Unsupervised domain adaptation with multi-level distillation boost and adaptive mask for medical image segmentation

Abstract

The mean-teacher (MT) framework has emerged as a commonly used approach in unsupervised domain adaptation (UDA) tasks. Existing methods primarily focus on aligning the outputs of the student and teacher networks by using guidance from the teacher network’s multi-layer features. To build upon the potential of the MT framework, we propose a framework named Multi-Level Distillation Boost (MLDB). It combines Self-Knowledge Distillation and Dual-Directional Knowledge Distillation to align predictions between the intermediate and high-level features of the student and teacher networks. Additionally, considering the complex variability in anatomical structures, foregrounds, and backgrounds across different domains of medical images, we introduce an Adaptive Masked Image Consistency (AMIC) approach. It provides a customized masking strategy to augment images for source and target domain datasets, using varying mask ratios and sizes to improve the adaptability and efficacy of data augmentation. Our experiments on fundus and polyp datasets indicate that the proposed methods achieve competitive performances of 95.2%/86.1% and 97.3%/89.0% Dice scores for optic disc/cup on REFUGE$\to$RIM, REFUGE$\to$Drishti-GS, and 78.3% and 86.2% for polyp on Kvasir$\to$ETIS and Kvasir$\to$Endo, respectively. The code is available at https://github.com/Yongze/MLDB_AMIC.