Image Analysis of Skin Diseases Using DenseNet-121 Architecture

Abstract

Skin diseases such as dermatitis, psoriasis, and tinea often exhibit similar visual characteristics, which can lead to frequent errors in early diagnosis. Accurate diagnosis is critical, as each disease requires different treatment approaches. This study aims to develop an automated classification model for these three skin diseases using a deep learning approach based on the DenseNet-121 architecture, which consists of 121 layers designed to facilitate efficient feature reuse and gradient flow. The dataset consists of 300 labeled images, evenly distributed among the three disease classes. To enhance model generalization, preprocessing steps were applied, including data normalization and augmentation techniques such as image rotation (±20°), horizontal and vertical flipping, random zooming (range 0.8-1.2×), and brightness adjustment (±20%). The model was trained and validated using a stratified 5-fold cross-validation strategy. Experimental results demonstrated an overall classification accuracy of 94.59%, with high precision and recall scores across all classes. These results indicate the potential of using DenseNet-based deep learning models as decision support tools for early skin disease diagnosis. Further validation with larger datasets and clinical input from dermatologists is recommended to ensure reliability in real-world healthcare settings. Visual comparison through Grad-CAM heatmaps was also conducted to enhance interpretability and validate model focus on relevant skin features.