Evidence-based uncertainty-aware semi-supervised medical image segmentation
Yingyu Chen,Ziyuan Yang,Chenyu Shen,Zhiwen Wang,Zhongzhou Zhang,Yang Qin,Xin Wei,Jingfeng Lu,Yan Liu,Yi Zhang
DOI: https://doi.org/10.1016/j.compbiomed.2024.108004
IF: 7.7
2024-01-27
Computers in Biology and Medicine
Abstract:Semi-Supervised Learning (SSL) has demonstrated great potential to reduce the dependence on a large set of annotated data, which is challenging to collect in clinical practice. One of the most important SSL methods is to generate pseudo labels from the unlabeled data using a network model trained with labeled data, which will inevitably introduce false pseudo labels into the training process and potentially jeopardize performance. To address this issue, uncertainty-aware methods have emerged as a promising solution and have gained considerable attention recently. However, current uncertainty-aware methods usually face the dilemma of balancing the additional computational cost, uncertainty estimation accuracy, and theoretical basis in a unified training paradigm. To address this issue, we propose to integrate the Dempster–Shafer Theory of Evidence (DST) into SSL-based medical image segmentation, dubbed EV idential I nference L earning ( EVIL ). EVIL performs as a novel consistency regularization-based training paradigm, which enforces consistency on predictions perturbed by two networks with different parameters to enhance generalization Additionally, EVIL provides a theoretically assured solution for precise uncertainty quantification within a single forward pass. By discarding highly unreliable pseudo labels after uncertainty estimation, trustworthy pseudo labels can be generated and incorporated into subsequent model training. The experimental results demonstrate that the proposed approach performs competitively when benchmarked against several state-of-the-art methods on public datasets, i.e., ACDC, MM-WHS, and MonuSeg. The code can be found at https://github.com/CYYukio/EVidential-Inference-Learning .
engineering, biomedical,computer science, interdisciplinary applications,mathematical & computational biology,biology