Session Item

Improvements in radiotherapy techniques and precision has meant there is more demand for delineation accuracy locally. Target volume (TV) delineation is subject to inter-observer variability (IOV) and is a major contributor to uncertainty in radiotherapy planning. Delineation for the TVs and organs at risk (OARs) is a time-consuming process and is primarily done manually. Automated methods can be more efficient and reduce IOV, but its use within radiotherapy is limited. This project aimed to produce an auto-segmentation model to delineate the TVs and OARs for patients receiving radiotherapy to the prostate and evaluate the efficiency and accuracy of using automated delineation methods in the workflow.

The model was first created with 150 anonymised prostate datasets.

The OARs and TV were manually delineated on 20 datasets by five radiographers and three clinicians, and the timings were manually recorded. The model was loaded onto the same datasets, and the OARs on 5 cases were manually modified; by 1 radiographer; as required, and the time manually recorded for comparison. Seven radiographers scored OARs for 20 cases using a 1-5 scale; one being completely acceptable, and five completely unacceptable. Comparison metrics, such as DSC were used to evaluate statistical significance, comparing the auto structures to the participants’ and to compare the participants.

The TVs generated were not acceptable for use in the first version of the model and required additional time to be trained, so they were excluded in the full analysis. Six hundred seventy structures were manually delineated, and 628 times were manually recorded. The overall mean time to complete all structures for a prostate and node patient is 80.4 minutes (33.8 minutes for radiographers and 46.5 minutes for clinicians). The mean time for a radiographer to edit 5 cases was 21.0 minutes, a time-saving of 12.8 minutes. There is a 79% agreement between clinicians, and 86% agreement between the radiographers, and an 83% agreement between the auto and the radiographers. The median scores were 2 for the femoral heads and 3 for the bladder, bowel, and rectum, with two requiring insignificant modification and three minor modification.

A significant time-saving can be made by using auto-segmentation methods to delineate the structures for prostate patients. Although the model requires refinement, as a first version, it does offer a good starting point for OARs. Training is still required to develop the model so that the TVs are more acceptable, to see a more significant time-saving. Although the model could not create structures that do not require at least some manual editing, a time saving was seen for OARs in this study. The magnitude of this saving is currently limited by the requirement for practitioners to check and modify the structures to ensure they are acceptable for planning.