BCL-Former: Localized Transformer Fusion with Balanced Constraint for polyp image segmentation

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Polyp segmentation in medical imaging is difficult due to variable polyp size/shape and poor contrast with surrounding mucosa.
• Existing methods struggle with generalization across diverse datasets.

Purpose of the Study:

• To develop an advanced polyp segmentation method that addresses variability and improves generalization.
• Introduce the Localized Transformer Fusion with Balanced Constraint (BCL-Former) for enhanced colon polyp detection.

Main Methods:

• Proposed the Strip Local Enhancement (SLE) module for capturing enhanced local features.
• Introduced the Progressive Feature Fusion (PFF) module for smoother feature aggregation.
• Developed the Tversky-based Appropriate Constrained Loss (TacLoss) to balance true positives and false negatives.

Main Results:

• BCL-Former achieved state-of-the-art performance on four benchmark datasets for polyp segmentation precision.
• Demonstrated superior generalization ability across different datasets.
• The method showed a 5%-8% increase in training and inference speed compared to benchmark methods.

Conclusions:

• BCL-Former effectively overcomes challenges in polyp segmentation, offering high precision and generalization.
• The proposed method provides a faster and more robust solution for automated polyp detection in medical imaging.