Improved CNN architecture for automated classification of skin diseases
Rosacea, atopic dermatitis and bullous disease are significant skin conditions with distinct characteristics and impacts, which affect millions of people globally. Early and quick diagnosis is crucial for these conditions to prevent complications, alleviate symptoms and improve quality of life for a...
Saved in:
| Main Authors: | , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Taylor & Francis Group
2025-12-01
|
| Series: | Computer Methods in Biomechanics and Biomedical Engineering: Imaging & Visualization |
| Subjects: | |
| Online Access: | https://www.tandfonline.com/doi/10.1080/21681163.2024.2420727 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Rosacea, atopic dermatitis and bullous disease are significant skin conditions with distinct characteristics and impacts, which affect millions of people globally. Early and quick diagnosis is crucial for these conditions to prevent complications, alleviate symptoms and improve quality of life for affected individuals. This research article presents an innovative approach to automated classification of common skin diseases using Convolutional Neural Network (CNN) models. The study focuses on diagnosing rosacea, atopic dermatitis and bullous disease, leveraging CNN technology. Four pre-trained CNN models – DarkNet-53, ResNet-18, SqueezeNet and EfficientNet-b0 – were investigated. Additionally, two improvised versions of DarkNet-53 and an improvised version of ResNet-18 were developed, integrating a specific number of fully connected neural network layers and adjusting other CNN layers such as batch normalisation and activation function layers accordingly. The performance of these improvised models surpassed that of the original architectures, demonstrating superior accuracy in identifying the targeted skin diseases. In terms of overall accuracy, improvised versions of DarkNet-53 and ResNet-18 achieved an accuracy of 98% and 98%, respectively, while the original models could achieve an accuracy between 75% and 80%. This research contributes to advancing automated diagnostic systems for dermatological conditions, potentially improving early detection and treatment outcomes. |
|---|---|
| ISSN: | 2168-1163 2168-1171 |