.
AI actually is very consistent with its interstitial patterns, so it actually helps people kind of reinforces what is, you know, an interstitial pattern based on a consensus of radiologists. When I'm looking at a radiograph and determining whether there is or is it an interstitial pattern. What I'm looking for is an increase in the number of lines in the thorax.
So there's normally lines associated with the vessels, there's some lines associated with the airways, but as you start to see, there's more lines out there and you can start to lose visibility of the vascular margins because of all the excessive lines, that's termed an interstitial pattern. Now, the interstitial pattern may be incidental, it may be due to hyperinflation, it may be just due to obesity. Or it may be more due to fibrosis or interstitial lung disease.
So it's a very non-specific pattern, but also can be challenging in terms of what exactly is, when do we call it interstitial pattern. So, the rule of thumb is if you see increased number of lines in there, that's essentially what we're talking about with an interstitial pattern. The other thing that can cause an interstitial pattern is rotation, and I've been fooled a lot of times when there's rotation because there's increased number of vascular markings you normally don't see, because a lot of times on a perfect lateral view, the vessels are superimposed or really right near.
So when you rotate an animal 30, 40 degrees, there can actually be a lot more appearing lines in there, resulting in an artifactual interstitial pattern. So again, proper positioning has a lot to do with interpretation of the radiographs. Pathology's performance on the bronchial pattern.
Our most recent results are a little perplexing to me, and I'm still waiting for the team to explain why we have this low sensitivity on there. Our previous results were around 70% for sensitivity and specificity. But if the system comes back as a bronchial pattern.
What are you looking for to confirm that there's a bronchial pattern? Well, on a lateral projection, what you're looking for is you're looking at these airways here to see if they're thickened. And if they're on the lateral and the bronchuss is heading in a cranial caudal direction, that's where you get those typical train tracks or tram lines.
And on bronchus, you'll actually see this donut, and in the middle of that donut is the lucinary, and that's the actual bronchus, and around that is the bronchial wall. So a little tip about bronchial patterns, at least how I interpret them. The bronchial pattern is pretty classic on its appearance, right?
You see that end on airway, you see the thickening of the wall around it, or the interstitium around it, and say, OK, well, there's a bronchial pattern. But the other thing you want to look for is whether or not there's hyperinflation, and in this case, you can see that there is not any hyperinflation. And hyperinflation is important in both dogs and cats because hyperinflation indicates that there is a bronchoconstriction or bronchial obstruction associated with the bronchial pattern, and when that happens, the lungs become hyperinflated because during inspiration, the lungs open, allowing air to get into the lung, and during expiration, the.
Pressure on the outside actually causes that, that bronchus to collapse down, reducing the diameter and causing air trapping. So when I see hyperinflation with a bronchial pattern, I'm then I'm concerned more about active airway disease and bronchial constriction. When I see a bronchial pattern without hyperinflation, it's more of a chronic or inactive process.
But either way, whether this is active or chronic, it's still that, what you're looking for is that bronchial pattern where you have the bronchus in the centre, and then the margin around there is thickened. How does vitology do on generalised cardiomegaly? Does pretty well, has a sensitivity in the mid 80s and a specificity in the low 70s.
So there's a couple of things that can actually cause the false reading of a cardiomegaly. So let's first talk about the true cardiomegaly, then we'll talk a little bit about what can cause a false cardiomegaly. So if you're looking at cardiomegaly, you're looking really for this left atrial margin here, and, again, you always want to assess this margin here to see if there's visualisation of that full margin or if it's obscured.
And if it's obscured, you start to be worried about a perihilar edoema. Then as we, follow this ventrally, you look at that left ventricular margin to see if there's rounding, and then looking for increased sternal contact, as well as rounding to that right heart. The other things you want to look for are elevation of the trachea, it's particularly the carina, indicative of left atrial enlargement or left sided heart enlargement, as it actually starts to get closer to the thoracic spine.
And whenever I'm looking at the radiographs, not only am I looking at the lung and the perihiar region, but I'm also looking at the vessels, or if I'm worried about right heart disease, I'm looking at the caudal cable diameter comparing to the aorta. So, whenever you have these reports, and they, if they say there's cardiomegaly, you want to confirm that by looking at those structures, but also look at the other structures, because those can be indicators of whether or not there's more severe cardiac disease. What things cause a pseudocardiomegaly or for the system to report cardiomegaly or a radiologist to report cardiomegaly, when there's not cardiomegaly.
Well, dexamethaomidine sedation will cause the heart to enlarge, so that can make your heart look large, it's actually a normal heart. Hypoinflation can actually cause a relative cardiomegaly because the lung volume's less, so the heart looks bigger within that space. AI can get tricked when there's actually part of that image is cropped.
It also reduces the lung volume, making the heart look bigger within the thoracic space. Some breeds, and particularly athletic dogs will have large hearts normally, so they can appear enlarged, but they, on echo, they, they would be normal. Other things are a phase of the cardiac cycle, positioning, laterality, right lateral versus left lateral, and VD versus DV.
The things that, that really impact, the AI are going to be whether the techs include that entire thorax or if they actually crop the image too closely. That will result in the heart being appearing larger to the system because of the relative lung volume. So, again, it's the reason why you want to make sure that the texts actually include all of the tholas, all of the lung lobe.
Left sided cardiomegaly, the sensitivity is a little bit lower, but the specificity is higher, and the sensitivity actually is affected by the minimal left atrial enlargements. When those are found on echo, they're, they're easier to identify than on radiographs. But as the left atrium increases, there's some pretty typical changes of elevation of the carina, which is what we look for a lot of times on that, whether it's on the lateral or VD projection.
And the left side of cardiomegaly is unpublished data. So, what's the cyst I'm looking for for left sided cardiomegaly? It's looking again at this left atrial margin to see if there's rounding to that, enlargement to it.
It's also looking to see if there's elevation to the trachea. And whenever you're looking at this left atrial margin, again, you want to make sure that you can or cannot see that area through here. In this case, this interstitial opacity is partially obscuring that left atrial margin.
So when you lose visualisation of that left atrial margin. You want to be thinking that there's perihilar edoema, because that's the first place that perihiar edoema will set up. It will also set up in this area here, just cauddle to it.
So any increase in interstitial pattern in this region, along with loss of that left atrial margin, you really want to be thinking about left heart failure. The other thing you want to think about that'll cause this soft tissue opacity would be hilar lymphadenopathy. And so hilar lymphadenopathy, though, will cause the trachea to deviate ventrally versus left atrial enlargement will cause the trachea to deviate dorsally.
So make sure to take a look at what the trachea is doing when you have a case of that soft tissue opacity. On the VD projection, you're looking to see if there's a soft tissue bulge in, along midline. And the reason I actually included, only half of this thorax is because again, you want to make sure that you look at the opacity of the lungs from cranial to caudal.
You want to see that the lungs actually, do they or do they not become more opaque as you move caudally. And here you can see that there's an increase in opacity in that right caudal lung lobe, which is actually resulting in an alveolar infiltrate. And so we have cardiomegaly, left atrial enlargement.
And a caudal lung lobe alveolar infiltrate, all of those are hallmarks for left heart failure. So whenever you have a case of cardiomegaly, be sure to look at the VD and look to see if there's an increasing opacity extending caudally. And if you have any questions about that, we're always available to send the images in to, have evaluated by the radiologist.
Right-sided cardiomegaly, how does, right-sided cardiomegaly, how does vitology do on right-sided cardiomegaly? And this is a published study we did with Oregon State University. Well, right-sided cardiomegaly is difficult to identify on radiographs.
It's difficult for radiologists, and as you can see, it's difficult for the AI system. So the sensitivity is actually around 60%, and the specificity is around 70%. And as right-sided cardiomegaly becomes more severe, it will improve detection, but if you, if the result comes back as a right-sided cardiac enlargement, things you're really gonna want to look at are the rounding to the right sided heart margin, but also make sure to look at this trachea, because if this trachea is elevated just over that heart base, that's a hallmark of right-sided cardiac enlargement.
And then the last thing that I'll look at with a right-sided cardiomegaly is the presence of pleural fluid, looking to see if there is or isn't pleural effusion. The vasculature I'll also evaluate when I see that there's right-sided cardiomegaly, and as you can see, the vasculature is quite enlarged. There's the main pulmonary artery sitting right there.
So this is a really severe case of pulmonary artery enlargement and causing this right-sided cardiac enlargement, and it's also affected the position of where the cardiac silhouette is actually contacting that sternum. So pleural effusion, as I mentioned on the last slide, whenever you have a case that's concerning for right-sided cardiomegaly, be sure to look for pleural effusion. The pathology system will actually evaluate for pleural effusion.
This is a case, or this is a study that we did with Tufts University, and we had a sensitivity of 90% in terms of identifying pleural effusion and a specificity around 80%. So, what is the system looking for? What's a radiologist looking for with pleural effusion?
The most telling sign of pleural effusion is this rounding of the caudal lung tips on the VD projection. A lot of times people will end up looking for pleural fissure lines, and I have an example of that, but I'll tell you why that can actually be a little tricky, because most commonly this is going to be pleural fluid sitting in this pleural fissure line. It will sit between the right middle and right caudal lung lobes, but fat opacity in really obese dogs can actually also sit there creating the artifactual appearance of pleural effusion.
So, whenever I'm looking to confirm pleural effusion or looking for a scant amount of pleural effusion, or the system's looking for pleural effusion, it's gonna be looking here at the causaphrenic angle, and it's gonna be . Looking to see if there's rounding of those caudal lung tips. So look right at that margin of the lung diaphragm interface on the VD projection for evidence of rounding.
The other area that will be affected by pleural effusion is Fluid will collect next to the cranial mediastinum, but that can also happen with fat deposition in the cranial mediastinum where it can look a little bit widened. So both the cranial mediastinal appearance and the appearance of a pleural line can be mimicked by fat, but this caudal long tip rounding will not be mimicked by fat. So, always look for that.
On the lateral projection, you're looking for the same things, you're looking for rounding of the lung lobes, as well as pleural fissure lines. All right, and then the owner discussion. So always start that discussion if you think you're gonna have radiographs that are gonna need to be taken or using any AI product, make sure the owner has given consent for GDPR reasons, and that way you can actually have the study immediately sent off for AI review and have the results back to you when you're sitting down with the owner.
Now, it's a personal preference as to whether you wanna go over the images with an AI tool that actually does measurements on the radiograph versus a pro-style report. The image in the bottom here is showing, AI actually reviewing the, images of the thorax and providing measurements for a vertebral heart score. And this is an example of a vitology report that has a conclusions, recommendations, as well as findings in a pro style.
So I recommend trying all these products out and seeing which one actually works better for your workflow. Patology is always available for questions. You can always send us an email at [email protected].
You can go to our website at pathology.net, and online, there is a recently updated guide to AI tools for veterinary medicine, and that is the title of it, and it was just updated for April of 2025. I think that's a good resource to look at products that are in the veterinary space.
And as I mentioned in other talks, AI should really help with efficiency and streamlining workflows, and if AI doesn't save you time, it's not worth paying for. Spend that time with your two and four-legged family members because happy pets equal a happy life. OK, so this is the 5th lecture.
This is Radilogy AI use cases for the canine abdomen. My name is Seth Wallach. I'm a founder and CEO of Vitology and Vitology Innovations.
Vitology does follow GDPR. Make sure that whatever product you're looking at also follows GDPR. The other disclaimer for this talk is that some of the images are of a single laterality example, but make sure when you're obtaining radiographs that you obtain orthogonal radiographs, if not 3 views.
OK, so accurate abdominal AI interpretation begins with proper radiographs. What vitology angels are looking for on the radiographs. And the same things as what we look for in the thoax, which are technique.
Positioning, are the images actually obtained at full expiration? And remember, for the thorax, it's full inspiration, but abdominal radiographs should be obtained at full expiration. Is the entire abdomen included, including that caudal part of the thorax all the way to the coxofemoral joints, and making sure that there's nothing superimposed over the anatomy that could impact the AI evaluation.
On this radiograph, nearly everything was done correctly in terms of including the entire abdomen, but what you will see is in the cranial abdomen, there's a little bit of spinal rotation, and that means that the abdomen's a little bit tweaked, and that can actually impact the AI results. So be aware of that when you're evaluating those AI results. The lateral projection, the rotation that you can see here is based on those ribs.
They're not superimposed. The rest of the image actually is also well done in terms of the caudal thorax being included, so we get all of that diaphragm, all of that liver, all the way to the coccofemoral joints. So, the whole abdomen's included and can be evaluated by the AI.
What I will say is if the animal's actually too big and doesn't fit entirely on the plate, you want to include the cranial 2/3 of the abdomen and obtain the caudal third of the abdomen on a separate radiograph. AI does better when including the diaphragm all the way to the caudal third of the abdomen than it does if you actually don't include the cranial, abdomen. And an example of that is this, where the cranial aspect of the liver was collimated out of the image, and that will actually create an artificially, an artificial result of micropatia on the report.
So, as I mentioned, always include the entire liver and part of that lung on the radiograph to obtain the best AI results. So where is AI really important? Well, it's really important in abdominal emergencies, and small intestinal obstructions or segmental bowel dilation is challenging for everybody.
That's the biggest thing that veterinarians say that they could really use assistance with is small intestinal obstructions. Abdominal fluid also can be difficult, particularly when there's only a small amount of abdominal fluid or a little bit of mesenteric inflammation. And the last thing we'll talk about is the different approaches to the AI results when talking to the owners.
Yeah, I'll give you a couple of examples of a pros report and then a measurement tool for when you're looking at the admin. So how does Vitology AI perform on small intestinal obstructions? The sensitivity is around 65%.
Specificity is a little over 60%. This is unpublished data, but what I can tell you is there are certain reasons when the AI is very accurate about small intestinal dilation. And when it actually is a little less accurate.
I'm gonna show you some examples, and so that when you're getting those results, you can actually correlate that with what the report is, is saying and what are actually on the images. So, this is a case of small intestinal segmental dilation. And when we're talking about segmental bowel dilation, we're talking about bowel that the maximum diameter is 2 times or more, 2 or more times the width of the smallest bowel loop.
So this is an example of segmental bowel dilation. And the AI had a greater than 75% likelihood of there being small intestinal segmental dilation on this image. Why did this result come up as so highly positive for segmental bowel dilation?
The reason is that the actual loops that are dilated. Are well visualised. So, we're not having any superimposition of any of the structures.
So just like a, a clinician or a radiologist looking at the images, AI needs to get a, a, does best when the actual image of the pathology is well defined and clearly visualised. When does AI actually hedge a little bit more in terms of whether there is or isn't small intestinal dilation? Well, when there's actually superimposition, when there's overlying bowel, when there's overlying abdominal fluid, in this case, the abdominal fluid partially obscures the belly, the, bowel dilation, and therefore, the AI came in around 50 to 75% likelihood of segmental bowel dilation.
Same for this example where the small bowel dilation is actually in the cranial abdomen. This is a duodenal foreign body, but the duodenum resides in an area where the transverse colon is, and there can, where the stomach sits as well. So it can be difficult to distinguish between the duodenum.
And what's in the stomach and what's in the colon. So in this case, there was a 50 to 75% likelihood of a segmental bowel dilation, but as you can tell in this cranial abdomen, this area here is obscured by a lot of the other structures. The VD projection duodenalform bodies also can be difficult to identify.
This is a loop of duodenum that's dilated, but that can also be a loop of colon, because the colon will also sit in that area. So it can make it very difficult for the AI to differentiate between the two, as well as for a radiologist or for a clinician. What can you do to assist on that?
You can actually do a pneumocolon. So if you get a result here where you see a loop of bowel in a vomiting dog, that is, could be duoum, could be colon. Do a pneumo colon, recheck the radiographs, and then you'll be able to distinguish whether that's colon or duodenum.
You can also send that in, the system will then be able to distinguish whether that's small or large bowel. In this case, the system actually said that there was a less than 50% likelihood of segmental bowel dilation. And again, because this is in the area where the ascending colon resides.
All right, so let's look at another example where AI actually has a little more difficulty in identifying segmental bowel dilation. In this case, this lateral abdomen, there was a less than 50% likelihood of small intestinal bowel dilation. The reason that AI actually gets confused by this, and it's a tough radiograph for anybody looking at it, is because there's rotation, which actually puts the cecum dorsal.
And the descending colon more ventral. So this is all large bowel, but the other thing for those who are very astute on this, we'll see this area here that's all fluid dilated. Well, that's actually also the colon coming around there.
And so, all of this is dilated colon. And so the segmental bowel dilation is really all of the other bowel that is uniform, in terms of diffusely gas-filled. And that will happen when you have a more distal small intestinal obstruction.
So these can be challenging cases for both AI and for radiologists. Another example of where AI could actually incorrectly or not identify small intestinal obstruction or small intestinal form material is when the contents within the small bowel actually look like contents in the colon. So in this case, the material here in this small intestinal loop looks a lot like faecal material.
And if there was a, a rotation, and sometimes the colon can actually dip a little bit ventral, this could easily be interpreted as colon. And so, the trick again on these types of cases where questioning whether it's small bowel or it's large bowel is to do a pneumocolon. Pneumocoons are.
Extremely helpful. And if you're not doing pneumocolons when you're wondering whether there's segmental small bowel dilation, it's something to add to the, they're easy to do and it's something to add to your list of, diagnostics. The other thing would be to do, an abdominal ultrasound.
So, typically, when I'm looking at the images, such as this, I'll recommend a pneumocolon or a small intestine, or, or an abdominal ultrasound to determine if there is or isn't small intestinal obstruction. Small intestinal fibrous form material, how does AI do on that? This is unpublished data, and it's about 70%.
And if it, you get a report that comes back and says that there's fibrous form material in the small bowel, what would you be looking for? You're looking for a Linear soft tissue opacity with gas, gas lines incorporated within it, so it creates this appearance and this is typical of cloth. Any type of cloth form body will create that.
It'll trap the gas within the soft tissue and it'll create that very fibrous look as you can tell, this also could look like colonic material. But the difference is that this bowel loop is actually heading cranial in two directions, and the colon should actually make its turn cranially and then go caudally. So those are the kind of indications that this is small intestinal, and the appearance is very typical for fibrous form material.
So if the report comes back with concern for fibrous form material, that's what you'd be looking for to see if there's those linear striations. Abdominal fluid, the AI performance on abdominal fluid, depending on severity, of course, the more severe the abdominal fluid, the more likely it will be correct. But, in general, the abdominal fluid is, sensitivity around a little below 70% specificity, also a little bit below 70%.
Abdominal fluid can be challenging and very similar to the interstitial pattern in the thorax. What you're looking for is extra wisps in the abdomen, and those wisps will obscure the bowel margins. And so if you're, Not able to clearly identify the bowel margins in parts of the image, that's when you start thinking about mesenteric inflammation or abdominal fluid.
Compare that image. To the image earlier that we looked at where there was segmental bowel dilation. Now there is a little bit of gas highlighting around this, which really makes this bowel stand out, but even some of the other bowel loops, you can see the bowel margins.
Right here there's a very clear bowel margin, and that's because there really isn't any fluid in that area. And when we go back to this other image. You can see how the bowel margin actually is not well defined.
That bowel margin typically is well defined when there's an abdominal fat. So, the other thing that will create that appearance can be a lack of intraabdominal fat. So, if you're not seeing bowel margins, you want to think about lack of intraabdominal fat, mesenteric inflammation, abdominal fluid.
So, as I mentioned in the thoracic study, or the thoracic talk. When you're talking with the owners, make sure to get GDPR consent about sending images off for radiology evaluation or AI radiology evaluation prior to obtaining the study. And when you're obtaining the radiographs, you have a couple of options.
You can either use a pro-style report, as the vitology report is above, where it's talking about the conclusions, or you can use AI tool for measurements on the images. Both are, are very helpful for, you're talking to the owner. It just depends which one you're actually more comfortable with and which one provides a better workflow for your clinic.
If you have any questions, you can always reach out to us at [email protected]. Our website is vitology.net, and for a guide to AI tools that was updated in April of 2025, talks about all different products that use AI in the veterinary space.
And if AI is not making your workflow, more streamlined and providing efficiency in the clinic, and if it's not saving you time, it's not worth paying for it. So make sure that whatever AI product you choose does give you more time with your family because happy pets equal a happy life. And I am Seth Wallach, and thank you for joining these talks, and please get in touch with us if you have any questions.
