Authors - Isaac Baffour Senkyire, Benjamin Ghansah, Emmanuel Freeman Abstract - With the rapid development of deep learning, CNN-based medical im-age segmentation algorithms have been successful. However, study on the pancreas in 3D CT and MRI images is limited due to the excess use of computer memory and the complexity of the pancreas. In this paper, we present a memory-efficient cascaded 3D network for pancreas segmentation in CT and MRI. We develop a novel Lightweight 3D Bond (L3D-Bond) Layer to reduce filter size, and maintain performance while lowering memory usage, and a novel Light-weight 3D Asymmetric Corollary Atrous Spatial Pyramid Pooling Module (L3D-aCASPP) that captures multi-scale 3D context with lower computational cost. Our experiments were done using the public NIH pancreas segmentation dataset, MRI pancreas segmentation dataset, and MSD spleen segmentation dataset achieving a competitive segmentation performance of 80.12 DSC on the NIH dataset with parameters less than 0.5 million.
Open Zoom