Multi-center validation of Ladybug Beetle Optimized Convolutional Capsule Neural Networks with Explainable AI for skin cancer classification using dermography images

Abstract

Skin cancer especially melanoma is one of the most common cancers around the world and needs early and accurate detection to improve the results for patients. Traditional methods for diagnosis have problems like being subjective and inconsistent which shows the need for better computer-based solutions. While deep learning techniques show promise in automating detection through dermoscopic image analysis existing models struggle with limited generalizability high computational demands and class imbalance in datasets. To address these limitations, this work proposes the Multi-Center Validation of Ladybug Beetle Optimized Convolutional Capsule Neural Networks with Explainable AI (LOCapsNet-XAI) for skin cancer classification using dermography images. The proposed workflow for skin cancer classification using LOCapsNet-XAI begins with the acquisition of multi-center clinical data contains diverse dermoscopic images from various healthcare institutions to create a representative training dataset. Next, image preprocessing techniques such as Anisotropic Diffusion and Kuwahara Filtering are employed to enhance image clarity by reducing noise while preserving important features like lesion boundaries. Following preprocessing, feature extraction is performed using the innovative Convolutional Capsule Neural Network (CapsNet) architecture, which effectively captures complex patterns and spatial relationships within the images. The model's parameters are then optimized using the Ladybug Beetle Optimization Algorithm (LBOA), which enhances the exploration and exploitation capabilities to improve classification performance. To foster trust in AI-assisted diagnoses, Explainable AI methodologies specifically Grad-CAM++ are integrated into the framework, that provides the clinicians with visual insights into the model's decision-making processes. Finally, the workflow culminates in the classification of skin lesions by accurately identifying them as benign or malignant and facilitating informed clinical decision making. The proposed LOCapsNet-XAI method achieves exceptional performance metrics including 99.992% accuracy; 99.99% precision; 99.988% specificity; 99.99% recall; 99.98% F1-score; 91ms computation time and an AUC value of 0.99. These findings underscore the capability of the proposed model to enhance early skin cancer detection and improve clinical outcomes. In conclusion, LOCapsNet-XAI represents a significant advancement in the automated detection of skin cancer facilitating reliable and interpretable diagnostics in diverse clinical environments.