Analysis of Skin Cancer feature pattern using GNN incorporating a Novel segmentation method.

Abstract

This study presents a comprehensive approach for the classification of skin lesions into benign and malignant categories using a combination of k-means clustering segmentation and advanced computational models. The methodology begins with the segmentation of lesion images through k-means clustering to extract relevant regions, followed by the computation of various features, including SIFT key points, WH-ratio, ECL-ratio, Harris corner, circularity, solidity, asymmetry score, color variation, and compactness index. Feature importance is determined through ranking and ANOVA tests. For model selection, both feature-based (GNN, GAT, 1D CNN) and image- based (VGG16, VGG19, Inception V3, MobileNetV1, MobileNetV2) models are evaluated, with the Graph Neural Network (GNN) emerging as the best model. The GNN model achieved a test accuracy of 98.62%. An ablation study is conducted to assess the contributions of individual components, and additional experimental analysis explores different thresholds in graph generation. The results are thoroughly analyzed using evaluation metrics, loss and AUC curves, statistical analysis, linear regression analysis, and 5-fold cross-validation. The proposed methodology demonstrates superior accuracy and robustness in the classification of skin lesions, offering a powerful tool for early detection and treatment planning.