A deep learning-based dual-branch framework for automated skin lesion segmentation and classification via dermoscopic...

Area of Science:

• Dermatology
• Medical Imaging
• Artificial Intelligence

Background:

• Early detection of skin diseases is crucial for patient survival.
• Limited access to dermatological expertise necessitates automated diagnostic solutions.
• Dermoscopic images are key for analyzing skin lesions.

Purpose of the Study:

• To develop a dual-branch deep learning framework for simultaneous skin lesion segmentation and classification.
• To integrate visual appearance and morphological features for enhanced diagnostic accuracy.
• To evaluate the framework's performance on multiple benchmark datasets.

Main Methods:

• A dual-branch deep learning architecture was employed.
• The segmentation branch utilized an EfficientNet-B7 encoder with ASPP and transformer blocks.
• The classification branch fused DenseNet-121 features with mask-derived morphology.
• Attention gates and Squeeze-and-Excitation blocks were incorporated for feature refinement.

Main Results:

• The framework achieved high segmentation performance across five datasets, with Dice scores up to 0.9568 and IoU up to 0.9242.
• Classification accuracy exceeded 0.95 across all datasets, with specificity above 0.93.
• The model demonstrated robustness and generalization capabilities on diverse datasets like HAM10000, PH2, ISIC 2016, ISIC 2017, and ISIC 2018.

Conclusions:

• The dual-branch framework effectively integrates visual and morphological features for comprehensive skin lesion analysis.
• State-of-the-art accuracy was achieved, highlighting the model's potential.
• Consistent high performance suggests strong clinical applicability as a diagnostic support tool.