RecU-Net++: Improved Utilization of Receptive Fields in U-Net++ for Skin Lesion Segmentation

Recently skin cancer has emerged as one of the most threatening diseases with an alarming rate of fatality. Though the early detection of this disease is quite useful for proper treatment planning, the task is quite intimidating. Several factors such as obscure lesion boundaries, variable contrast, the difference in lesion shape, size and colours and the presence of undesirable artefacts (such as hair) make the segmentation task extremely challenging. To address those challenges and to augment the efforts of clinicians, in this paper we propose RecU-Net++, a novel segmentation framework that utilizes multiple Receptive Field Block (RFB) to extract spatial context information at multiple scales to preserve discriminative features of the lesion images. RecU-Net++ also introduces an extremely pliable feature fusion scheme by agglomerating features of multiple semantic scales with a redesigned skip connection. We also introduce dense blocks after every downsampling operation for effective reuse of features and strengthening feature flow, thus mitigating the problem of exploding gradient. Our proposed method is evaluated on two publicly available skin lesion segmentation datasets: HAM10000 and ISIC2017. Experimental results show that our proposed model outperforms the existing segmentation methods quite significantly.