Two-Stage Multi-Organ Automatic Segmentation with Low GPU Memory Occupancy

Abstract

Abdominal multi organ segmentation is of great significance in medical diagnosis and research. As the abdominal CT usually has a high resolution and a high image size, automatic segmentation of the abdominal organs demands a high configuration of hardware. In this paper, we proposed a low GPU memory occupied two stage fully supervised automatic segmentation framework for abdomina113 organs: liver, spleen, pancreas, right kidney, left kidney, stomach, gallbladder, esophagus, aorta, inferior vena cava, right adrenal gland, left adrenal gland, and duodenum, and designed a lightweight 3D CNN refer to as Tiny-CED Net. The proposed Tiny-CED Net can accurately complete the automatic segmentation of the whole abdominal CT with the GPU memory occupation <2GB. The results show that the average DSC of our method reached 0.83. The average time consumption and max GPU memory occupied are less than 25s and 2GB.