Skin lesion segmentation and classification: A unified framework of deep neural network features fusion and selection

Abstract Automated skin lesion diagnosis from dermoscopic images is a difficult process due to several notable problems such as artefacts (hairs), irregularity, lesion shape, and irrelevant features extraction. These problems make the segmentation and classification process difficult. In this research, we proposed an optimized colour feature (OCF) of lesion segmentation and deep convolutional neural network (DCNN)‐based skin lesion classification. A hybrid technique is proposed to remove the artefacts and improve the lesion contrast. Then, colour segmentation technique is presented known as OCFs. The OCF approach is further improved by an existing saliency approach, which is fused by a novel pixel‐based method. A DCNN‐9 model is implemented to extract deep features and fused with OCFs by a novel parallel fusion approach. After this, a normal distribution‐based high‐ranking feature selection technique is utilized to select the most robust features for classification. The suggested method is evaluated on ISBI series (2016, 2017, and 2018) datasets. The experiments are performed in two steps and achieved average segmentation accuracy of more than 90% on selected datasets. Moreover, the achieve classification accuracy of 92.1%, 96.5%, and 85.1%, respectively, on all three datasets shows that the presented method has remarkable performance.