Authors: M.Swathi, B. Ranjana
Abstract: The impact of facial skin problems on one's look and general health makes them a significant dermatological issue. Timely treatment relies on early detection of many illnesses, but manual diagnosis may be time-consuming and frequently requires clinical skill. Convolutional Neural Networks (CNNs) and other deep learning advancements have shown exceptional promise in reliably analysing medical pictures. This study examines the efficacy of convolutional neural network (CNN) models for automated facial skin disease categorisation using clinical face photos. The illnesses under investigation include Rosacea, Actinic Keratosis, Seborrhoeic Keratosis, Lupus Erythematosus, Basal Cell Carcinoma, and Squamous Cell Carcinoma. One of the biggest publicly disclosed collections of face dermatological photos in China, the Xiangya-Derm dataset is used in the experiments. A subset of 2,656 facial photographs was created from this dataset. We took a look at the classification performance of five popular CNN architectures: ResNet-50, Inception-v3, DenseNet121, Xception, and Inception-ResNet-v2. In addition, pretrained models were able to enhance their feature extraction and classification capabilities via the use of transfer learning, which included drawing on an independent dataset that had the same illness categories gathered from other body areas. Experiments show that transfer learning always improves model performance, with Inception-ResNet-v2 doing the best overall. The model achieved recall values of 92.9% for Lupus Erythematosus, 89.2% for Basal Cell Carcinoma, and 84.3% for Seborrhoeic Keratosis on the facial-image test set of 388 clinical images. On average, the model achieved 77.0% recall and 70.8% precision across all six disease categories. The results show that transfer learning has the ability to improve automated skin disease detection in clinical practice and validate that CNN-based facial skin analysis may provide trustworthy assistance for computer-aided dermatological diagnosis.
DOI: http://doi.org/10.5281/zenodo.21375722
