HCHS-Net: A Multimodal Handcrafted Feature and Metadata Framework for Interpretable Skin Lesion Classification.

Area of Science:

• Dermatology and Artificial Intelligence
• Medical Image Analysis
• Computational Pathology

Background:

• Accurate skin lesion classification is vital for early cancer detection.
• Current deep learning models face challenges with computational cost, interpretability, and transparency.
• Clinical deployment requires efficient and understandable diagnostic tools.

Purpose of the Study:

• To present HCHS-Net, a lightweight and interpretable multimodal framework for six-class skin lesion classification.
• To improve upon the limitations of existing deep learning approaches in terms of computational efficiency and transparency.
• To enable accurate and timely classification for potential point-of-care applications.

Main Methods:

• HCHS-Net extracts visual features using Color, Haralick (GLCM), and Shape (Hu moments) modules.
• A biomimetic architecture processes information hierarchically, mimicking human visual and dermatological diagnostic workflows.
• Visual features are combined with clinical metadata and classified using an ensemble of gradient boosting algorithms (XGBoost, LightGBM, CatBoost).

Main Results:

• HCHS-Net achieved 97.76% accuracy with only 0.25 M parameters, significantly outperforming deep learning baselines.
• Inference time is 0.11 ms per image, enabling real-time classification on standard CPUs.
• The model demonstrated perfect melanoma and nevus recall (100%) with high specificity (99.55%).

Conclusions:

• HCHS-Net offers a computationally efficient, interpretable, and accurate alternative to deep learning for skin lesion classification.
• The domain-informed handcrafted features combined with clinical metadata provide superior performance and transparency.
• The framework shows significant potential for clinical deployment and point-of-care diagnostics.