Abstract:
Skin diseases are prevalent health problems around the world and there effects on patients can be not only painful but disturbing. Misdiagnosis or late diagnosis may cause complications, chronic propagation and recurrence as its early detection is crucial. Classical diagnosis is relying on dermatologists, however there are a scarce number of specialists in many rural and underserved areas. Artificial Intelligence (AI) and Deep Learning (DL) are promising for automated dermatological assistance, with recent developments. However, their practical applicability is hindered by challenges including dataset imbalance, high inter-class similarity, variety of appearances of lesions, and the “black-box” nature of DL models. Furthermore, the overwhelming majority of studies have centred on cancer related to the skin and less common non-cancerous inflammations or infections are underrepresented.In this work, a hybrid deep learning model with XAI using Grad-CAM for multiclass skin disease classification on macroscopic images is proposed. They were divided into five groups: Psoriasis, Warts, Vitiligo, Nail Fungus and Healthy skin. A CNN with channel attention was proposed and several transfer learning models (VGG16, MobileNetV2, ResNet50, DenseNet121) were pretrained. To make best use of the complementary feature representations, three fusion-based hybrid architectures were presented. Of these, the FusionNet_Dense_ResNet (DenseNet121 + ResNet50) model achieved the best performance by simultaneously exploiting detailed local textures and deep hierarchical features. Hybrid algorithm has achieved an overall accuracy of 96.36% with good sensitivity and specificity between classes. The incorporated XAI module generated interpretable Grad-CAM heatmaps: it helped to elucidate disease-specific areas and enhanced the explanation of decisions. In general, our work presents an efficient and explainable framework which bolsters the applicability of AI system for the early detection of multiclass skin diseases.
