Semi-supervised medical image classification via Cross-Training and Dual-Teacher fusion model

Abstract

Pseudo-labeling approaches, a powerful paradigm for semi-supervised learning in medical image analysis, involve a teacher network to generate pseudo-labels and a student network to utilize the generated pseudo-labels. However, the generation and utilization of pseudo-labels are tightly coupled as both the teacher and student model share the same network. The inability of a single model to self-correct effectively can cause confirmation biases and potential error accumulation as the training proceeds. To address the problems, a novel semi-supervised framework fusing Cross-Training and Dual-Teacher (CTDT) is proposed in this paper. Firstly, a novel cross-training strategy is introduced, which adopts distinct architectural inductive biases within semi-supervised learning framework, enabling different models to mutually correct each other due to their varying learning capabilities and effectively preventing the direct accumulation of errors. Further, a dual-teacher fusion module is proposed to alleviate confirmation biases, which fuses complementary knowledge from diverged teachers to capture distinctive feature representations from unlabeled data and co-guide the student model. Extensive experiments on two public medical image classification benchmarks, i.e. skin lesion diagnosis with ISIC2018 challenge and colorectal cancer histology slides classification with NCT-CRC-HE, justify that our method (CTDT) achieves an average improvement of 2.48% on the NCT-CRC-HE and 3.13% on the ISIC2018.