Scientists from Skoltech, Philips Research, and Goethe University Frankfurt have trained a neural network to detect anomalies in medical images to assist physicians in sifting through countless scans in search of pathologies. Reported in IEEE Access, the new method is adapted to the nature of medical imaging and is more successful in spotting abnormalities than general-purpose solutions.
Image anomaly detection is a task that comes up in data analysis in many industries. Medical scans, however, pose a particular challenge. It is way easier for algorithms to find, say, a car with a flat tire or a broken windshield in a series of car pictures than to tell which of the X-rays show early signs of pathology in the lungs, like the onset of COVID-19 pneumonia.
“Medical images are difficult for several reasons,” explains Skoltech Professor Dmitry Dylov, the head of the Institute’s Computational Imaging Group and the senior author of the study. “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,” the researcher adds. “Machines are good at something called a two-class problem. That’s when you have two distinct classes, each of them populated with lots of examples for training — like cats and dogs. With medical scans, the normal case is always grossly overrepresented, with just a few anomalous examples cropping up here and there. And even those tend to be different between themselves, so you just don’t have a well-defined class for abnormalities.”
Dylov’s group studied 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. 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.
What also sets the study apart is the proposed recipe for standardizing the approach to the medical image anomaly detection problem so that different research groups could compare their models in a consistent and reproducible way.
