Unsupervised Local Discrimination for Medical Images

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Unsupervised learning, particularly contrastive learning, is crucial for medical image analysis due to annotation costs.
• Existing methods primarily learn global features, neglecting fine-grained local details essential for medical image interpretation.

Purpose of the Study:

• To develop a novel unsupervised representation learning framework, local discrimination (LD), for medical images.
• To capture local discriminative features by embedding semantically similar pixels and identifying similar structures across images.

Main Methods:

• The proposed local discrimination (LD) framework utilizes a pixel-wise embedding module and a clustering module for segmentation.
• A novel region discrimination loss function unifies these modules, enabling reflection of structural information and measurement of pixel-wise and region-wise similarity.
• LD is extended to develop a center-sensitive one-shot landmark localization algorithm and a shape-guided cross-modality segmentation model.

Main Results:

• The representation learned by LD demonstrated superior generalization across 12 downstream tasks, outperforming 18 state-of-the-art methods.
• LD achieved significant performance gains, particularly in challenging medical lesion segmentation tasks.
• The framework successfully integrates pixel-level and region-level feature learning.

Conclusions:

• Local discrimination (LD) offers a powerful unsupervised approach for learning discriminative features in medical images.
• The LD framework enhances the performance and generalizability of downstream medical image analysis tasks, including landmark localization and cross-modality segmentation.
• Focusing on local details is critical for advancing unsupervised representation learning in medical imaging.