An integrated deep learning framework using adaptive enhanced vision fusion and modified mobilenet architecture for precision classification of skin diseases with enhanced diagnostic performance
Due to challenges such as illumination variability, noise, and visual distortions, machine learning (ML) and deep learning (DL) approaches for skin disease evaluation remain complex. Traditional methods often neglect these issues, leading to skewed predictions and poor performance. This research lev...
Saved in:
| Main Authors: | , , , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Elsevier
2025-10-01
|
| Series: | SLAS Technology |
| Subjects: | |
| Online Access: | http://www.sciencedirect.com/science/article/pii/S2472630325000895 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Due to challenges such as illumination variability, noise, and visual distortions, machine learning (ML) and deep learning (DL) approaches for skin disease evaluation remain complex. Traditional methods often neglect these issues, leading to skewed predictions and poor performance. This research leverages a diverse dataset and robust image processing techniques to enhance diagnostic accuracy under such demanding conditions. We propose Dermo-Transfer, a novel architecture that combines MobileNet with dense blocks and residual connections to improve skin disease severity classification by addressing problems such as vanishing gradients and overfitting. Our method incorporates multi-scale Retinex, gamma correction, and histogram equalization to enhance image quality and visibility. Furthermore, a quantum support vector machine (QSVM) classifier is employed to improve classification performance, providing confidence scores and effectively handling multi-class problems. The proposed approach significantly enhances diagnostic accuracy and outperforms previous models. Dermo-Transfer not only improves pattern recognition and classification accuracy but also robustly handles varying image quality and lighting conditions. Dermo-Transfer was trained on 77,314 images covering skin conditions such as molluscum, warts, eczema, psoriasis, lichen planus, seborrheic keratoses, atopic dermatitis, melanoma, basal cell carcinoma (BCC), melanocytic nevi (NV), benign keratosis, and other benign tumors. The Dermo-Transfer classification method achieved accuracies of 99 %, 98.5 %, 97.5 %, and 89 % across four datasets, demonstrating its effectiveness and potential utility for clinical diagnostics. Additionally, Dermo-Transfer outperformed SkinLesNet and MobileNet V2-LSTM in terms of classification accuracy. Experimental results also highlight how IoT devices and mobile applications can enhance the computational efficiency and practical deployment of the Dermo-Transfer model. |
|---|---|
| ISSN: | 2472-6303 |