A Two-Stage Deep Learning Framework for Skin Lesion Detection and Classification Using ResNet18 and EfficientNet-B4

Abstract

Skin diseases encompass a wide range of conditions that require an early and accurate diagnosis for effective treatment. This paper presents a two-stage deep learning framework for automated skin lesion detection and classification using deep convolutional neural networks. The first stage uses a ResNet18 model to detect the presence of a lesion in dermoscopic images. If a lesion is detected, the image is transferred to an EfficientNetB4 model for multiclass classification. Our approach integrates data augmentation, hair removal preprocessing, learning rate scheduling, and early stopping to enhance model performance and robustness. The framework is trained and evaluated on the HAM10000 dataset, addressing challenges such as class imbalance, model fine-tuning, and overfitting. Experimental results demonstrate the effectiveness of this method in accurately identifying and categorizing skin lesions, contributing to the advancement of deep learning-based dermatological diagnosis.