Explainable AI for skin disease classification using gradient-weighted class activation mapping and transfer learning...

Area of Science:

• Computer Vision
• Medical Imaging
• Machine Learning

Background:

• Medical data privacy concerns limit data availability for training computer vision models.
• Skin disease diagnosis often requires specialized expertise and can benefit from automated systems.
• Limited resources and data scarcity pose challenges for developing robust diagnostic AI.

Purpose of the Study:

• To evaluate the feasibility of using low-resource computer vision techniques for skin disease diagnosis.
• To assess the impact of data preprocessing and augmentation on model performance.
• To compare the effectiveness of five transfer learning models for diagnosing various skin conditions.

Main Methods:

• Utilized five transfer learning models: VGG19, VGG16, GoogleNet, XceptionNet, and Inception.
• Created imbalanced (4092 samples) and balanced (5182 samples) datasets.
• Applied data augmentation, image resizing, noise removal, and blur techniques for preprocessing.
• Assessed model performance on preprocessed data, including augmented balanced data.

Main Results:

• VGG-19 achieved 95.00% accuracy on the imbalanced dataset.
• VGG-16-Aug reached 97.07% accuracy on the balanced, augmented dataset, outperforming other models.
• Preprocessing and augmentation enhanced image quality, leading to increased accuracy.

Conclusions:

• Low-resource approaches combined with preprocessing effectively diagnose skin diseases.
• The VGG-16-Aug model with a balanced, augmented dataset shows significant potential for skin disease diagnosis.
• Explainable AI (Grad-CAM) aids in interpreting model outcomes and addressing data scarcity.