An Interpretable DL Approach to Early Identification of Skin Diseases by Combined Using FC-ResNet with Lesion Indexing Framework
DOI:
https://doi.org/10.46647/rdems0205059Keywords:
Deep Learning, Skin Disease, Prediction Model, Convolutional Residual Network, Dice Score, Classification of Melanoma, Medical Imaging, Dermatological FeaturesAbstract
Skin diseases affect millions of people worldwide every year, and early detection is important to prevent severe health complications. Traditional skin disease prediction methods mainly depend on dermatology experts, making the diagnosis process time-consuming and dependent on expert availability. Deep Learning (DL) techniques can analyze skin disease symptoms with high accuracy and effectively handle inter-class similarities and inconsistent image patterns.
In this research, a hybrid deep learning model combining FC-ResNet and Lesion Index Calculation Unit (LICU) is proposed for accurate skin disease classification and lesion severity analysis. FC-ResNet extracts important spatial features, while LICU measures lesion intensity and pathological characteristics. Experimental results on different datasets show that the proposed FC-ResNet + LICU model achieved 95.9% accuracy, 95.4% precision, 95.1% recall, and 95.5% F1-score. The proposed model outperformed existing models such as VGG16, ResNet50, EfficientNet-B0, DenseNet121, DeepLabV3+, and Attention U-Net.