LTPNet: Lesion-Aware Triple-Path Feature Fusion Network for Skin Lesion Segmentation

Area of Science:

• Medical image analysis
• Computer vision
• Artificial intelligence in healthcare

Background:

• Accurate skin lesion segmentation is crucial for diagnosis but challenged by complex backgrounds, ambiguous boundaries, and low contrast.
• Existing methods struggle with precise delineation, necessitating advanced deep learning approaches.

Purpose of the Study:

• To introduce the lesion-aware triple-path feature fusion network (LTPNet) for improved skin lesion segmentation.
• To address limitations in current segmentation techniques by enhancing feature extraction, refinement, and aggregation.

Main Methods:

• The proposed LTPNet framework utilizes an end-to-end approach with distinct extraction, refinement, and aggregation stages.
• Key modules include general foreground-background attention, attentive spatial modulator (ASM), lesion-aware lite-gate attention (LALGA), and triple-path feature fusion (TPFF).
• TPFF employs common, saliency, and difference paths to model multi-scale feature relationships.

Main Results:

• LTPNet demonstrated superior segmentation accuracy on both in-domain and cross-domain datasets.
• The model achieved competitive results with reasonable inference efficiency and manageable model complexity.
• Experiments confirmed the effectiveness of the proposed attention and fusion modules in enhancing segmentation performance.

Conclusions:

• LTPNet offers an effective solution for accurate and reliable skin lesion segmentation.
• The proposed network shows significant potential for clinical decision support systems.
• The lesion-aware design and multi-path fusion strategy contribute to robust performance across diverse datasets.