Evaluation of deep learning-based target auto-segmentation for Magnetic Resonance Imaging-guided cervix brachytherapy

The target structures for cervix brachytherapy are segmented by radiation oncologists using imaging and clinical information. At the first fraction, this is performed manually from scratch. For subsequent fractions the first fraction segmentations are rigidly propagated and edited manually. This process is time-consuming while patients wait immobilized. In this work, we evaluate the potential clinical impact of using population-based and patient-specific auto-segmentations as a starting point for target segmentation of the second fraction.

For twenty-eight patients with locally advanced cervical cancer, treated with MRI-guided brachytherapy, auto-segmentations were retrospectively generated for the second fraction image using two approaches: 1) population-based model, 2) patient-specific models fine-tuned on first fraction information. A radiation oncologist manually edited the auto-segmentations to assess model-induced bias. Pairwise geometric and dosimetric comparisons were performed for the automatic, edited and clinical structures. The time spent editing the auto-segmentations was compared to the current clinical workflow.

The edited structures were more similar to the automatic than to the clinical structures. The geometric and dosimetric differences between the edited and the clinical structures were comparable to the inter-observer variability investigated in literature. Editing the auto-segmentations was faster than the manual segmentation performed during our clinical workflow. Patient-specific auto-segmentations required less edits than population-based structures.

Auto-segmentation introduces a bias in the manual delineations but this bias is clinically irrelevant. Auto-segmentation, particularly patient-specific fine-tuning, is a time-saving tool that can improve treatment logistics and therefore reduce patient burden during the second fraction of cervix brachytherapy.