MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

A Deep-Learning Approach Toward Automated Digitization of High Dose-Rate Brachytherapy Applicators

H Jung1*, P Klages2 , C Shen3 , Y Gonzalez4 , K Albuquerque5 , X Jia6 , (1) University of Texas Southwestern Medical Center, Dallas, TX, (2) Memorial Sloan-Kettering Cancer Center, New York, NY, (3) University of Texas Southwestern Medical Center, Dallas, TX, (4) University of Texas Southwestern Medical Center, Dallas, TX, (5) University of Texas Southwestren Medical Center, Dallas, Tx, (6) University of Texas Southwestern Medical Center, Dallas, TX

Presentations

(Wednesday, 8/1/2018) 7:30 AM - 9:30 AM

Room: Davidson Ballroom A

Purpose: Applicator digitization is one of the most critical steps in high dose-rate brachytherapy (HDRBT) treatment planning, as it affects accuracy of source positions and therefore dosimetric accuracy due to sharp dose gradient. Current standard method is manual digitization. Not only is this labor-insensitive, the accuracy is subject to human experience and available time. In this work, we develop a tool to automatically digitize different HDRBT applicators in CT images using deep-learning techniques.

Methods: Our tool employed a two-step process. It first used U-net to segment applicator regions. We trained the U-net using 95000 two-dimensional CT images (including data augmentation) of HDRBT cases with a tandem-and-ovoid (T/O) applicator. The second step applied a polynomial curve fitting method to extract applicator central lines based on the segmentation results. Applicator tip position is determined as the point with the highest CT-number gradient along the channel. We evaluated our tool using five T/O cases that were not used in network training. To evaluate robustness of our tool, we also applied it to Y-tandem and cylinder-applicator cases, and T/O applicator cases scanned in cone beam CT (CBCT) with a low image quality.

Results: In test cases with a T/O applicator, average Dice similarity coefficient between automatic and manual segmented applicator regions was 0.93. Average distance between tip positions determined by our tool and manually was 0.63 mm and average Hausdorff distance (HD) between applicator channels determined by our tool and manually was 0.68 mm. Computation time was 5 secs per patient. Although trained in T/O applicator cases, our tool can also digitize Y-tandem, cylinder applicator, and T/O applicator scanned in CBCT. Accuracy of tip position and HD were < 1 mm.

Conclusion: We developed a deep-learning based applicator digitization tool for HDRBT. The achieved accuracy, efficiency, and robustness made our tool clinically attractive.

Keywords

Brachytherapy, Treatment Planning, Segmentation

Taxonomy

TH- Brachytherapy: Dose optimization and planning

Contact Email