A Metadata-Driven Approach to Malignant Tumor Identification in Dermoscopy Images based on Genetic-Inspired Crow Search Optimized Convolutional Multilayer Perceptron Network

Abstract

Introduction: Skin cancer is frequent and deadly thus, early detection enhances patient results. Skin lesions can be magnified using dermoscopy to detect cancerous tumors.
Objective: This paper proposed utilizing an Improved Convolutional Multilayer Perceptron Network (ICMPN) to detect malignant tumors in Dermoscopy Images (DI). Adding lesion features with medical data enhances the model's accessibility and usefulness.
Methods: The suggested model is trained and evaluated by the research using datasets from the ISIC 2019 and ISIC 2020. The preprocessing step focused on noise reduction to enhance image quality, which is critical for accurate analysis. Local Binary Pattern (LBP) separates important sections of images, enhancing feature extraction. Genetic-inspired Crow Search Optimization (GICSO) identified the most dissimilar traits. The approach entails extracting key metadata characteristics from healthcare records and using this data in the ICMPN architecture training process.
Results: A customized convolutional neural network, the ICMPN, dynamically learns and extracts hierarchical structures from DI to distinguish benign and malignant tumors. Compare the proposed method's accuracy, precision, sensitivity and specificity to evaluate its tumor detection efficacy.
Conclusions: The experimental findings illustrate the model's higher performance in identifying DI, highlighting its potential for clinical applications.