EncU-Net: A Modified U-Net for Dermoscopic Image Segmentation

Abstract

In this study, we proposed the EncU-Net model based on U-Net architecture for the segmentation of lesions in dermoscopic images. The EncU-Net model has two encoder sections and one decoder part. The proposed model took two channels (green, L) obtained from two separate color spaces (RGB, LAB) as input images. We gave L channel images to the first encoder section and green channel images to the second encoder section. The model processed them with convolution and down-sampling operations on two separate encoder paths. The feature maps obtained from each block of the first encoder section were concatenated with the second encoder section feature maps at the same level. In the continuation, the segmentation result was obtained through the decoder path. The model was evaluated on the PH2 dataset. According to the accuracy, sensitivity, specificity, dice, Jaccard (IoU) metrics, the model achieved more than 90% success.