Explainable AI-Based Skin Cancer Detection Using CNN, Particle Swarm Optimization and Machine Learning.

Area of Science:

• Dermatology
• Medical Imaging
• Artificial Intelligence

Background:

• Skin cancer is a global health concern, necessitating improved diagnostic methods.
• Traditional visual skin cancer diagnosis is subjective and time-consuming.
• Existing AI methods for skin cancer detection have limitations in efficiency and interpretability.

Purpose of the Study:

• To develop a comprehensive and efficient AI pipeline for automated skin cancer diagnosis.
• To enhance the accuracy and interpretability of AI-driven skin cancer detection.
• To address the computational and interpretability challenges of current AI approaches.

Main Methods:

• Utilized transfer learning with pretrained Convolutional Neural Network (CNN) models, selecting Xception as optimal.
• Implemented Particle Swarm Optimization for feature dimensionality reduction from 1024 to 508.
• Integrated machine learning classifiers (Subspace KNN, Medium Gaussian SVM) and Explainable AI (XAI) techniques (Grad-CAM, LIME, Occlusion Sensitivity).

Main Results:

• Achieved high diagnostic accuracies of 98.5% on the ISIC 2018 dataset and 86.1% on the HAM10000 dataset.
• Significantly improved computational efficiency through feature dimensionality reduction.
• Enhanced model interpretability using XAI techniques, providing insights into diagnostic decisions.

Conclusions:

• The proposed AI pipeline offers a robust, efficient, and interpretable solution for automated skin cancer diagnosis.
• This approach has the potential to significantly aid clinicians in early and accurate skin cancer detection.
• The integration of transfer learning, feature selection, and XAI represents a significant advancement in dermatological AI.