CBAM-Xception: An Attention-Guided Framework for Skin Cancer Classification

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Dermatology

Background:

• Skin cancer diagnosis requires early and accurate detection for improved patient outcomes.
• Current deep learning models for skin lesion classification face challenges like limited interpretability and class imbalance.
• Automated diagnosis systems need to focus on clinically relevant features while minimizing noise.

Purpose of the Study:

• To introduce CBAM-Xception, an explainable attention-guided deep learning model for enhanced skin lesion classification.
• To improve the accuracy and interpretability of automated skin cancer diagnosis.
• To address limitations of existing models, including class imbalance and irrelevant feature extraction.

Main Methods:

• Developed an attention-guided deep learning model (CBAM-Xception) integrating Xception backbone and Convolutional Block Attention Module (CBAM).
• Applied Contrast Limited Adaptive Histogram Equalization (CLAHE) for image contrast enhancement.
• Utilized geometric and color augmentation to mitigate class imbalance on HAM10000 and ISIC 2019 datasets.
• Froze initial Xception layers for fine-tuning and employed Grad-CAM++ for visualization.

Main Results:

• CBAM-Xception achieved high accuracy: 98.62% (AUC: 0.9997) on HAM10000 and 93.66% (AUC: 0.9939) on ISIC 2019.
• The model demonstrated superior performance compared to MobileNet and EfficientNet baselines.
• Grad-CAM++ visualizations confirmed the model's focus on clinically relevant lesion areas, enhancing interpretability.

Conclusions:

• CBAM-Xception offers a reliable solution for automated skin cancer diagnosis by combining high accuracy, interpretability, and robustness.
• The model effectively addresses class imbalance and focuses on discriminative lesion features.
• Further research may be needed to assess generalizability across diverse clinical settings and computational resources.