Hierarchical skin lesion image classification with prototypical decision tree.

Area of Science:

• Dermatology
• Artificial Intelligence
• Machine Learning

Background:

• Traditional disease classification models often ignore misclassification costs and lack interpretability.
• Accurate and transparent skin lesion classification is crucial for effective clinical decision support.

Purpose of the Study:

• To introduce a novel Hierarchical Prototypical Decision Tree (HPDT) for skin lesion classification.
• To enhance diagnostic accuracy and reduce the severity of misclassifications.
• To improve the interpretability of classification models.

Main Methods:

• HPDT integrates prototypical networks and decision trees within a class hierarchy.
• A hierarchy-based distance matrix is employed to prioritize less severe misclassifications.
• The model predicts from general to specific categories, enhancing interpretability.

Main Results:

• HPDT demonstrated superior performance over flat classifiers and existing hierarchical methods on a large dataset (235,268 images, 65 conditions).
• The model achieved higher accuracy, reduced error severity, and improved interpretability.
• HPDT showed effective generalization to unseen classes.

Conclusions:

• Integrating clinical hierarchies into model design significantly improves diagnostic reliability and transparency.
• HPDT offers a promising approach for clinical decision support in skin lesion diagnosis.
• The model's ability to handle misclassification costs and provide interpretable results addresses key limitations of traditional methods.