Diagnosis and proper treatment planning and work-flow using CT images for abdominal cavity can be supported by automatic segmentation of abdomen. A limitation of current methods of segmentation using MALF (multi-atlas label fusion) and statistic models is the need of inter-subject image registrations. Image registrations for abdominal cavity is difficult and existing registration-free methods suffer in accuracy. The need is to achieve a segmentation algorithm that does not require registration. Gibson et. al. have developed a segmentation algorithm based on deep learning which does not require registration and can be used in endoscopic procedures for 8 organs: pancreas, esophagus, stomach, duodenum, liver, spleen, left kidney and gallbladder. This dense V-network FCN (DenseVNet) uses feature reuse and memory-efficient dropout to facilitate high resolution activation maps. The computation and memory costs of dropout are lowered using batch-wise spatial dropout scheme. The model has higher accuracy compared to existing state-of-the-art methods.
Input variables : CT image
Output Variables : Segmented organs
Visit Model : github.com
Additional links : ncbi.nlm.nih.gov
| Model Category | : | Public |
| Date Published | : | February, 2018 |
| Healthcare Domain | : |
Medical Technology
Provider |
| Code | : | github.com |
| Health Risk Management |
| Disease Detection |