Click here to

Session: AI-Based Auto-Segmentation and Auto-Contouring - I [Return to Session]

Multi-Organ Auto-Segmentation of Abdominal Structures From Contrast-Enhanced and Non-Contrast-Enhanced CT Images

C Yu1,2*, C Anakwenze2, Y Zhao1,2, R Martin2, J Niedzielski5, E Ludmir2, J Yang1,2, L Court1,2, C Cardenas1,2, (1) University of Texas MD Anderson UTHealth Graduate School of Biomedical Sciences, (2) University of Texas MD Anderson Cancer Center, Houston, TX


TU-E-TRACK 6-4 (Tuesday, 7/27/2021) 3:30 PM - 4:30 PM [Eastern Time (GMT-4)]

Purpose: To develop a deep-learning-based tool for accurate and robust upper abdominal organs-at-risk(OAR) autosegmentation suitable for treatment planning and adaptive replanning.

Methods: Forty pancreas SBRT patients with contrast-enhanced breath-hold planning CT images were selected. All OARs (duodenum, small bowel, large bowel, stomach, liver, spleen, left kidney, right kidney and spinal cord) were edited or recontoured under physician supervision to ensure accuracy and consistency. We used the self-adapting nnUNet framework to customize a 3D U-Net for our data to predict all organs concurrently. We tested our model using five-fold-cross-validation and calculated Dice-similarity-coefficients(DSC) and 95th-percentile Hausdorff(HD95) distances. A radiation oncologist assessed the performance of the tool on 8 randomly selected patients. This assessment was then repeated on non-contrast-enhanced daily CT images of the same patients.

Results: The mean dice-similarity-coefficient/95th-percentile Hausdorff distance between the automatic segmentation and the ground truth on contrast-enhanced CT images were 0.88±0.07/8.94±10.29mm for small bowel, 0.89±0.08/12.46±15.09mm for large bowel, 0.80±0.08/12.29±8.10mm for duodenum, 0.91±0.04/5.98±4.59mm for stomach, 0.96±0.02/5.43±9.93mm for liver, 0.97±0.01/2.28±2.09mm for spleen, 0.95±0.05/2.78±2.02mm for left kidney and 0.96±0.02/2.46±1.42mm for right kidney. For clinical performance evaluation, all contours on contrast-enhanced planning CT images and non-contrast-enhanced daily CT images were clinically acceptable. The radiation oncologist suggested minor edits for the following OARs: large bowel (1 patient), small bowel (1 patient), liver (1 patient), and duodenum (7 patients). All other OARs were rated as “use-as-is”. For non-contrast-enhanced daily CT images, large bowel (2 patients), small bowel (2 patients), liver (1 patient), stomach (1 patient), and duodenum (7 patients) required minor edits. All other OARs were rated as “use-as-is”.

Conclusion: We developed an auto-contouring tool for upper abdominal OAR segmentation for both contrast-enhanced and non-contrast-enhanced CT images. Our tool achieved a high-level of clinical acceptability on both types of images, which can provide accurate and consistent contours for treatment planning and adaptive replanning.



    Not Applicable / None Entered.


    IM/TH- image Segmentation: CT

    Contact Email