Deep Learning for Skin Disease Classification: A Comparative Study of CNN and CNN-LSTM Architectures

Abstract

Skin diseases, particularly melanoma and other types of pigmented lesions, constitute a significant portion of global health concerns due to their prevalence and potential severity. In recent years, deep learning (DL) has revolutionized image classification tasks in the medical domain, particularly using Convolutional Neural Networks (CNNs) for skin lesion analysis. However, traditional CNNs are limited to capturing spatial features, often overlooking sequential patterns and complex contextual cues inherent in dermatological imagery. This study explores the automated classification of pigmented skin lesions using the HAM10000 dataset, a diverse collection of 10,015 dermatoscopic images spanning seven diagnostic categories. Addressing challenges in computational dermatology, we leverage MobileNet-V2 and InceptionV3 deep learning architectures, optimized via transfer learning and advanced preprocessing techniques. Comparative evaluation is performed between baseline CNN models and their Long Short-Term Memory (LSTM)-augmented variants to assess improvements in classification performance through sequential feature modeling. Results indicate that LSTM integration enhances contextual feature learning, improving accuracy for underrepresented lesion classes, with InceptionV3+LSTM achieving the highest classification accuracy.