ИСТИНА

Background & objectives. Complete manual labelling is extremely time-consuming process which limits possibility of developing deep learning algorithms for histological images analysis (e.g. the problem of segmentation of mucous glands). Development of semi-automatic interactive annotation tools helps a lot to solve this problem. Methods. The proposed algorithm for semi-automatic image segmentation works with scribble annotations and is based on a graph model with 2-stage label propagation. The weights of graph edges are predicted with a CNN, that was trained on a fully-labeled dataset. Results. The developed algorithm was trained and tested on the Warwick-QU dataset as well as PATH-DT-MSU dataset (https://imaging.cs.msu.ru/en/research/histology/path-dt-msu) containing 80 full-size annotated images of colon epithelial neoplasms. The proposed algorithm works in real time and allows to add new scribbles during the annotation which makes the labelling process interactive. Despite the high accuracy of the proposed algorithm we also offer a “classical” set of manual annotating tools to postprocess the results of the algorithm thus allowing annotator to finalize the annotation down to the smallest details. The developed algorithm allows to perform interactive labelling with scribbles and reduce the annotation time of one image from 150 minutes to 25-30 minutes. Conclusion. Using this semi-automatic interactive tool will significantly increase the number of fully annotated images of colon epithelial neoplasms in PATH-DT-MSU which is necessary for the development of real diagnostic algorithms. The work was supported by RFBR grant 19-57-80014 (BRICS2019-394). Ссылка на доклад: https://www.esp-congress.org/portal/congress/video-s.html?sessionId=5fa03aa42ceee79a64406e7b Ссылка на программу конгресса: https://www.esp-congress.org/scientific-programme/online-programme.html#!/by-topic/16