Purpose: Evaluate the accuracy of organ-at-risk (OAR) autosegmentation by a deep-learning model (DLM) for head-and-neck (H&N) radiotherapy and assess need for revision of contours before treatment planning.
Methods: H&N planning CTs (pCTs) were retrospectively identified, and a new set of OAR contours was generated for each pCT by experienced radiation oncologists (ROs) to form a database of highly-curated, institutional “gold-standard” contours (GS-OARs) consistent with international consensus guidelines. 312, 51, and 82 GS-OAR sets were used in training, validation, and testing of a DLM (custom 3D U-net architecture). The DLM operates on 32x512x512-voxel subvolumes and generates 42 binary H&N-OAR labels for each voxel which are subsequently vectorized (DLM-OARs). It was trained on a TPUv3 pod with spatial partitioning using a hybrid loss function that automatically accounts for large variability in OAR sizes. For 19 additional “hold-out” pCTs, both GS-OARs and DLM-OARs were generated; DLM-OARs were subsequently reviewed by ROs who had not been involved with GS-curation and revised to meet the RO’s expectations for acceptability for planning. DLM-OARs from before and after RO editing (PRE-OARs and POST-OARs, respectively) were compared to each other and GS-OARs via Dice Similarity Coefficient (DSC). Using the PRE-OARs and target volumes from the clinical treatment plan, a viable treatment plan was created under supervision of two expert ROs; resultant dose was the basis for dose-volume statistical comparisons and an associated plan quality metric (PQM) was computed for each contour-set.
Results: Mean DSC relative to GS-OARs was greater than 0.8 for 37 (88%) PRE-OARs and 36 (86%) POST-OARs, and greater than 0.9 for all OARs comparing PRE-OARs to POST-OARs. The mean PQM change comparing PRE-OAR to POST-OAR dose-volumes was 0.8%.
Conclusion: DLM-autosegmented contours agreed well with institutional gold standards and has the potential to be used for clinical treatment planning with only minor or no revisions.