A couple of years back, I and 350 other Veterinary Innovation Summit attendees witnessed a debate between Dr. Eli Cohen, a boarded radiologist from North Carolina State University, and Dr. Neil Shaw, representing the veterinary radiology company SignalPet. Others were on the panel, but none were as adamant about their opposing positions on artificial intelligence as those two men. Since then, I’ve heard everything from “It’s irresponsible for Dr. Shaw to have such a product on the market” to “AI is good for radiologists because it can help write reports faster.”
As someone who considers himself a techno-optimist, I wanted to better understand where the veterinary industry stands on using and developing artificial intelligence in veterinary radiology. The Gartner Hype Cycle can help us make sense of where we’ve come from, where we are and where we’re going.
Understanding the Cycle
The Gartner Hype Cycle represents in graphical form the maturity, adoption and social application of specific technologies. Detailed at gartner.com, the cycle consists of these five key stages:
- Innovation Trigger: A breakthrough product launch or other event generates significant interest.
- Peak of Inflated Expectations: Early publicity yields success stories and failures.
- Trough of Disillusionment: Interest wanes as experiments and implementations fail to deliver.
- Slope of Enlightenment: How the technology can benefit the enterprise is widely understood.
- Plateau of Productivity: Mainstream adoption starts to take off.
Where Are We Now?
Let’s review the state of radiology artificial intelligence through the Gartner Hype Cycle.
Innovation Trigger
AI in radiology first caught the attention of the veterinary community with its potential to revolutionize diagnostic imaging. Early research and pilot projects showcased AI’s ability to identify abnormalities in X-rays, CT scans and MRI scans with remarkable accuracy. For example, in 2017, researchers in Australia trained a system on 53,000 X-rays and demonstrated “diagnostic performance equivalent to a human radiologist.”
The more commonly cited case is CheXNet, which, according to the authors, uses an algorithm that has exceeded the detection capabilities of practicing radiologists.
This initial excitement marked the Innovation Trigger phase. People using the technology at this stage are typically called “innovators.”
Peak of Inflated Expectations
As word spread, expectations soared. Many believed AI would soon replace human radiologists, offering faster and more accurate diagnoses. This period of heightened expectations saw a surge in AI-driven radiology tools entering the market and promising to transform veterinary diagnostics. We were at the Peak of Inflated Expectations, a region populated by “early adopters.” This period is characterized by:
- Excessive enthusiasm and unrealistic projections. (“AI is going to replace all radiologists.”)
- A proliferation of success stories.
- Increased financial investment.
- A lack of complete understanding, leading to confusion and blowback from established industry members.
- A discrepancy between expectation and reality.
In this phase, we saw a proliferation of companies hit the market — for example, SignalPet and, to a less public extent, Vetology. SignalPet, being one of the first companies to traverse the breach, experienced the most community flak. Not being one to balk at the challenges of being first to market,
Dr. Shaw and his team persisted and continued to develop a product that is seeing success throughout the market.
One of the mistakes in this phase is the overgeneralization of results. For example, an algorithm identifying pneumonia doesn’t necessarily find pleural effusion. There are no reasons, in principle, why an algorithm wouldn’t be able to diagnose both. Still, a lot of work needs to be done (the generation of datasets, training the algorithm and verification) before it can accurately move from one to another.
What typically happens, however, is that the overinflated expectations are unfulfilled, and the impending disappointment leads to a disillusionment with technology.
Trough of Disillusionment
At this point, as with many emerging technologies, reality begins to set in. Veterinary practices that adopted AI radiology tools faced challenges with integration, high costs and training. Moreover, AI systems sometimes produced false positives or missed subtle abnormalities that a trained radiologist would catch. These setbacks led to a period of skepticism and disappointment, known as the Trough of Disillusionment, which is filled with people considered the “early majority.”
Despite the efforts of radiologists like Dr. Ryan Appleby and Dr. Ira Gordon to flatten the downward trend of disillusionment, several hurdles meant the Trough of Disillusionment would be less like a soft landing and more like a squirming patient falling off an exam table. The challenges included:
- Hasty generalizations.
- Not enough boarded radiologists.
- Recent veterinary school graduates who thought they were inadequately trained.
Fortunately, there is reason to believe we’re emerging from the trough and headed toward enlightenment.
Slope of Enlightenment
Today, AI in radiology is climbing the Slope of Enlightenment. This phase’s users, known as the “late majority,” are veterinary professionals who have gained a more nuanced understanding of AI’s capabilities and limitations. Rather than viewing artificial intelligence as a replacement for human radiologists, they increasingly see it as a valuable tool to augment their skills. Case studies have demonstrated how AI can assist in the early detection of diseases, improve workflow efficiency and enhance diagnostic accuracy when used with human expertise.
I recently watched a webinar in which Antech’s Dr. Diane Wilson discussed the diagnostic company’s successful use cases. (See AI’s Many Talents.) The development of these cases helps narrow the focus from overinflated expectations to when the technology works well, providing a solid foundation for the technology’s continued growth.
As we enter this phase, more companies will enter the market. Antech Imaging Services, PicoxIA, Radimal and others are gaining market share.
It’s worth mentioning again that the effort to develop use cases is monumental. Not only do we need massive, accurately annotated datasets — more than 50,000 is a good start — but we also must train the algorithm through supervised learning. In addition, we have to continue to validate the algorithm so it doesn’t drift and start diagnosing in unwarranted areas.
No matter how much progress we make, there is always a concern about confabulation — a particular result looks good but is, in fact, a mistake. Because of this, the radiology companies’ validation and supervision are always necessary.
Plateau of Productivity
While we are not yet at the Plateau of Productivity, the trajectory is promising. As AI technology evolves and integrates more seamlessly into veterinary practice, it is expected to become a standard component of radiological diagnostics. The focus is now on refining the tools, improving their accuracy and making them more accessible to veterinary practices of all sizes.
Validation studies continue to be published, identifying the areas where the technology will work well and needs refinement. For example, during Dr. Wilson’s talk, she said validation studies now underway will permit their algorithms to be used in veterinary dentistry.
What It All Means
For veterinarians, understanding the current stage of AI in radiology on the Gartner Hype Cycle is crucial because it helps set realistic expectations and informs decision-making regarding the adoption of AI tools.
Here are a few takeaways for veterinary professionals:
- Stay informed: Keep up with the latest developments in AI technology and its applications in veterinary radiology.
- Invest wisely: Consider the costs and benefits of integrating AI tools into your practice. Look for solutions that offer robust support and training.
- Collaborate: Work closely with radiologists and AI developers to ensure their tools meet your practice’s needs. I have yet to meet an AI radiology company that isn’t open to feedback.
- Educate your team: Provide ongoing training to ensure your employees can effectively use AI tools and interpret the results.
Here’s my final take:
The shortage of boarded radiologists will worsen in the years to come. The solution is not to take fewer radiographs and help fewer patients. Equally, the solution is not to blindly trust everything from an algorithm. A lot of work remains to be done, whether it means dataset creation and annotation, algorithm development and validation, and assiduous verification of the algorithms. The benefits to patients and veterinary professionals and the costs associated with using an algorithm point to the need for practitioners to use AI tools.
It’s been seven years since CheXNet’s article triggered innovation in veterinary radiology. It’s taken seven years for us to start climbing out of the Trough of Disillusionment toward the Slope of Enlightenment. Getting to the same point with ultrasound, endoscopy, cytology, scribing and differentials will take less than seven years.
THE HONEST TRUTH
It is equally foolish for me, the author, to avoid using artificial intelligence in writing this article as it is for veterinary professionals to avoid the responsible use of AI in their practices.
AI’S MANY TALENTS
According to Dr. Diane Wilson, a staff radiologist with Antech Imaging Services, successful use cases have trained artificial intelligence in these areas:
- Cardiovascular: Cardiomegaly, left atrial enlargement, left ventricular enlargement, right atrial enlargement, right ventricular enlargement, main pulmonary artery enlargement, aortic arch enlargemen, heart base mass effect, microcardia, vertebral heart score
- Pleural space: Pleural mass effect, pneumothorax, signs of pleural effusion, thin pleural fissure lines
- Pulmonary structures: Pulmonary interstitial nodules, pulmonary masses, tracheal narrowing or opacity, alveolar pattern, bronchial pattern, unstructured interstitial pattern, pulmonary vascular
- Mediastinum: Mediastinal widening/opacity, sternal lymph node enlargement
- Gastrointestinal: Gastric dilatation volvulus, small intestinal obstruction, gastric foreign material (debris), esophageal dilation, gastric distention
- Other areas: Hepatobiliary, additional abdominal, skeletal, soft tissue, urogenital