Improving skin lesion classification through saliency-guided loss functions.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer-Aided Diagnosis

Background:

• Deep learning significantly advances medical image analysis, especially for skin lesion classification.
• Challenges persist in achieving high classification accuracy and explainability in deep learning models for medical imaging.

Purpose of the Study:

• To enhance deep learning classifier performance and explainability in skin lesion classification.
• To introduce a method integrating saliency scores from explainable AI (XAI) into the loss function.

Main Methods:

• Developed a custom loss function by incorporating penalization weights derived from XAI saliency scores.
• Evaluated the method on HAM10000 and PH2 datasets using Inception-ResNet-v2, EfficientNet-B3, and ResNeXt classifiers.
• Tested with various XAI methods to assess their impact on classification performance.

Main Results:

• Achieved 94.3% accuracy on HAM10000 and 98% on PH2 datasets, outperforming baseline methods.
• Demonstrated accuracy improvements of 7% and 6% respectively, using an LRP-guided loss function over standard loss functions.
• Showcased substantial performance enhancements compared to state-of-the-art methods.

Conclusions:

• The proposed integration of saliency scores into the loss function effectively enhances deep learning model performance and reliability.
• This method provides a quantitative assessment of XAI technique effectiveness in improving classification outcomes.
• The approach addresses key challenges in medical imaging, offering more trustworthy AI-driven diagnostic tools.