Click here to

Session: Therapy: AI in Radiation Therapy [Return to Session]

Accurate Automatic Rectum Segmentation for High-Dose Rate Brachytherapy of Cervical Cancer Treatment Planning

Y Gonzalez1,2*, C Shen1,2, K Albuquerque2, X Jia1,2, (1) UT Southwestern Medical Center, Dallas, TX (2) iTORCH Lab, UT Southwestern Medical Center, Dallas, TX


SU-IePD-TRACK 5-2 (Sunday, 7/25/2021) 12:30 PM - 1:00 PM [Eastern Time (GMT-4)]

Purpose: Despite advances in deep-learning (DL) based automatic organ segmentation, automatic rectum segmentation is still challenging in high-dose-rate brachytherapy (HDRBT) of cervical cancer treatment planning, because of a lack of well-defined rectosigmoid boundary. DL models often cannot accurately determine this boundary, causing large errors when evaluating D2cc, as the superior part of rectum falls in the high dose region and 2cc volume is sensitive to segmentation results. This study develops a method to address this problem.

Methods: We first performed an initial segmentation of rectum using a 3D U-Net that mapped a CT image to a rectum volume. We then used the superior-most slice of the initial rectum volume as initial condition to segment the rectum/sigmoid structure in a slice-by-slice fashion for 5cm along the superior direction to ensure rectosigmoid junction was included in the segmented volume. As such, a 2.5D U-Net was established to predict the sigmoid on a given slice based on the connectivity between previous adjacent CT slice and segmentation. We finally calculated the centroid of segmented volume in each slice and determined the centerline of the structure. The final rectum structure was determined by keeping the part inferior to the rectosigmoid junction defined as the place where the structure centerline turned for an angle over a threshold value.

Results: We evaluated our method and a simple 3D U-Net-based rectum segmentation method against ground truth drawn by physicians. The error in superior rectum boundary position was 5.0±5.0 mm in our method, while the error of 3D U-net was 15.7±12.1 mm. When using the segmented structure for D2cc calculation, the error of our method and 3D U-net was 4.8%±3.9% and 9.8%±10.2%.

Conclusion: The proposed method overcame the challenge of conventional DL-based rectum segmentation, which allowed accurate computations of D2cc in HDRBT of cervical cancer treatment planning.



    Not Applicable / None Entered.


    Not Applicable / None Entered.

    Contact Email