EVALUATING THE TIME SAVINGS OF TARGET VOLUME AUTO-CONTOURING ASSISTANCE IN CERVICAL CANCER HDR BRACHYTHERAPY

Purpose:

To assess the impact on contouring efficiency when an in-house machine learning (ML) model provides a starting point for the high-risk clinical target volume (HR-CTV) definition in high-dose-rate (HDR) cervical brachytherapy.

Materials and Methods:

T2-weighted MRIs with HR-CTV contours from patients receiving HDR cervical brachytherapy between 2016 and 2024 were used to develop and test an ML model. The model was built using PyTorch and architectures from the Medical Open Network for AI (MONAI). The final model was used to generate an HR-CTV contour on previously unseen MRIs, serving as a starting point for the radiation oncologist to edit until a clinically acceptable contour was achieved. Efficiency was assessed by having four radiation oncologists individually contour the HR-CTV offline, with and without model support, two months apart. Contouring time for both scenarios was compared to quantify the model’s impact on efficiency. The quality of the contour made with model support was assessed using the Sorensen-Dice similarity coefficient (DSC) against the same radiation oncologist’s contour without model support.

Results:

The retrospective dataset for model development included 103 patients (151 MRIs) and the testing dataset consisted of 5 patients (11 MRIs). During development and testing, the model achieved an average DSC of 0.75 and 0.70, respectively, when compared to the clinical contours used for brachytherapy. Contouring time with and without model support in the testing set was 5.1±2.7 and 8.7±4.5 minutes, respectively (p<0.01), corresponding to a 3.6-minute absolute reduction, or a 38% decrease in contouring time with model support. The average DSC between the final contours made with and without support was 0.77±0.07.

Conclusions:

Target volume auto-contouring assistance with an ML model reduced the average time spent contouring the HR-CTV by 38% while maintaining contour quality. Future work will include a prospective study to validate the efficiency of this ML model in a real-time clinical setting.