Perhaps not surprisingly, it’s easier for AI to identify a car with a flat tire in a series of care photographs than it is to detect the early signs of a diseased lung cell, so it can be tricky for radiologists to spot abnormalities when sifting through numerous scans in search of signs of disease.

But a multi-national team of scientist has trained a neural network that may be able to help. Specifically, scientists from the Russian Skolkovo Institute of Science and Technology (Skoltech), Philips Research and Goethe University Frankfurt have trained a neural network that is adapted to the nature of medical imaging and is more successful in spotting abnormalities than other, general-purpose solutions. 

“In recent years,” the team wrote in their report in IEEE Access, “deep learning techniques achieved important advances in image anomaly detection. However, these efforts were primarily focused on artificial problems with distinct anomalies in natural images. The medical anomalies, however, differ from those in the natural images. Contrary to the natural images, the anomalies in the medical domain tend to strongly resemble the normal data.”

Adding to the report, Skoltech Professor Dmitry Dylov, the head of the Institute’s Computational Imaging Group and the senior author of the study, explained,  “Medical images are difficult for several reasons.  For one thing, the anomalies look very much like the normal case. Cells are cells, and you usually need a trained professional to recognize something’s amiss. Besides that, there’s the shortage of anomaly examples to train neural networks on.”

For their study, Dylov’s group examined four datasets of chest X-rays and breast cancer histology microscopy images to validate the universality of the method across different imaging devices. While the advantage gained and the absolute accuracy varied widely and depended on the dataset in question, the new method consistently outperformed the conventional solutions in all of the considered cases.

According to the team, “what distinguishes the new method from the competitors is that it seeks to ‘perceive’ the general impression that a specialist working with the scans might have by identifying the very features affecting the decisions of human annotators.”

The authors also noted that their approach — dubbed “Deep Perceptual Autoencoders” — is easy to carry over to a wide range of other medical scans, beyond the two kinds used in the study, because the solution is adapted to the general nature of such images. Namely, it is sensitive to small-scale anomalies and uses few of their examples in training.

Photo by Alvarez/Getty Images