A semi-automated workflow for cohort-wise preparation of radiotherapy data for dose-response modeling, including autosegmentation of organs at risk.

BACKGROUND: Preparing retrospective dose data for risk modeling using large study cohorts can be time consuming as it often requires patient-wise manual interventions. This is especially the case when considering organs at risk (OARs) not systematically delineated historically. Therefore, we aimed to develop and test a semi-automated workflow for cohort-wise preparation of radiotherapy data from the oncology information system (OIS), including OAR autosegmentation, for risk modeling purposes. METHODS: A semi-automated workflow, including cohort-wise data extraction from a clinical OIS, cleanup, autosegmentation, quality controls (QCs), and data injection into a research OIS was iteratively developed using 106 patient cases. We evaluated two deep learning (DL)-based methods and compared with four atlas-based methods for autosegmentation of the proximal bronchial tree (PBT), the heart, and the esophagus that were possible to integrate into the workflow. One method was an in-house DL-based model using OARs manually contoured by experts for 100 cases. Geometric and dosimetric agreements with manually contoured OARs were evaluated for 20 independent cases. The final workflow was tested on 50 independent cases. RESULTS: The DL-based methods were better than the atlas-based at segmenting the PBT (mean Dice similarity coefficient (DSC) 0.81-0.83 versus 0.59-0.80) and the esophagus (mean DSC 0.76-0.77 versus 0.39-0.46). The methods performed similarly for the heart (mean DSC 0.90-0.95 (DL-based) and 0.84-0.90 (atlas-based)). Our in-house autosegmentation model had the highest mean DSC for all OARs. The final version of the workflow successfully prepared data for 80% of the test cases without case-specific manual interventions. CONCLUSIONS: The semi-automated workflow enabled efficient cohort-wise preparation of OIS data for risk modeling purposes. Our in-house DL-based segmentation model outperformed the other methods.