A dual-stream deep learning framework for skin cancer classification using histopathological-inherited and...

Area of Science:

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

Background:

• Skin cancer, especially melanoma, is a significant global health concern.
• Current diagnostic methods face limitations in accuracy, objectivity, and resource availability.
• Early detection is crucial for effective melanoma treatment and improved patient survival rates.

Purpose of the Study:

• To develop and evaluate a novel dual-stream deep learning framework for accurate and efficient skin lesion diagnosis.
• To integrate histopathological-inherited and vision-based features for a comprehensive lesion analysis.
• To overcome the limitations of traditional diagnostic approaches in skin cancer detection.

Main Methods:

• A dual-stream deep learning framework utilizing U-Net for segmentation.
• Feature extraction via Virchow2 (histopathological embeddings) and Nomic (vision-based features).
• Fusion of features followed by classification using a multilayer perceptron (MLP) on the HAM10000 dataset.

Main Results:

• Achieved a mean accuracy of 96.25% and a mean F1 score of 93.79% across 10 trials.
• Ablation studies confirmed the critical contribution of both feature streams.
• The proposed framework demonstrated superior performance compared to existing single-modality methods.

Conclusions:

• The dual-stream deep learning framework offers a robust and interpretable solution for skin cancer diagnosis.
• This approach has the potential to significantly enhance clinical applications for early and accurate skin lesion detection.
• The study highlights the efficacy of combining diverse feature extraction methods for improved diagnostic accuracy.