A semi-supervised deep convolutional framework for signet ring cell detection

Early detection of signet ring cell carcinoma (SRCC) can significantly improve patient survival rate. Pathological image analysis is applied as the golden standard for SRCC diagnosis. Automatic detection of pathological cells with deep learning methods can greatly reduce the burden of pathologists. Deep learning methods are commonly trained using large amounts of annotated data. However, due to the uneven distribution of medical resources and tedious manual examination procedure of high -resolution images, annotation data are usually insufficient and incomplete for deep learning model train-ing. In this paper, we propose a new semi-supervised deep convolutional framework to address the data annotation problem for signet ring cell detection. Specifically, we propose a self-training strategy to gen-erate pseudo bounding boxes based on Test Time Augmentation and modified Non-Maximum Suppression to re-train our detector. Our framework achieves 0.8774 in Valid Recall and 100.00 in FPs, winning the 1st place in the signet ring cell detection task of the Digestive-System Pathological Detection and Segmentation Challenge 2019. Code has been made publicly available at: https://github. com/ooooverflow/DigestPath2019. (c) 2021 Published by Elsevier B.V.

Automated summary

Early detection of signet ring cell carcinoma can improve patient survival rates significantly, and deep learning methods are effective in automatically detecting pathological cells. However, limited and incomplete annotated data for deep learning model training are often due to uneven distribution of medical resources and tedious manual examination procedures for high-resolution images.