A Novel Patch-Based Ensembling Approach with Perceptual Attention for Skin Lesion Classification

Most of the time biopsy has been the gold standard for skin lesion evaluation. However, specialists evaluate signs and symptoms for the final decision. Shortage of specialist definitely adds the adverse effect on effective and early detection. Recently, CNN has extended the helping hand for the specialist during the final decision. Also, many pre-trained CNN models have been designed to be used as transfer learning. But, a common approach of random resizing of input images are required before training to get fit to the input layers. This is because the approximate size of most of the available skin lesion images and pre-trained models are of 1000×1000 and 224×224 respectively. Hence the required resizing though solves one problem of size mismatch, it may eliminate principal feature for classification leading to poor accuracy. Hence, in this work, we propose a novel patch-based ensembling approach for the early diagnosis of melanoma and nevus skin lesions. Here the effect of applying patches over classification has been studied on an incremental basis. In the ensembling approach, the resultant features from different patches have been combined for further processing with perceptual attention to maintain the spatial relationship. The proposed model was evaluated on a set of 748 dermoscopy images collected from the ISIC 2017 data set (374 melanoma and 374 nevus images). Our result demonstrates that using image patches as input improves accuracy instead of image scaling. The proposed model performed well enough to serve as a baseline for further studies of sin lesion classification.