An explainable hybrid deep learning framework for precise skin lesion segmentation and multi-class classification

Area of Science:

• Dermatology
• Artificial Intelligence
• Medical Imaging

Background:

• Skin disease diagnosis is challenging due to visual complexity and subjective manual examination.
• Malignant tumors like melanoma require accurate and timely diagnosis.
• Deep learning offers potential for objective and accurate skin lesion analysis.

Purpose of the Study:

• To develop a hybrid deep learning framework for simultaneous skin lesion segmentation and multi-class classification.
• To enhance model interpretability and clinical trust using explainable AI (XAI).
• To improve diagnostic accuracy for various skin conditions using dermoscopic images.

Main Methods:

• A dual-task architecture combining U-Net for segmentation and EfficientNet-B0 for classification was developed.
• Gradient-weighted Class Activation Mapping (Grad-CAM) was integrated for model interpretability.
• The model was trained and validated on the HAM10000 dataset.

Main Results:

• The model achieved a Dice coefficient above 0.85 for segmentation.
• Classification accuracy reached approximately 85%, demonstrating robust performance.
• Reliable results were obtained across diverse skin lesion types, despite class imbalance.

Conclusions:

• The hybrid deep learning model shows significant promise for accurate skin lesion segmentation and classification.
• Explainable AI (XAI) integration enhances transparency, crucial for clinical adoption.
• This approach can support dermatologists, particularly in resource-limited settings, by improving diagnostic efficiency and accuracy.