Artificial intelligence use rising in veterinary radiology

Tech companies sign up thousands of users amid shortage of specialists

Published: March 02, 2021
By Ross Kelly

Image courtesy of SignalPET
Can computer software be taught to read an X-ray image, or radiograph, as effectively as a person? At least two commercial artificial intelligence products are now available in the veterinary sphere.

The number of veterinarians using artificial intelligence to interpret radiographs is growing rapidly, stirring questions about what role radiology specialists will play in the future of veterinary medicine.

Executives of the only two U.S.-based startups that appear to be offering the novel technology — SignalPET and Vetology AI — each told the VIN News Service that thousands of practitioners are now buying their products.

The computer software offerings employ artificial intelligence (AI) to read radiographs, commonly called X-rays, and provide an interpretation within minutes, for as little as $10 per study. Users access the software by signing into a website and have the option of linking it to their practice management systems.

Apart from purported efficiency benefits, growth in AI diagnostics is being driven by a shortage of veterinary radiology specialists, the companies say. The shortage, particularly in academia, has been recognized by the American Veterinary Medical Association and the American College of Veterinary Radiology.

"With so few radiologists in universities, this impacts not only the production of new radiologists, but also the radiology training of the DVM students," Dr. William Blevins, a professor emeritus of diagnostic imaging at Purdue University's veterinary school, said via email. "Both of these issues produce a 'vacuum' that AI will attempt to fill." (Blevins wrote a commentary in 2018 about the radiologist shortage.)

Veterinarians contacted by VIN News who have used the AI services offered mixed evaluations of their capabilities, ranging from enthusiastic praise to questioning their accuracy or applicability to certain conditions. The companies claim their products provide readings that correlate closely to interpretations by radiologists. They acknowledge the technology can be improved, including through constant feedback from users.

Adoption happening faster than expected

Vetology AI, based in California and launched in late 2017, charges a subscription of $200 per month for unlimited interpretations. SignalPET, based in Texas and launched in early 2020, charges $10 per reading. By comparison, interpretations by human specialists range from about $60 to more than $100 per image.

The vast majority of both companies' AI customers are based in the United States, although they said they're starting to expand to other markets, including in Canada, Europe, Asia and Australia. Dr. Neil Shaw, founder of SignalPET, is naturally bullish on the technology's prospects. "Undoubtedly, this is the future of veterinary medicine," he said.

Before he went into AI, Shaw, with his brother Darryl, founded BluePearl, an emergency and referral veterinary hospital chain, which they sold to Mars Inc. in 2015. Shaw worked for Mars until 2017, then started SignalPET.

Shaw said the company "significantly exceeded" a goal of having 1,000 paying veterinary customers in its first year of operation. He declined to say how many practices use the service, offering only that they include primary care, emergency and specialty practices, the majority independently owned. "We initially wanted to prove that the market really wanted this, and our goal wasn't actually to scale for the first year," Shaw said. "We've now got strong validation, and the goal is now to scale from where we are today, by quite a bit.”

Vetology AI is an arm of Vetology, a company founded in 2010 by radiologist Dr. Seth Wallack as a traditional teleradiology service. Its AI business is growing "in leaps and bounds," Wallack said, adding 30 customers in January alone.

Vetology AI currently has 2,935 clinics in its system, including some that are doing a free trial, according to the company president, Eric Goldman.

How it works

AI diagnostic tools are produced through a technique known as deep learning, which trains the tools to identify abnormalities that may indicate disease. Training involves feeding the AI thousands of radiographic images. Accuracy can then be strengthened via reviews by radiology experts and user feedback.

In human medicine, AI radiology diagnostic tools have shown "excellent accuracy," according to an article published last August in The Lancet. "However," the authors wrote, "outcome assessment in AI imaging studies is commonly defined by lesion detection while ignoring the type and biological aggressiveness of a lesion, which might create a skewed representation of AI's performance.” Moreover, the AI tools might enhance estimated sensitivity "at the expense of increasing false positives and possible over-diagnosis … " the paper says.

The application of AI to veterinary radiology seems relatively limited so far, with SignalPET and Vetology AI appearing to be the only companies dedicated to offering it, at least in the U.S., Blevins said. Blevins is a past ACVR president and a current diagnostic imaging consultant at the Veterinary Information Network, an online community for the profession and parent of VIN News.

Blevins noted that more limited forms of interpretative technology have been applied elsewhere. For instance, VIN offers a vertebral heart score calculator — developed by California-based Metron Imaging Software — that scans thoracic radiographs to assess heart size. An enlarged heart, or cardiomegaly, is a condition that can indicate any of several diseases. 

AI technology in veterinary medicine also has been assessed in academia. Last August, scientists from the University of California, Davis, published results of research applying AI to radiographs for the detection of canine left atrial enlargement. The analysis showed accuracy, sensitivity and specificity of 82.7%, 68.4% and 87.1%, respectively. "This study documents proof‐of‐concept for application of deep learning techniques for computer‐aided diagnosis in veterinary medicine," the researchers wrote. 

SignalPET and Vetology AI say their technology has been trained with hundreds of thousands of radiographic images. Shaw said SignalPET accumulated its image database through his friends in the veterinary community. Vetology AI tapped the vast trove of images reviewed by its teleradiology business.

SignalPET provides its service through a collection of apps, each dedicated to a specific condition, such as vertebral heart score, pleural fluid, pulmonary nodules, bone fractures and gastric distension. The company releases for commercial use only apps that are 95% accurate in recognizing what is normal or abnormal in a radiograph, Shaw said.

He added that clinical decisions should be based on a combination of factors, which could include clinical signs and laboratory diagnostics, not on AI assessments alone. "So, for example, the presence of lung infiltrate in a dog that is coughing is of different significance to a dog that is showing no pulmonary signs,” Shaw said.

Vetology AI, which on its website claims 92% agreement with interpretations by veterinary radiologists, has applied the technology a little differently. Wallack said its customers want a more holistic impression of what could be wrong with an animal rather than checking off a list of AI findings.

"Our core model is to take the synthesis of what's going on in the radiograph and give it back to the veterinarian in a format that agrees with what a radiologist would say for that entire body region," Wallack said.

Veterinarians give mixed reviews

Dr. Brad Singleton, medical director of South West Vet in Austin, Texas, first came across Vetology AI in April 2019 while attending a veterinary innovation conference at Texas A&M University. Drawn by its potential, he brought the technology to the company's three hospitals that summer. So far, he's impressed.

"This modality is fun for the novelty aspect of it, but has made radiology review for certain cases much more efficient, as well," Singleton said.

The product is especially helpful for junior colleagues in providing peace of mind, he said, calling the report accuracy "very impressive." Still, the program isn't perfect. One particular Vetology AI feature, which allows users to ask specific questions, doesn't always answer sufficiently. "The questions must be very basic, such as 'Is there evidence of obstruction?' rather than 'What is the soft tissue in the cranial abdomen?' " Singleton said.

SignalPET users include Dr. Kerry Peterson, medical director of Indianapolis-based Pet Wellness Clinics. Peterson said all nine of the company's clinics tap the technology daily. "Many of our locations have a single doctor, and it provides those veterinarians with a second set of eyes on radiographs for faster decision-making, thus providing faster treatment to patients," she said.

The software has picked up incidental conditions that might otherwise have been overlooked, she added. For example, when Peterson was imaging a cat for suspected bladder stones, the AI identified severe back arthritis, helping the veterinarian develop a suitable treatment plan.

Singleton and Peterson each provided testimonials to the respective companies about their positive experiences. Both said they received no financial inducements for their reviews and have no personal relationships with company executives.

Some practitioners have been less impressed with the technology. "We cancelled as it really was not worth the $10," Dr. Kathi Heiber, owner of the South Putnam Animal Hospital in Mahopac, New York, wrote on a message board of VIN. 

Heiber was among several veterinarians in the online forum who discussed their experiences using SignalPET. Heiber posted that in one particular radiograph, the AI had circled the entire thorax of a dog as having possible pulmonary nodules. She said the algorithm reads were weak and "more likely to produce further fruitless inquiry" if not used with a high degree of human evaluation.

Another practitioner who shared impressions was Dr. Bruce Henderson, hospital director of Valley Animal Hospital in Clifton, New Jersey. In a recent interview, Henderson said: "AI did not point out any real lesions that we did not see ourselves. More concerning, however, was that they sometimes identified lesions that were not really there."

Dr. Allison Ward, a relief veterinarian based in Fort Lauderdale, Florida, said one of the practices she worked with was so pleased with the SignalPET trial, it started paying for the service and is now using it with every radiograph.

"I think it's helpful, but it has its limitations," Ward said in an interview. "Right now, the technology will in no way, shape or form replace a radiologist's report — but it can give some peace of mind in situations where you are needing to quickly decide next steps for the patient."

Where will human interpreters fit in the future?

The radiology expert Blevins has no firsthand experience with the SignalPET or Vetology products but based on users' reports, he's concerned about their apparent limitations. Any number of variables might skew an AI interpretation, he said, including variations in breed, age, patient positioning and image quality.

"Both products are in the very early stages of development," he said. "Unfortunately, in my opinion, they have been pushed to the market too soon."

At the same time, Blevins said AI could be "very beneficial" with certain screening tests, such as hip dysplasia. "AI is also very good at measuring anatomical features," he said.

As for fears that AI could put radiologists out of work, Blevins believes such worries are unfounded. Radiologists could even end up welcoming AI's ascendency for making their jobs easier, he said.

"Yes, there may be a few vets that will accept whatever AI says about an image," he said. "But I think the real asset from AI will be that the system will pre-read the image studies and list the alterations that were found. That will speed up the interpretation process by the human viewing the study."

Shaw, responding to the concern that variables could skew AI results, said SignalPet's database intentionally is made up of real-world images of various breeds that don't always have a perfect positioning or exposure, to teach the system to adjust. And the system, he stressed, will only get smarter as the database expands.

"Now, are there some films that simply can't be read because they're overexposed, they're underexposed, the positioning is so poor?" he said. "Yes, but very similar to a human radiologist, the system can identify that, and those films don't get read."

At Vetology AI, if a customer is convinced the technology has made a mistake, Wallack said, they can send the image to the company for feedback. "The technology isn't going to replace the human radiologist, and that's because the technology will be good at a lot of things, but there's going to be some oddballs," he said. "There's different things that can have a similar look, or confuse the system, which is why radiologists are going to have a job."

How large a role radiologists will play in the future versus AI is an open question. "It's going to fall somewhere in between," Shaw predicted. "We'll determine where over time."

VIN News Service commentaries are opinion pieces presenting insights, personal experiences and/or perspectives on topical issues by members of the veterinary community. To submit a commentary for consideration, email

Information and opinions expressed in letters to the editor are those of the author and are independent of the VIN News Service. Letters may be edited for style. We do not verify their content for accuracy.