Abstract

Accurate and non-invasive diagnosis of dermatological diseases remains a significant clinical challenge due to visual similarity among different skin diseases and variability due to imaging environment. Thermal imaging has emerged as a promising diagnostic technique as it captures physiological variations related to inflammation and vascular abnormalities, independent of illumination conditions. In this study, a deep learning–based framework is proposed for dermatological disease identification using thermal images. The proposed framework employs two strategies, one DenseNet-169 based and another VGG19 for feature extraction, both followed by a classifier consisting of either convolutional neural network (CNN) or bidirectional long short-term memory (Bi-LSTM). The system has pre-processing unit, transfer learning-based deep feature extraction, and classification with CNN for investigating the importance of spatial learning and Bi-LSTM for the sequential learning. Six general categories of dermatological diseases, namely Acne, Allergy, Melasma, Milia, Psoriasis and Vitiligo, were used in the evaluations of experiments. The system performance is measured by common measures like accuracy, precision, recall, F1-score, confusion matrix, and ROC–AUC plots. We can see that CNN classifier outperforming Bi−LSTM classifier in both feature extractors. With DenseNet-169 features, the CNN classifier reaches almost perfect performance with an accuracy of 99.68%, which indicates a stable training behaviour, fast convergence, and good discriminative capacity. In comparison, the DenseNet-169 and Bi-LSTM model show significantly worse performance overall, reaching an accuracy of only 73.05% with a low generalization capability. In the case of VGG19 as feature extractor, CNN based classifier achieves a good accuracy of 99.35% while having robust ROC characteristics, while VGG19 and Bi-LSTM configuration achieved a relatively lower performance as 89.61% since they confound the data between classes. In general, the results provide evidence that convolutional classifiers perform better in using the spatially diverse representations that DenseNet-169 and VGG19 generated. Such a framework could be a scalable and reliable solution for automated thermal image–based dermatological diagnosis, with a strong promise for facilitating real-world clinical decision support.

Statistics