Skin lesion segmentation using two-phase cross-domain transfer learning framework.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Dermatology

Background:

• Deep learning (DL) models in medical imaging are limited by insufficient computing power and data scarcity.
• Domain shifts and ineffective data fusion hinder DL model generalization across multiple data sources.
• Current DL models face challenges in clinical decision-making due to complexity and low interpretability.

Purpose of the Study:

• To present a novel two-phase cross-domain transfer learning system for effective skin lesion segmentation from dermoscopic images.
• To address limitations in DL model generalization, data fusion, and interpretability for clinical use.

Main Methods:

• A two-phase cross-domain transfer learning approach (model-level and data-level) was employed, fine-tuning on MoleMap and ImageNet datasets.
• A high-performing DL network, nSknRSUNet, was developed, incorporating broad receptive fields and spatial edge attention for feature fusion.
• Model generalization was evaluated using skin lesion image datasets (MoleMap and HAM10000) from varied clinical contexts.

Main Results:

• The proposed model achieved 94.63% DSC and 99.12% accuracy on the HAM10000 dataset during data-level transfer learning.
• On the Molemap dataset, the model obtained 93.63% DSC and 97.01% accuracy in cross-examination at data-level transfer learning.

Conclusions:

• The developed system demonstrates excellent performance in skin lesion segmentation.
• The proposed method shows significant improvements over state-of-the-art techniques on both qualitative and quantitative metrics.