While in epilepsy pinpointing the location of a lesion is necessary to perform surgical intervention, in multiple sclerosis monitoring the appearance of new disease foci of various sizes allows doctors to lay out or adapt a treatment plan. The problem here is that models are good at finding large lesions on scans but often miss small ones. Thus, in both epilepsy and multiple sclerosis, the model has to be taught that sometimes pathologies manifest atypically, and the features observed in those cases are significant — not merely random noise. This is precisely what the authors of the study managed to achieve.
“By integrating biological features into the neural network, we make it work more in line with the actual reasoning of physicians,” said the study’s lead author, Nadezhda Alsahanova, a research engineer at Skoltech and a PhD student in the Institute’s Computational and Data Science and Engineering program. “Further development of this approach will make the results of computer vision analysis more interpretable. The model will justify its conclusions in a comment, making the findings more trustworthy in the eyes of a physician.”
The researchers tested their approach on clinical data. For both diseases, the efficiency of pathological area detection increased: by 2% for scans showing signs of multiple sclerosis, and by 8% — from 72% to 80% — on scans showing signs of epilepsy. Training also became more robust: The new model could match the performance of its predecessors with 30% less training data.
Technically, the innovation behind the improvement involved supplementing segmentation, which is a standard task in automated medical image analysis, with an auxiliary procedure, classification. Now, the model does not simply segment the image in the manner of “these areas are healthy tissue, and these other ones are pathological,” but also categorizes the latter regions based on which part of the brain they appear in, whether the suspected lesion has blurred boundaries, and so on. Thanks to this, the model begins to “pay attention” to biomedical features, which improves segmentation quality and — in the long run — allows the result to be accompanied by a comment for the doctor’s benefit.
Assistant Professor Maxim Sharaev, who co-directs the joint Skoltech-University of Sharjah BIMAI-Lab and was a principal investigator in the study, said: “Searching for small and atypical pathologies on MRIs is a critically important task, since missing them can cost patients the right treatment. What makes this study’s approach original is that when detecting a pathology, the model begins to rely on biomedical features, much like a human physician. Consequently, the output becomes interpretable and more suitable for real-world diagnostics. Importantly, Nadezhda, who came up with this idea and brought it all the way to MICCAI level, managed to confirm the findings on clinical data.”