Assessing the Dosimetric Accuracy and Time Savings of MR-Based Deep Learning Auto-Segmentation Implemented In a Gynecological High Dose Rate (HDR) Brachytherapy Clinical Setting
Abstract
Purpose
Gynecological high dose rate (HDR) brachytherapy patients receiving treatment for cervical cancers are under anesthesia during planning and treatment. Minimizing planning time without compromising safety and effectiveness would improve patient care, improve planner workload, and reduce the burden on hospital resources. In this study, we assess the time savings and accuracy of a deep learning auto-segmentation (DLAS) model when applied to daily MRI for clinical planning.
Methods
A single-institution DLAS model trained on 61 T2 Space MR images with U-net, and tandem and ring (TR) applicator in situ was implemented in the clinic. This model was used to contour organs at risk (OARs), the bladder, rectum, sigmoid, and small bowel, on daily MRIs from eight patients (with Tandem and Cylinder (TC), Advanced Gynecological Applicator (AGA), and AGA with needles cervical brachytherapy applicators in situ). The DLAS and edited DLAS contours (e-DLAS) used for planning were compared using contour volumes and dosimetric criteria (D0.1cc, D1.0cc, and D2.0cc). Editing times of the e-DLAS were compared with clinical contouring times acquired prior to MR-DLAS implementation.
Results
Mean volume differences (DLAS – e-DLAS volume), for bladder, rectum, sigmoid, and small bowel were (-2.31±12.77, -1.69±11.74, -19.38±26.19, -7.87±48.64cc), respectively. Mean dosimetric differences in bladder, rectum, sigmoid, and small bowel for D0.1cc (-8.43±9.06, 16.93±47.16, -1.16±88.49, 13.19±151.11cGy), D1.0cc (-6.82±9.65, 9.58±35.60, -14.44±61.11, 26.11±104.19cGy), and D2.0cc (-23.94±51.64, 2.78±28.68, -30.04±66.77, 25.74±89.19cGy), respectively. Segmentation time was <1 minute for all cases. The average time for e-DLAS contours was 30.1±12.5 minutes, with a range of 13.5–47.5, versus 52.5 minutes and a range of 15–110 before the MR-DLAS implementation.
Conclusion
This study demonstrates the potential of MR-DLAS for gynecological HDR brachytherapy planning. While the introduction of MR-DLAS expedited and standardized the contouring part of the planning process, it was limited by atypical anatomy, as seen with longer editing times.