Automating the problem of predicting cervical cancer recurrence using a conditional generative adversarial network

This paper presents an intelligent model based on the Pix2Pix conditional generative adversarial network that automates the process of predicting the recurrence of cervical malignancy in patients who have not yet undergone surgery. The implemented model accepts a pelvic MRI image as input data and provides an output probability of tumor recurrence and a generated image for the "post-operative" perspective. The presented model differs from its basic analogue by modifying the loss function for the problem conditions and replacing the standard generator with a convolutional neural network U-Net. Since the formulated problem belongs to the class of medical diagnostic tasks, the presence of false negatives of the intelligent model was reduced to zero by slightly increasing the number of false positives. In the process of comparative analysis of prognostic and real postoperative images, it was experimentally proven that the model not only accurately predicts the recurrence of the disease, but also generates almost identical centers of tumor foci and their relative areas on the magnetic resonance tomography image. The feasibility of modifying the basic version of Pix2Pix was confirmed by comparing the results of the two models using common quality metrics – precision, recall and their harmonic mean. The modification developed makes it possible to obtain prediction data in the shortest possible time, allowing it to be used in real-time mode.