Deep retinanet for melanoma lesion detection

Abstract

Ever since the automation of melanoma detection, there is a huge challenge pertaining to irregularity in shape, size, location and color of dermoscopy images. Moreover, melanoma treatment seems a complicated task owing to inadequate details for diagnosis and limited visual inspection. Therefore, an auto-mated process of detection is required in dermoscopic images for efficient and timely detection and diagnosis of melanoma lesion. Consequently, we have localized melanoma using one stage object detector named RetinaNet. The proposed model is evaluated by conducting experiments on PH2 dataset. RetinaNet serves a single step object detector that efficiently and precisely detects melanoma region. Moreover, focal loss is also evaluated to avoid class imbalance between normal skin pixels and melanoma foreground segmentation. The proposed system showed a significant performance gain up-to 97% i.e. the average precision using PH2 sample images. Our system can be effectively utilized in automation of clinical decision support systems for practical diagnosis and prognosis of melanoma.