Skin Lesion Classification Through Test Time Augmentation and Explainable Artificial Intelligence.

Area of Science:

• Dermatology
• Medical Imaging
• Artificial Intelligence

Background:

• Physician skepticism towards AI in skin lesion classification stems from a lack of model transparency and explainability.
• Widespread clinical adoption of AI diagnostic tools requires enhanced trust and confidence.

Purpose of the Study:

• To develop a highly accurate AI algorithm for skin lesion classification with visual explainability.
• To investigate the impact of Test Time Augmentation (TTA) on Convolutional Neural Network (CNN) performance.
• To foster physician trust in AI diagnostic tools through improved transparency.

Main Methods:

• Employed t-distributed Stochastic Neighbor Embedding (t-SNE) for visualizing high-dimensional CNN features.
• Utilized Gradient-weighted Class Activation Mapping (Grad-CAM) to generate heatmaps for prediction interpretability.
• Evaluated six CNN architectures (EfficientNet, ResNet, ResNeXt) with and without TTA on the ISIC 2019 dataset.

Main Results:

• Test Time Augmentation (TTA) improved balanced multi-class accuracy of CNN models by up to 0.3%.
• Achieved a balanced accuracy rate of 97.58% on the ISIC 2019 dataset.
• Performance was comparable to or surpassed more complex methods like Vision Transformers (ViTs).

Conclusions:

• The developed AI algorithm offers high accuracy and visual explainability for skin lesion classification.
• TTA is an effective technique for enhancing CNN performance in dermatological AI applications.
• The study demonstrates the potential of explainable AI to increase physician trust and facilitate clinical adoption.