Room: Stars at Night Ballroom 2-3
Purpose: Segmentation of organs at risk (OARs) remains one of the most time-consuming tasks in radiotherapy treatment planning. Atlas-based segmentation have been used with success in brain or head and neck anatomy, but fail in areas like the abdomen where large interpatient variations are present. The goal of this study is to design a segmentation algorithm based on a deep-convolutional neural network to segment 33 organ and tissue structures in the abdomen.
Methods: The algorithm was trained on 56 and tested on 18 CT examinations that had been manually segmented. Organs included liver, kidneys, stomach, spleen, pancreas, small and large bowel, uterus, prostate, fat, skeletal muscle, bone, aorta, vena cava, diaphragm, adrenal glands, renal vasculature, spinal cord, and intervertebral disks. A 3D convolutional neural network based on the U-Net architecture was trained to perform pixel-wise segmentation. The results of this algorithm were compared to manually segmented cases in terms of dice coefficient. The interobserver variability for the organs considered was also estimated.
Results: The Dice similarity coefficients between our model and manually corrected segmentations ranged from 0.95 (liver) to 0.50 (renal arteries). Full Dice scores are reported in the supplemental table. Model performance was similar to work reported in the literature for several organs, and exceeded the accuracy of the best reported model in the literature for pancreas and stomach.
Conclusion: We have demonstrated automated segmentation of multiple organs of the abdomen at an accuracy that approaches manual tracing. To date, this is the largest number of abdominal organs segmented simultaneously. This work shows potential to reduce the time required for radiotherapy treatment planning in a clinical setting.