A novel two‐staged network for skin disease detection using atrous residual convolutional networks

Abstract

SummaryThe preferred modality for the identification of skin disease has predominantly been dermoscopy. Skin lesion identification is the primary diagnostic basis for the dermatologist to evaluate the severity and impact of the disease. Current diagnostic procedures for the identification of skin lesions may be subjected to misdiagnosis among observers. Also, these procedures are conventionally labour‐intensive and prone to delays in treatment. Deep learning has proven to be beneficial for automated computer‐aided diagnosis in the medical field. This research presents a two‐stage approach involving segmentation and classification architectures for the effective detection of skin disease. The proposed approach aims to combine the architectural benefits of residual learning with the contextual retention of atrous convolutions. The classification is performed on seven different classes of skin lesions namely, melanoma (MEL), melanocytic nevus (NV), basal cell carcinoma (BCC), actinic keratosis/Bowen's disease (AKIEC), benign keratosis (BKL), dermatofibroma (DF), and vascular lesion (VASC). To the best of our knowledge, this is the first research that reports the impact of a two‐staged detection network using atrous residual convolutions for these seven classes. Experimental observations indicate that the proposed model is characteristically balanced in terms of inter‐class performance, with precise segmentation. The proposed network was trained with the HAM10000 dataset and reported a classification accuracy and precision of 89.27% and 89.06%, respectively.