STRUDEL: Self-Training with Uncertainty Dependent Label Refinement\n across Domains

2021/04/23 by Fabian Gröger, Gröger, Fabian, Anne-Marie Rickmann +3
Computer Science · #Computer Vision and Pattern Recognition (cs.CV) #Digital Imaging for Blood Diseases #Domain Adaptation and Few-Shot Learning #FOS: Computer and information sciences #Machine Learning and ELM

paper · pdf · doi:10.48550/arxiv.2104.11596

openalex publication_date 2021/04/23 · openalex created_date 2022/07/25 · openalex updated_date 2026/07/28

Abstract

We propose an unsupervised domain adaptation (UDA) approach for white matter\nhyperintensity (WMH) segmentation, which uses Self-Training with Uncertainty\nDEpendent Label refinement (STRUDEL). Self-training has recently been\nintroduced as a highly effective method for UDA, which is based on\nself-generated pseudo labels. However, pseudo labels can be very noisy and\ntherefore deteriorate model performance. We propose to predict the uncertainty\nof pseudo labels and integrate it in the training process with an\nuncertainty-guided loss function to highlight labels with high certainty.\nSTRUDEL is further improved by incorporating the segmentation output of an\nexisting method in the pseudo label generation that showed high robustness for\nWMH segmentation. In our experiments, we evaluate STRUDEL with a standard U-Net\nand a modified network with a higher receptive field. Our results on WMH\nsegmentation across datasets demonstrate the significant improvement of STRUDEL\nwith respect to standard self-training.\n

Related

• Two-phase Pseudo Label Densification for Self-training based Domain Adaptation
• Domain Adaptation for Medical Image Segmentation using Transformation-Invariant Self-Training
• Constraining Pseudo-label in Self-training Unsupervised Domain Adaptation with Energy-based Model
• Energy-constrained Self-training for Unsupervised Domain Adaptation
• Understanding Self-Training for Gradual Domain Adaptation