ИСТИНА

Background & objectives. Determination of foci of microinvasion of adenocarcinoma in polyps with low and high grade dysplasia is a rather difficult task, which can be tried to solve with the help of deep learning based methods. Methods. We use two datasets that are developed for the purpose of whole-slide images (WSI) segmentation and tissue type recognition: NCT-CRC-HE-100K and PATH-DT-MSU. Our PATH-DT-MSU dataset contains 20 H&E WSI of digestive tract tumors with pixel-level annotation of 6 tissue types. We solve the segmentation problem via classification approach, with a simple AlexNet-based CNN trained for patch classification. Results. The main goal for developing these algorithms is to automatically recognize the layers of the wall of the stomach and colon on WSI, namely the lamina propria, muscularis mucosa, submucosa, own muscle layer, subserous layer, serous membrane and adjacent areas adipose tissue. Since pixel-wise annotation of typical WSI is too time-consuming, we developed the patch classification model, applying which to overlapping patches results in getting coarse segmentation with reasonable accuracy. To adopt the model trained on NCT-CRC-HE-100K to PATH-DT-MSU we replace the last fully-connected layer and perform fine-tuning. The overall test accuracy of WSI classification is 0.93 on NCT-CRC-HE-100K and 0.8 on PATH-DT-MSU. Conclusion. Thus, we managed to develop an algorithm that detects layers of gastric mucosa and depth of invasion of intestinal-type gastric tumors with acceptable accuracy. The use of developed post-processing methods of segmentation contour analysis allows to detect depth of invasion in some cases of diffuse-type tumors. Also the next step is to train deep learning algorithms to segment tubular and papillary structures, low and high grade dysplasia, foci of invasive adenocarcinoma. This work was supported by RFBR grant 19-57-80014 (BRICS2019-394).