A Lightweight and Deployable CNN Framework with Interactive Preprocessing for Skin Cancer Detection

Abstract

Skin cancer is one of the most prevalent forms of cancer worldwide, and early detection plays a critical role in improving patient outcomes. This paper presents a lightweight convolutional neural network (CNN)-based framework for automated skin cancer detection, built upon MobileNetV2 and enhanced with adaptive preprocessing and augmentation strategies. Unlike prior works that rely on heavy architecture or large proprietary datasets, this study emphasizes practical deployability and transparency by combining modular preprocessing, a compact CNN backbone, and real-time deployment through a Gradio interface. The framework was evaluated on 400 dermoscopic images (200 malignant, 200 benign) sampled from the publicly available HAM10000 dataset. Data augmentation and dropout regularization were ap-plied to mitigate overfitting, while preprocessing operations—such as resizing, contrast enhancement, and denoising—were systematically compared through an ablation study. To establish baselines, we also trained ResNet50 and VGG16 models under the same conditions. Results show that Mo-bileNetV2 achieves competitive accuracy (87.5%) with lower memory requirements compared to ResNet50 (88.9%) and VGG16 (89.2%), while offering significantly faster inference times. ROC-AUC analysis confirms Mo-bileNetV2’s robustness (AUC = 0.91) under noisy and small-sample conditions. The findings highlight that carefully designed preprocessing and light-weight architectures can deliver reliable performance with reduced computational overhead, making them suitable for deployment in clinical and mobile settings. Future extensions will include multi-class lesion classification, lay-out-aware data augmentation, and integration with federated learning approaches for privacy-preserving medical AI.