A Novel Pseudo-Labeling Approach for Cell Detection Based on Adaptive Threshold

Abstract

The cell detection is not only significant to clinical diagnosis, but a challenging task in the field of computer-aided diagnosis. One of reasons for this challenge is that it is difficult to obtain sufficient labeled samples to train an excellent detection model. Labeling all cells in the image manually is a time-consuming task. In this article, we propose a semi-supervised learning approach that generating pseudo-labels through dealing with unlabeled samples to automatically extract additional information for the retraining of the model to reduce the manual labor cost. Differing from former pseudo-labeling methods, great efforts are made to boost the reliability of pseudo-labels. In our model, pseudo-labels are generated according to adaptive threshold to reduce the noisy labels and retain the effective information. Moreover, our model effectively avoid the impact of difficult-to-detect cells and inhomogeneous background in the image by distilling the training data with the implementation of “patch attention” when leveraging samples with pseudo-labels for retraining. Extensive experiments have been conducted on two datasets to verify the performance of our method. We obtain a performance close to that of 2+M labeled images in supervised learning with only 2 labeled images and M unlabeled images in a semi-supervised learning manner. It is worth mentioning that the state-of-the-art results are achieved by our model compared with other existing semi-supervised methods.

This work is supported by the Development Project of Jilin Province of China (Nos.20200801033GH, YDZJ202101ZYTS128), Jilin Provincial Key Laboratory of Big Data Intelligent Computing (No.20180622002JC), The Fundamental Research Funds for the Central University, JLU.