An optimal graph convolutional vision neural network with explainable feature optimization for improved skin cancer...

Area of Science:

• Dermatology
• Artificial Intelligence
• Medical Imaging

Background:

• High misclassification rates in early skin cancer detection lead to delayed treatment and reduced survival.
• Manual diagnosis suffers from inter-observer variability and human error.
• Traditional machine learning models face challenges with imbalanced datasets and feature generalization.

Purpose of the Study:

• To propose an Optimal Skin Cancer Classification Network (OSCC-Net) for improved early skin cancer detection.
• To enhance the robustness of classification models for minority lesion classes.
• To improve the interpretability and localization capabilities in skin lesion analysis.

Main Methods:

• Developed OSCC-Net using the International Skin Imaging Collaboration-2019 (ISIC-2019) dataset.
• Integrated Adaptive Minority Over-Sampling Procedure (AMOP) for dataset balancing.
• Employed Stochastic Neighbourhood T-Distilling driven Score-Weighted Class Activation Mapping (STND-SWCAM) for feature analysis and interpretability.
• Utilized Grizzly Bear Fat Increase Optimizer with Density-Based Spatial Neighbourhood Discovery Algorithm (GBFIO-DSNDA) for feature selection.
• Implemented a Graph Convolutional Vision Neural Network (GC-VNN) for classification leveraging spatial dependencies.

Main Results:

• OSCC-Net achieved 98.32% accuracy.
• The model demonstrated 98.43% precision, 98.40% recall, and 98.39% F1-Score.
• Significant improvements were observed compared to baseline methods in experimental evaluations.

Conclusions:

• OSCC-Net offers a substantial advancement in automated skin cancer classification.
• The proposed methods effectively address challenges of imbalanced datasets and feature generalization.
• The model shows potential for improving early detection rates and patient outcomes in skin cancer diagnosis.