Image-based evaluation of a commercial AI synthetic CT generator for brain and prostate MR-only radiotherapy.

Abstract

BackgroundSynthetic CTs (sCT) were developed to provide electron density information for MR-only treatment planning. Some AI-based sCT generators have received FDA-clearance and better capture tissue heterogeneity through voxel-wise mappings to improve image guidance and dose calculation.PurposeTo quantitatively characterize the image quality and geometric surrogate agreement of an FDA 510(k) cleared AI-based synthetic CT generator (TheraPanacea) for brain and prostate MR-only radiotherapy, using per-structure volumetric analysis, Hounsfield unit (HU) assessment, and multi-metric geometric evaluation.MethodsPaired CT/MR datasets from 20 brain and 20 prostate patients were retrospectively analyzed. sCTs were generated from T1w (brain) and T2w (pelvis) MR images acquired on the same day with identical immobilization. Auto-contours from TheraPanacea and MIM Contour ProtégéAI were compared against MIM Contour ProtégéAI contours on the reference CT. Bone was segmented via a fixed-threshold method. Evaluation metrics included contour volumes, cross-contour HU comparisons, centroid offsets, overlap and surface-based similarity metrics, and manual bony landmark measurements.ResultsSoft-tissue mean HU differences were ConclusionTheraPanacea sCT demonstrated acceptable soft-tissue HU (± 50 HU) and geometric surrogate agreement (< 2-5 mm) relative to CT, with larger errors in cortical bone, air interfaces, and implants. Site-specific QA must verify bone HU compression, peripheral geometric agreement via landmark measures, and manually review small/high-contrast contours (optic nerves, cochleae, rectum). Institutional image and dosimetric validation, particularly with implanted devices, remains essential prior to clinical deployment.