2560 - Development of a Brain Sct Hallucination Screening Workflow In Mim for MR-Only Radiotherapy Quality Assurance

Purpose/Objective(s): Deep learning (DL)-generated synthetic CT (sCT) images are increasingly used in MR-only radiotherapy workflows to provide electron density and reference anatomy. However, DL-generated sCT images may contain hallucinations or anatomic misassignments that may affect dose calculation accuracy and image guidance. We developed and implemented a brain-specific hallucination screener to identify suspicious bone-related discrepancies in AI-generated brain sCT images for clinical quality assurance.

Materials/Methods: A brain bone DL auto-contouring model was developed using MR brain data from 150 patients to generate MRI-based bone contours within MIM. For hallucination screening, a second bone representation was generated directly from the sCT using whole-volume threshold-based contouring (HU > 150). A MIM screening extension compared the sCT threshold-derived bone contour with the MRI DL bone contour and generated a flagged contour representing discordant regions for visual review. The workflow was qualitatively evaluated in 5 brain patient cases using multiplanar review (axial, sagittal, coronal) to assess whether flagged regions corresponded to clinically meaningful suspicious anatomy versus false-positive detections.

Results: The screening workflow was successfully implemented and executed in MIM, producing reviewable flagged regions across all tested cases. The screener consistently highlighted discordance in anatomically complex regions, most commonly the paranasal/ethmoid sinus region, nasal cavity/septal region, and central skull base/midface region (including adjacent sinonasal interfaces), which are clinically relevant areas of sCT uncertainty due to thin bony structures and air-bone transitions. In several cases, flagged regions were focused and plausible, supporting the utility of the approach. False-positive flagging was also observed, primarily along skull boundary edges (calvarial contour mismatch/thickness differences) and, in some cases, regions outside the cranial anatomy (e.g., lower neck/inferior image areas). These out-of-region flags likely reflect limited training contour coverage in inferior non-cranial anatomy.

Conclusion: We developed and integrated a brain sCT hallucination screening workflow in MIM that compares sCT threshold-derived bone contours (HU > 150) with MRI-based deep-learning bone contours to generate discordance flags for QA review. In a pilot evaluation of 5 brain patients, the workflow demonstrated feasibility and clinical relevance, while also highlighting predictable false-positive patterns at contour boundaries and out-of-region anatomy. Future work will focus on improving specificity through retraining with expanded contour labeling coverage and/or anatomy-restricted processing.