3D EdgeSegNET: a deep neural network framework for simultaneous edge detection and segmentation of medical images

Deep learning has been a mainstream choice for computer-aided medical diagnosis in recent years. Medical practitioners need accurate and fast diagnosis results to monitor the extent of the disease and devise an efficient treatment plan. This article proposes a deep neural network-based 3D EdgeSegNET architectural framework for simultaneous segmentation and edge detection of brain tumors. It is essential to extract and analyze the critical information about the lesion’s shape and volume from the brain’s magnetic resonance imaging protocol for accurate tumor segmentation. A radiologist keeps a mental map of both edges and segmented regions while performing the brain tumor segmentation task, so shapes and segmented regions are essential parameters of any segmentation task. By recognizing the importance of boundary and area, an automated 3D EdgeSegNET model is devised by combining edge detection and semantic segmentation tasks in a single model architecture with a twofold output of interest of edges and tumor volume. The proposed model achieves more accurate and robust performance on the benchmark dataset provided by Brain Tumor Segmentation Challenge (BraTS 2020) compared to a few top-ranked submissions. The edge detection task obtains an F-measure at optimal dataset scale of 0.7704. An average dice score of 0.89595 and a Hausdorff distance (95th percentile) of 4.375 is achieved on the whole tumor for the semantic segmentation task.