A Collaborative Self-supervised Domain Adaptation for Low-Quality Medical Image Enhancement

Area of Science:

• Medical image analysis
• Artificial Intelligence
• Computer Vision

Background:

• Poor medical image quality and inconsistent illumination hinder accurate clinical decision-making and can lead to misdiagnosis.
• Existing image enhancement methods often require high-quality reference images, which are difficult to obtain in clinical settings.

Purpose of the Study:

• To develop a fully self-supervised learning approach for enhancing medical image quality without requiring paired or high-quality reference images.
• To investigate the combination of self-supervised learning and domain adaptation for robust medical image enhancement.

Main Methods:

• Proposed a Domain Adaptation Self-supervised Quality Enhancement (DASQE) framework.
• Established multiple domains at the patch level using a rule-based quality assessment and style clustering.
• Formulated image quality enhancement as a collaborative self-supervised domain adaptation task to disentangle image quality, content, and style.

Main Results:

• DASQE achieved state-of-the-art performance on six benchmark medical image datasets.
• Demonstrated significant advantages in downstream clinical tasks, including segmentation (retinal fundus, nerve fiber, polyp, skin lesion) and disease classification.
• The method effectively enhances image quality while maintaining style consistency.

Conclusions:

• The DASQE framework offers a powerful solution for medical image quality enhancement in a fully self-supervised manner.
• This approach is beneficial for a wide range of clinical image analysis applications, improving diagnostic accuracy and efficiency.
• Self-supervised learning combined with domain adaptation shows great potential for addressing challenges in medical imaging.