A Lightweight Convolutional Neural Network for Browser-Based Skin Disease Classification Using the HAM10000 Dataset

Introduction: Malignant skin lesions present a severe global health risk, necessitating rapid and accessible diagnostic screening tools. While deep learning models offer high classification performance, traditional transfer learning architectures are computationally intensive, requiring GPU acceleration that limits edge deployment. This study proposes a custom, lightweight Convolutional Neural Network (CNN) engineered specifically for resource-constrained, browser-based environments.

Methods: The proposed architecture was trained on the multi-source HAM10000 dataset, encompassing seven distinct diagnostic categories of pigmented skin lesions. To accommodate deployment efficiency, images were preprocessed and resized to $64\times64\times3$ pixels. A streamlined CNN pipeline featuring localized convolutional layers, max-pooling, and strategic dropout (0.5) was trained entirely in a local CPU environment using the Adam optimizer and categorical cross-entropy loss over 10 epochs.

Results: The optimized lightweight model achieved an overall test accuracy of approximately 72%. Due to the inherent class imbalance within the HAM10000 dataset, performance was rigorously verified using macro-averaged metrics, yielding a macro precision of 0.70 and a macro recall of 0.69. The model demonstrated strong stability on majority classes while maintaining a low computational footprint.

Conclusions: The trained model was successfully integrated into an interactive, web-based graphical user interface (GUI), enabling real-time image uploads and diagnostic predictions. While limited by raw accuracy compared to heavy architectures, this work validates the feasibility of deploying custom deep learning systems on CPU-level frameworks, serving as a scalable foundation for accessible, educational telemedicine tools.