LGGC-Net: a local-global graph and color attention-based lightweight CNN for skin cancer classification

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Dermatology

Background:

• Clinical deployment of AI for skin cancer classification faces challenges like limited robustness, interpretability, and computational constraints.
• Existing deep learning models often suffer from large sizes, extensive training needs, and poor generalizability, hindering practical application.

Purpose of the Study:

• To develop LGGC-Net, a lightweight convolutional neural network (CNN) incorporating LGGC attention for enhanced discriminative feature learning and computational efficiency.
• To evaluate the robustness and generalizability of LGGC-Net on diverse skin tone datasets and under domain shift conditions.

Main Methods:

• Proposed LGGC-Net, a lightweight CNN utilizing Local, Global Graph, and Color (LGGC) attention mechanisms.
• Conducted experiments with various CNN backbones, assessed performance on external datasets with diverse skin tones, and performed ablation studies.
• Employed Gradient-weighted Class Activation Mapping++ (Grad-CAM++) and SHapley Additive exPlanations (SHAP) for model interpretability.

Main Results:

• LGGC attention consistently improved performance across evaluated CNN backbones.
• LGGC-Net achieved 88.05% accuracy (binary) and 76.1% accuracy (multiclass on HAM10000) with high accuracy per epoch and per million parameters.
• The model demonstrated an area under the curve exceeding 0.93 in both settings and outperformed existing methods on deployment-oriented metrics.

Conclusions:

• LGGC-Net is an effective, interpretable, and computationally efficient solution for skin cancer classification.
• The model's robustness and generalizability across diverse skin tones suggest its potential for practical clinical deployment.
• LGGC-Net offers a promising advancement for AI-driven dermatological diagnostics.