Artificial intelligence in diagnosis of prostate cancer using magnetic resonance imaging. New approach

Aim. To improve the diagnosis of prostate cancer by training a neural network to identify malignant tumor lesions using the results of magnetic resonance imaging (MRI) studies with the same or greater accuracy than an experienced radiologist, using as the truth histological mapping of slides performed by a morphologist. Materials and methods. The work was performed at the "Zdorovie" Clinical and Diagnostic Center in Rostov-on-Don. Patients selected for the study underwent MRI in the Philips Ingenia 3.0T machine according to the prostate multiparametric MRI protocol, which complies with the requirements of PI-RADS v.2.1. The obtained data was used to train a convolutional neural network based on the U-Net architecture. The correct map of the actual locations of prostate cancer lesions was obtained using the "Morphologist's digital mapping tool" software. Results. The research part of the work consisted of following stages: center dot development of the "Morphologist's digital mapping tool" software for virtualization of lesions; center dot analysis of MRI data archive, retrospective selection of patients; center dot mapping of data by a morphologist to identify lesions in the prostate with layer-by-layer transfer of visualized lesions in the histological preparation to the image of the prostate gland in the "Morphologist's digital mapping tool", as well as training of the neural network to identify the presence of a malignant neoplasm in the prostate, location of the lesion(s), clinically significant disease; center dot data validation. For a certain amount of input data and high-quality mapping of this data, the neural network is capable of detecting prostate cancer lesions with the same accuracy as an experienced radiologist. Validation showed that the neural network correctly localized prostate cancer in 78 % of cases, while the radiologist did so in 55 % of cases. Comparative analysis also revealed the ability of the neural network to detect prostate cancer in areas of the prostate where the radiologist could not recognize any visual patterns indicating the presence of prostate cancer. Conclusion. Training a neural network without the participation of a radiologist is a fundamentally new approach that allows to sidestep the experience and qualifications of a radiologist in interpreting the obtained multiparametric MRI images.