Dynamically aggregating MLPs and CNNs for skin lesion segmentation with geometry regularization.

Area of Science:

• Medical Image Analysis
• Artificial Intelligence in Dermatology
• Computational Pathology

Background:

• Accurate skin lesion segmentation is crucial for melanoma diagnosis.
• Challenges include blurred boundaries, variable shapes, and low contrast in dermoscopy images.

Purpose of the Study:

• To develop a novel framework, CFF-Net (Cross Feature Fusion Network), for supervised skin lesion segmentation.
• To improve the accuracy of delineating skin lesions in dermoscopy images for computer-aided diagnosis.

Main Methods:

• Proposed CFF-Net with dual-branch encoder (CNNs for local features, MLPs for global dependencies).
• Introduced a feature-interaction module for enhanced representation and an auxiliary task for geometric information.
• Utilized supervised learning for segmentation.

Main Results:

• CFF-Net outperformed state-of-the-art models on ISIC 2018, ISIC 2017, ISIC 2016, and PH2 datasets.
• Significant improvements in Jaccard Index scores were observed compared to U-Net.
• Ablation and cross-validation studies confirmed the effectiveness and generalizability of CFF-Net.

Conclusions:

• CFF-Net demonstrates strong performance in segmenting challenging skin lesions, including those with blurred edges and low contrast.
• The framework shows potential for broader applications in segmentation tasks requiring precise boundary delineation.