Source-free domain adaptive segmentation with class-balanced complementary self-training.

Area of Science:

• Computer Vision and Machine Learning
• Medical Image Analysis

Background:

• Unsupervised Domain Adaptation (UDA) is vital for applying models to new datasets, but labeled source data access is often restricted due to privacy.
• Source-Free UDA (SFUDA) offers a solution but struggles with class imbalance ('winner takes all') and noisy pseudo-labels in self-training.
• Existing SFUDA methods often fail to adequately segment minority classes and are susceptible to over-confident pseudo-label noise.

Purpose of the Study:

• To propose a novel framework, Class-Balanced Complementary Self-Training (CBCOST), to overcome limitations in Source-Free Unsupervised Domain Adaptation (SFUDA) segmentation.
• To enhance segmentation performance by addressing class imbalance and pseudo-label noise in SFUDA.
• To enable effective knowledge transfer from labeled source domains to unlabeled target domains without requiring source data during adaptation.

Main Methods:

• Developed a CBCOST framework that jointly optimizes pseudo-label self-training with two key components: Class-wise Balanced Pseudo-label Training (CBT) and Complementary Self-Training (COST).
• CBT utilizes fine-grained class-wise confidence and adaptive thresholds to select balanced pseudo-labeled pixels, mitigating the 'winner takes all' issue.
• COST employs a heuristic complementary label selection to filter out incorrect pseudo-labels, reducing noise and improving model robustness.

Main Results:

• The CBCOST framework demonstrated superior performance compared to existing SFUDA methods across various segmentation tasks.
• Evaluated on 2D/3D cross-modality cardiac and brain tumor segmentation, CBCOST achieved competitive results, comparable to traditional UDA methods that use source data.
• Experimental results validate the effectiveness of CBCOST in handling class imbalance and pseudo-label noise for improved SFUDA segmentation.

Conclusions:

• The proposed CBCOST framework effectively addresses critical challenges in SFUDA, namely class imbalance and pseudo-label noise.
• CBCOST offers a robust solution for domain adaptation when source data is unavailable, achieving high segmentation accuracy.
• This method shows significant potential for real-world applications in medical image segmentation where data privacy is a concern.