Superpixel-guided class-level denoising for unsupervised domain adaptive fundus image segmentation without source...

Area of Science:

• Computer Vision
• Machine Learning
• Medical Image Analysis

Background:

• Unsupervised domain adaptation (UDA) methods align data distributions between source and target domains.
• Traditional UDA requires both labeled source and unlabeled target data.
• Privacy concerns often prevent sharing of source data, hindering classical UDA.

Purpose of the Study:

• To develop a novel framework for adapting pretrained models to a target domain without access to the original source data.
• To address the challenge of noisy pseudo-labels generated by source models on target domains due to domain shift.
• To enable effective model adaptation in privacy-sensitive scenarios.

Main Methods:

• Proposed a Superpixel-guided Class-level Denoised self-training (SCD) framework.
• Implemented an adaptive class-aware thresholding for balanced pseudo-label generation.
• Utilized masked superpixel-guided clustering for enhanced feature centroids and prototypical label denoising.
• Employed adaptive learning schemes for both noisy and correctly labeled pixels.

Main Results:

• The SCD framework demonstrated superior performance in multi-site fundus image segmentation.
• Each component of the SCD framework was shown to be effective.
• The approach successfully adapted pretrained models without source data, outperforming existing methods.

Conclusions:

• The proposed SCD framework offers an effective solution for unsupervised domain adaptation when source data is unavailable.
• The method leverages self-training with novel denoising techniques to overcome domain shift challenges.
• SCD provides a practical approach for privacy-preserving model adaptation in medical imaging.