Optimizing CNN Kernel Sizes for Enhanced Melanoma Lesion Classification in Dermoscopy Images

Skin cancer, particularly melanoma, presents a significant global health challenge due to its increasing incidence and mortality rates. Current diagnostic methods relying on visual inspection and histopathological examination are subjective and time-consuming, often leading to delayed diagnoses. Recent advancements in machine and deep learning, particularly convolutional neural networks (CNNs), offer a promising avenue for transforming melanoma detection by automating precise classification of dermoscopy images. This study leverages a comprehensive dataset sourced from Kaggle, comprising 10,605 images categorized into benign and malignant classes. Methodologically, a custom CNN architecture is trained and evaluated using varying kernel sizes (3x3, 5x5, 7x7) to optimize melanoma lesion classification. Results demonstrate that smaller kernel sizes, notably 3x3, consistently yield superior accuracy of 93.00% and F1-scores of 96.00%, indicating their efficacy in distinguishing between benign and malignant lesions. The CNN model exhibits robust generalization capabilities with minimal overfitting, supported by high validation accuracy throughout training epochs. Comparative analysis with related studies highlights competitive performance, suggesting potential enhancements through advanced feature selection and optimization techniques. Despite these advancements, challenges such as dataset diversity and model optimization persist, particularly concerning underrepresented darker skin tones. The study underscores the transformative potential of CNNs in enhancing diagnostic accuracy and efficiency in dermatological practice, paving the way for improved patient outcomes through early detection and intervention strategies. Future research directions include refining segmentation techniques and expanding dataset evaluations to ensure the models applicability across diverse clinical settings. Ultimately, this research contributes to advancing melanoma diagnosis by integrating cutting-edge deep learning methodologies with clinical practice, thereby addressing current limitations and driving forward innovations in dermatological image analysis.