Biomedical Image Recognition in Pulmonology and Oncology with the Use of Deep Learning

The study is dedicated to solving the image processing tasks of different types, including detection and analysis of lesions, segmentation and recognition of biomedical images with the use of deep learning methods in computerized disease diagnosis. With the use of large datasets consisting of hundreds of thousands of medical images of different modalities we demonstrate the efficiency of the deep learning methods and state-of-the-art architectures of deep convolutional neural networks applied to solve different types of problems in pulmonology and oncology. Moreover, we show the examples of applying the neural networks for the generation of realistic, i.e., visually plausible and spatially close in feature space of medical images for their subsequent use in training classification neural networks and in any other areas with legal and/or ethical restrictions on the dissemination of personal data. We analyze the problem of the security of using the neural networks in the context of possible malicious adversarial attacks on deep neural networks in solving the tasks of diagnosing pulmonological and oncological diseases and present the corresponding quantitative estimates of their potential danger.