Class key feature extraction and fusion for 2D medical image segmentation

Area of Science:

• Medical image analysis
• Deep learning in healthcare
• Computer vision applications

Background:

• Medical image segmentation faces challenges due to size variations, complex semantics, and high feature similarity.
• These factors hinder the performance and generalizability of deep learning models.
• Accurate segmentation is crucial for medical diagnosis and treatment planning.

Purpose of the Study:

• To improve deep learning model performance and generalizability in medical image segmentation.
• To overcome limitations posed by image size variation and feature complexity.
• To enhance the accuracy of semantic object localization.

Main Methods:

• Proposed the Key Class Feature Reconstruction Module (KCRM) to rank channel weights and select class-specific key features (KFs).
• Implemented KCRM to reconstruct local features, establishing dependencies on KFs.
• Introduced the Spatial Gating Module (SGM) using KFs to suppress irrelevant regions and focus on semantic objects.
• Diversified the receptive field to adapt models to varying object sizes.

Main Results:

• Integrated KCRM and SGM into the Class Key Feature Extraction and Fusion Network (CKFFNet).
• Validated CKFFNet on CHAOS, UW-Madison, and ISIC2017 datasets.
• Demonstrated superior segmentation results and generalizability compared to mainstream methods.

Conclusions:

• The proposed modules significantly improve medical image segmentation accuracy.
• Enhanced model generalizability makes the approach suitable for diverse clinical applications.
• The method shows potential for further development and expansion in medical imaging research.