Bridging the Gap Between Theoretical Performance and Clinical Utility in Multi-Class Skin Lesion Diagnosis

Abstract

The escalating global incidence of skin cancer necessitates the development of robust, objective, and automated diagnostic systems capable of augmenting clinical decision-making. This study presents a rigorous comparative analysis of four landmark Convolutional Neural Network (CNN) architectures, ResNet-101, MobileNet-v3-Large, EfficientNet-B5, and Inception-v4, evaluated against the expansive and heterogeneous ISIC 2019 dataset. Comprising 25,331 high-resolution images across eight diagnostic categories, the dataset presents significant morphological challenges due to inherent visual ambiguity and class imbalance. Our findings reveal that EfficientNet-B5 achieves the highest predictive robustness with a peak accuracy of 0.8968 and an F1-score of 0.8458, leveraging its sophisticated compound scaling approach to capture subtle malignant markers. Concurrently, MobileNet-v3-Large demonstrated exceptional efficiency, yielding a nearly identical accuracy of 0.8965 with a minimal computational load of 0.4307 GFLOPs, making it a prime candidate for edge- computing applications. Despite its higher theoretical complexity, ResNet-101 provided the fastest real-world inference latency at 0.5032 ms, indicating superior hardware optimization. While these results underscore the transformative potential of deep learning in dermatology, misclassification patterns between melanoma and melanocytic nevi highlight persistent challenges in navigating fine-grained morphological boundaries. Ultimately, this research provides a holistic framework for selecting optimal architectural backbones based on specific clinical deployment constraints, bridging the gap between theoretical model performance and practical utility.