How Intelligent Is Artificial Intelligence?

AI For Human And Veterinary Radiology

Radiology Artificial Intelligence, commonly referred to as AI, is in full development, and the FDA is actively testing AI in human radiology. In veterinary radiology we’re not far behind and are quickly catching up. Veterinary medicine AI product development’s greatest strength is also its greatest weakness. That is, oversight, or more specifically, the lack of oversight by a governing body.

In human radiology, AI products require FDA oversight and approval prior to coming to market (see the American Association of Veterinary Radiologists GMLP write-up here to learn more). Some veterinary products do fall under FDA oversight BUT veterinary radiology AI isn’t one of them. This means a veterinary radiology AI company can develop quickly but also have no formal obligation to demonstrate their product actually works.

If a company doesn’t provide clinical test results, it is up to you, the veterinarian, to determine the worthiness of the product. A true 'caveat emptor.'

Assessment Is Critical

So how should a veterinarian assess a veterinary radiology AI product? The same way the entire medical community evaluates any diagnostic test for a condition, by measuring clinical performance.

The two standard measures of clinical performance are SENSITIVITY and SPECIFICITY. To better understand how these measures can assist us to asses clinical performance, we will briefly revisit a couple of formulas from that old favorite - statistics class.

Sensitivity And Specificity

SENSITIVITY is the probability that a test will identify a patient who HAS a condition (true positive). It is calculated by the following formula:

Sensitivity = True Positives / (True Positives + False Negatives)

SPECIFICITY is the probability a test will correctly identify a patient who DOES NOT have a condition (true negative). It is calculated by the following formula:

Specificity = True Negatives / (True Negatives + False Positives)

The Confusion Matrix

These two standard measures of clinical performance lead us to the four outcomes possible for each patient:

True Positive, False Positive, True Negative, False Negative

The four outcomes are typically reported in a 2 x 2 table called a confusion matrix showing the total numbers of true and false positives and negatives. A generic example is shown here next.

A Typical Confusion Matrix

Vetology's AI Testing

Vetology’s AI testing evaluates AI results against veterinary radiologist reports as a reference standard. The results and confusion matrix tables are displayed just below on this page.

The Truth Is In The Confusion Matrix

To help you evaluate a product’s performance, always ask an AI vendor for their confusion matrix tables.

Also, be aware that an AI product MUST be 100% autonomous to have a valid result. If a human intervenes during any part of the result creation, it’s not artificial intelligence, it’s human intelligence.

Next we show the confusion matrix for several diseases among 75 random cases: