Deep Learning-Based Automated Quality Assurance for Palliative Spinal Treatment Planning in Radiotherapy

Abstract

Listen

Translate article

Robustquality assurance (QA) for palliative spine radiation therapy (RT) remains critical due to the risk of wrong anatomic level treatment on account of human error in enumerating vertebral bodies accurately based on morphology with incomplete imaging of the spine and prevalence of anatomic variants (10%). We propose a rapid, fully automated deep learning-based QA (DL-QA) tool for segmenting and enumerating vertebral structures from image data, capable of identifying misalignment based on discrepancies in calculated dose coverage. Ina retrospective cohort of 514 patients who received palliative spine radiation treatment at a single institution for spinal metastases, vertebral volumes for each individual spine level were automatically segmented on RT planning computed tomography scans using a publicly available deep learning algorithm, Total Segmentator (TS) deployed in the treatment planning system (Wasserthal et al, 2022). Departmental policy requires that the prescription/plan name include all spinal levels that receive a prescribed dosimetric threshold of V50% > 50%. By comparing the intended spine level target in the prescription and plan name against the TS volumes, the DL-QA flagged all cases for which any target vertebrae did not receive this threshold dose and/or any non-targeted vertebrae that received V50% > 50%. To detect spine anomalies, cases were also flagged if any vertebrae volume was not within ±1σ of the entire population of vertebrae volumes. Flagged cases were either categorized as: (1) wrong spine level RT error; (2) documentation error, in which treatment was correct but the prescription/plan name did not follow Departmental policy; or (3) potential spinal geometric error. All flagged cases were verified manually by checking the original images and treatment planning data. Outof 514 patients, 29 cases were flagged as potential errors. Manual review revealed that one of these was a previously discovered true treatment error (due to anatomic variant with 4 lumbar bodies) while 10 were treated as intended but showed documentation errors due to variants in the number of vertebral bodies, kyphosis of the spine causing non-targeted vertebrae to appear in the treatment field, or improper observation of the Departmental plan naming policy. The remaining 18 cases were associated with flagged vertebrae volumes. Reviewing those patients, we identified spinal anomalies where TS attempted to account for extra or missing vertebrae (N = 9) and cases where TS made segmentation errors (N = 9). Theproposed automated DL-QA system successfully identified patients with spine anomalies, flagged documentation errors, verified the correct target levels of spine RT treatments, and detected a known misadministration. The next phase will involve prospective testing of the system in a clinical setting upstream of treatment delivery.