ИСТИНА

Background & Objectives: There are difficult tasks in the diagnosis of colon epithelial neoplasms: lack of quantitative criteria of basal dilation of the crypts and spread of the serration, determination of potential malignancy. Automatic mucous glands segmentation using a convolutional neural network (CNN) is the first step to real diagnostic algorithm development. Methods: We propose a two step algorithm for glands segmentation. The first step produces semantic segmentation using a UNet-based CNN, while the second performs gland instance segmentation using a novel CNN architecture that predicts parameters for active contour model. This allows to segment each individual gland. We have designated closed-contour glands and “open glands” (glands with open contour). Results: The network was first trained on Warwick-QU dataset (165 images), fine network tuning was performed on the collected PATH-DT-MSU dataset (19 images, colon biopsy material). 12 images were hyperplastic polyps; 6 images were SSA/P and one was normal colon tissue. Our segmentation algorithm is characterised by Dice coefficient 0.87 on Warwick-QU and 0.78 on PATH-DT-MSU dataset. Dice coefficient decreased because of presence of «open glands» and the glands with adhered contours. Conclusion: It is necessary to create alternative collections of annotated histological images of colon epithelial neoplasms and to use full-size images obtained in the pathology examination of the real colon biopsies because images are cut off and only contain closed-circuit glands (mag.x200, x400) in the Warwick-QU. In contrast there are full-size real images with the presence of “open glands” (mag.x100) in PATH-DT-MSU dataset.