A Robust Mean Teacher Framework for Semi-Supervised Cell Detection in Histopathology ImagesDownload PDF

Published: 04 Apr 2023, Last Modified: 14 Apr 2023MIDL 2023 PosterReaders: Everyone
Keywords: Cell detection, Histopathology image analysis, Semi-supervised learning
TL;DR: We propose a robust mean teacher framework for semi-supervised cell detection in histopathology images.
Abstract: Cell detection in histopathology images facilitates clinical diagnosis, and deep learning methods have been applied to the detection problem with substantially improved performance. However, cell detection methods based on deep learning usually require a large number of annotated training samples, which are costly and time-consuming to obtain, and it is desirable to develop methods where detection networks can be adequately trained with only a few annotated training samples. Since unlabeled data is much less expensive to obtain, it is possible to address this problem with semi-supervised learning, where abundant unlabeled data is combined with the limited annotated training samples for network training. In this work, we propose a semi-supervised object detection method for cell detection in histopathology images, which is based on and improves the mean teacher framework. In standard mean teacher, the detection results on unlabeled data given by the teacher model can be noisy, which may negatively impact the learning of the student model. To address this problem, we propose to suppress the noise in the detection results of the teacher model by mixing the unlabeled training images with labeled training images of which the ground truth detection results are available. In addition, we propose to further incorporate a loss term that is robust to noise when the the student model learns from the teacher model. To evaluate the proposed method, experiments were performed on a publicly available dataset for multi-class cell detection, and the experimental results show that our method improves the performance of cell detection in histopathology images in the semi-supervised setting.
4 Replies

Loading