- This study reports the development and evaluation of two CT-based deep learning models for automated contouring when planning radiation therapy for patients with prostate cancer who receive an absorbable hydrogel rectal spacer
- Model I was trained and validated on 135 patients injected with SpaceOAR and tested on 24 independent patients; model II was trained and validated on the same 135 patients and tested in 64 patients who received SpaceOAR Vue
- The automated contour for the transparent spacer was of little clinical value because the spacer wasn't visible on CT, and the inaccuracy degraded the quality of adjacent regions of interest, mainly the prostate and rectum
- By comparison, testing results confirmed the validity of model II, and editing was unneeded or only slightly needed in 59 cases (92%)
- At Massachusetts General Hospital, use of the opaque spacer and model II has provided standardized contouring, promoted consistency with network hospitals and shortened treatment turnaround time
When prostate cancer is treated with external beam radiation therapy, planning requires accurate contouring of the prostate, seminal vesicles, and surrounding critical organs at risk. Deep learning—a form of artificial intelligence that uses large artificial neural networks to learn patterns in images—is being developed to automate this tedious task.
Subscribe to the latest updates from Urology Advances in Motion
However, the application of deep learning models is limited by certain clinical protocols, particularly the immobilization method, which can affect the shape and position of key organs.
A new minimally invasive immobilization technique, the rectal spacer, is rapidly gaining popularity. Absorbable hydrogel is injected between the prostate and rectum, moving the rectum slightly aside and reducing its exposure to a high radiation dose. The spacer is designed to be absorbed entirely after six to 12 months.
The initial "transparent" version of the spacer, SpaceOAR (Boston Scientific, Marlborough, MA), is tissue-equivalent without contrast enhancement. An updated "opaque" version, SpaceOAR Vue, includes iodinated contrast.
Massachusetts General Hospital began using the transparent spacer in 2017 and the opaque version in 2020, so it has accumulated enough data to develop new deep-learning auto-segmentation models. Yi Wang, PhD, head of the Laboratory of Artificial Intelligence in the Department of Radiation Oncology, Jason A. Efstathiou, MD, DPhil, director of the Genitourinary Service in the Department and vice-chair for Faculty and Academic Affairs, and colleagues describe their value in Medical Physics.
The models were based on a commercial product, the Deep Learning Contouring Expert (Mirada Medical Ltd., Oxford, UK):
- Model I, developed for patients with the transparent spacer, was trained on data from 125 patients, validated on ten cases, and tested on 24 independent cases. The spacer was contoured on MRI and the prostate was contoured on CT with MRI guidance. The remaining regions of interest (ROIs) were drawn on CT without MRI guidance
- Model II—Later, Mass General switched to the opaque spacer. The team developed a new model by reusing the 135 training and validation cases from model I, without any transfer learning but with refinements provided by Mirada. This model was tested on 64 patients who underwent only CT
The automated contour for each ROI was assessed against the manual contour submitted by one of three genitourinary radiation oncologists during treatment planning. The assessor assigned 1 point for no or minor discrepancy; 2 for moderate discrepancy; 3 for major discrepancy; 4 for unacceptable.
The geometric differences were also evaluated using the dice similarity coefficient (DSC, the spatial overlap between the automatic and manual sets, ranging from 0–1) and the mean distance to agreement (MDA, the average distance between the two outlines on each slice). The thresholds recommended by the American Association of Physicists in Medicine are DSC ≥∼0.8 to 0.9 and MDA ≤∼2 to 3 mm.
The journal article provides results for each ROI. In summary:
- Method I had little clinical value because the transparent spacer was not visible on CT (mean score on the four-point scale, 3.63; DSC, 0.52; MDA, 2.9 mm). The inaccuracy degraded the quality of adjacent ROIs, mainly the prostate and rectum
- Method II—Test results confirmed the validity of method II (mean score, 1.30; DSC, 0.84; MDA, 0.9 mm). The mean score was 1 or 2 in 59 cases (92%)
The Mass General Experience
Mass General has phased out the transparent spacer except for patients with contrast allergy. Clinics still using it are advised to replace the spacer auto-contour with the manual contour drawn on MRI while fused to simulation CT, then use Boolean functions to fix adjacent ROIs quickly.
Model II has been used at Mass General and three network hospitals since January 2021. It provides standardized contouring, promotes consistency across the network, and can shorten treatment turnaround time. Model II can be applied to cases where patients do not receive a spacer or rectal balloon.
Learn about the Genitourinary Program at Mass General
Refer a patient to the Department of Radiation Oncology