Fully Automated Multi-Organ Segmentation in CT Images via Deep Neural Networks

Organ segmentation is essential in medical imaging applications in radiation therapy, diagnostic organ dosimetry, and computational quantitative calculations such as radiomics studies and theranostic. Manual segmentation of clinical data is labor-intensive and time-consuming, besides the inter-observer differences remain a problem. In this study, we tried to do the automatic segmentation of multiple organs from CT images. We collected CT images and organ segmentations for the Liver (157 cases), Pancreas (390 cases), Kidneys (310 cases), Spleen (70 cases), Bladder (112 cases), Aorta (30 cases), and Gallbladder (295 cases). After pre-processing, including resizing and rescaling, data were split into train/validation (80%) and test (20%) groups. A residual deep neural network was trained to segment images. The performance was evaluated by the Dice and Jaccard Coefficient of the output compared with the ground truth for each organ. The average dice values were 0.96±0.02, 0.78±0.2, 0.90±0.19, 0.94 ±0.08, 0.85±0.16, 0.85±0.06, and 0.79±0.2 for the Liver, Pancreas, Kidney, Spleen, Bladder, Aorta, and Gallbladder, respectively. Also, the Jaccard values were 0.92 ± 0.02, 0.65 ± 0.20, 0.84 ± 0.19, 0.90 ± 0.08, 0.77 ± 0.18, 0.75 ± 0.08, 0.69 ± 0.21 for the mentioned organs, respectively. The best dice were for the liver, followed by the spleen, while the weakest performance was shown in pancreas and gall bladder segmentation. The performance of our network was comparable with the recent studies in this field, while we included the data from multiple sources in the test set. This fully automatic segmentation pipeline can decrease the treatment planning time in radiation therapy and make the dosimetry in theranostic feasible. Also, it can be helpful to generate computational phantoms by anchor organ segmentation.