MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

A Deep-Learning Module in AutoBrachy for Organs-At-Risk Segmentation in High-Dose-Rate Brachytherapy of Gynecological Cancer

Y Gonzalez1,2,3*, C Shen1,2,3 , K Albuquerque3 , X Jia1,2,3 , (1) innovative Technology Of Radiotherapy Computations and Hardware (iTORCH) Laboratory, (2) Medical Artificial Intelligence and Automation (MAIA) Laboratory, (3) University of Texas Southwestern Medical Center, Dallas, TX

Presentations

(Thursday, 7/18/2019) 7:30 AM - 9:30 AM

Room: 301

Purpose: To automate the planning process of high-dose-rate brachytherapy (HDRBT) of gynecological cancer to improve efficiency and ensure plan quality, we have developed and implemented the AutoBrachy system at our institution. By automating the process from applicator digitization to determining dwell time, AutoBrachy has reduced the time from ~40min of manual approach to ~3min. We realized the remaining bottleneck is organ segmentation. Hence, we developed a deep-learning based segmentation module in AutoBrachy to segment major organs-at-risk (OARs) (bladder, rectum, and sigmoid colon).

Methods: We employed two 3D U-Nets for bladder and rectum segmentation, which map a patient CT volume to binary organ masks. Simple 3D U-Net cannot accurately segment sigmoid due to its complex topology and large variation in shape, size, and filling status. We developed an iterative 2D U-Net that maps a CT slice, known sigmoid contour in it, and an adjacent CT slice to the sigmoid contour in the adjacent slice. With an initial condition of the CT slice at the top of rectum and the known rectum contour, the U-Net iteratively predicted sigmoid contours slice by slice. Since the sigmoid may curve multiple times along superior-inferior direction, we repeatedly applied the U-Net multiple sweeps in different directions to segment the entire organ. Known bladder and rectum region were subtracted from the sigmoid contour. We trained the networks using 126 patient cases with 7 for validation. We evaluated the performance in 10 patients not seen in training using Dice similarity coefficient (DSC) with respect to manual segmentation results.

Results: 3D U-Net achieved DSCs of 0.91 and 0.85 for bladder and rectum. Iterative 2D U-Net achieved DSC of 0.84 for sigmoid. It took ~4min to complete segmentation of one patient case.

Conclusion: The deep-learning based segmentation tool in AutoBrachy can achieve accurate segmentation of major OARs in HDRBT.

Funding Support, Disclosures, and Conflict of Interest: Supported by NIH R37-CA214639S1

Keywords

Not Applicable / None Entered.

Taxonomy

Not Applicable / None Entered.

Contact Email