October 27, 2021 – Scientists from Skoltech, Philips Research and Goethe University Frankfurt have trained a neural network to detect abnormalities in medical images to help doctors sift through countless scans for research of pathologies. Reported inIEEE Accessthe new method is suited to the nature of medical imaging and is more successful in spotting abnormalities than general-purpose solutions.
Image anomaly detection is a recurring task in data analysis in many industries. Medical scanners, however, pose a particular challenge. It is much easier for the algorithms to find, for example, a car with a flat tire or a broken windshield in a series of car images than to tell which x-rays show early signs of pathology in the lungs, such as the outbreak of COVID -19 pneumonia.
“Medical imaging is difficult for several reasons,” explained Professor Skoltech Dmitry Dylov, head of the institute’s Computational Imaging Group and lead author of the study. “On the one hand, the abnormalities are very similar to the normal case. Cells are cells, and you usually need a trained professional to recognize that something is wrong.
“On top of that, there is a dearth of examples of anomalies on which to train the neural networks,” the researcher added. “Machines are good for what’s called a two-class problem. That’s when you have two separate classes, each of which has lots of training examples, like cats and dogs. With the medical scans, the normal case is still grossly overrepresented, with only a few abnormal examples popping up here and there, and even those tend to be different from each other, so you just don’t have a well-defined class for abnormalities .
Dylov’s group studied four datasets of chest X-rays and histological microscopy images of breast cancer to validate the universality of the method on different imaging devices. While the advantage achieved and absolute accuracy varied widely and depended on the dataset in question, the new method consistently outperformed conventional solutions in all cases considered. What sets the new method apart from its competitors is that it seeks to “perceive” the overall impression a specialist working with the scans might have by identifying the very characteristics 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 can compare their models in a consistent and reproducible way.
“We propose to use what is called weakly supervised training,” says Dylov. “Since two clearly defined classes are not available, this task usually tends to be handled with unsupervised or non-distribution models. That is, abnormal cases are not identified as such in the training data. However, treating the abnormal class as a complete unknown is actually very strange for a clinical problem, because doctors can always cite a few abnormal examples. So we showed anomalous images to the network to release the arsenal of weakly supervised methods, and that helped a lot. Even one abnormal scan for every 200 normal scans goes a long way, and that’s quite realistic.
According to the authors, their approach – Deep Perceptual Autoencoders – is easy to transpose to a wide range of other medical scanners, beyond the two types used in the study, because the solution is tailored to the general nature of these images. Namely, it is sensitive to small-scale anomalies and uses few of their examples in training.
Co-author of the study and director of the Philips Research branch in Moscow, Irina Fedulova commented: “We are delighted that the Philips-Skoltech partnership allows us to meet challenges like this which are of great importance for the health care industry. We expect this solution to dramatically speed up the work of histopathologists, radiologists and other healthcare professionals faced with the tedious task of spotting tiny abnormalities in large image sets. By subjecting scans to preliminary analysis, obviously non-problematic images can be eliminated, giving the human expert more time to focus on the more ambiguous cases.
For more information: https://www.skoltech.ru/