Skin Lesion Classification Using Additional Patient Information.

Area of Science:

• Dermatology
• Medical Imaging
• Artificial Intelligence

Background:

• Accurate skin lesion classification is crucial for early diagnosis and treatment.
• Existing methods often lack practicality for real-world clinical scenarios.
• The HAM (Human Against Machine) dataset provides diverse skin lesion classes for training and evaluation.

Purpose of the Study:

• To develop a simplified yet accurate skin lesion classification method.
• To improve classification performance by incorporating patient metadata and data augmentation.
• To achieve top performance in the ISIC 2018 skin lesion classification challenge.

Main Methods:

• A single-model deep learning approach for classifying skin lesions into seven categories (melanoma, nevus, basal cell carcinoma, etc.).
• Integration of patient metadata and augmented data into the classification network.
• Application of Test Time Augmentation (TTA) and Grad-CAM for enhanced inference and interpretability.

Main Results:

• Achieved a balanced multiclass accuracy of 88.7% with a single model and 89.5% with an embedding solution on the ISIC 2018 test set.
• The proposed method demonstrated superior accuracy compared to previous approaches.
• The algorithm ranked first on the ISIC 2018 live leaderboard.

Conclusions:

• Incorporating metadata significantly improves skin lesion classification performance.
• The simplified, single-model approach is practical for real-world clinical use.
• TTA and Grad-CAM integration aids in diagnosis by generating heatmaps for clinician assistance.