Semantic Segmentation of Tumors in Kidneys using Attention U-Net Models

Abstract

Accurate segmentation of tumors in kidneys (renal) will assist clinical experts to identify the occurrence of cancer. Physical kidney segmentation on CT images is tedious and varies between individual professionals due to its diverseness. Therefore, deep convolutional neural networks are widely used in renal segmentation tasks to aid in the early detection of renal cancer. Using Attention U-Net architecture, we propose an automated technique for delineating the kidneys and tumor in computed tomography (CT) images. Attention U-Net models place a greater emphasis on regions of interest (kidneys and tumors) and less emphasis on areas that are not in focus. With 19 pre-exercised model segments, the Attention U-Net Architecture is utilized to segment kidney tumors (KiTS 2019). To improve the kidney and tumor IOU scores we ultimately ensemble six Top Models.