KEY POINTS
- This retrospective evaluation assessed an FDA 510(k)-cleared TheraPanacea AI synthetic CT generator using same-day CT/MRI datasets. The quantitative cohort comprised 20 brain and 20 prostate patients, using T1-weighted brain and T2-weighted pelvic MRI inputs.
- Real-world interoperability was itself a failure mode: 5/25 brain and 2/28 prostate projects initially failed generation because of minor DICOM-header incompatibilities. A further six prostate cases containing brachytherapy seeds were excluded from quantitative analysis because the implants lay within the evaluated target region.
- Mean soft-tissue HU differences were generally <50 HU, while central homogeneous ROIs were usually within 20 HU of reference CT. Larger discrepancies appeared in heterogeneous bone and air-containing regions; femoral-head and sacral differences reached approximately 347 ±14 HU and 455 ±50 HU, respectively.
- Geometric surrogate agreement was generally small: contour centroids differed by <2 mm, pelvic centroid differences were all <1.5 mm, and manually measured skeletal dimensions differed by no more than 0.66 mm in the brain and 0.32 mm at the femoral heads. Mean distance-to-agreement remained below 5 mm for evaluated structures.
- Small cranial structures remained problematic. Optic nerves and cochleae were sometimes partially or completely missed by automated segmentation, while pelvic bladder, penile bulb and seminal-vesicle contours also showed poorer overlap than large rigid structures; this supports manual review rather than direct clinical acceptance of generated contours.
- Implant representation was a major limitation: dental implants produced MR-like susceptibility artifacts rather than CT streaking, with approximately 9,000 HU mean underestimation within implant ROIs. Prostate seeds were effectively absent from the synthetic CT, with maximum HU discrepancies approaching 9,500 HU.
- The study deliberately evaluated image and geometric performance rather than recalculated treatment dose. The authors therefore recommend site-specific commissioning, dedicated assessment of bone and high-density materials, and institutional dosimetric validation before clinical deployment.
CLINICAL TAKEAWAY
The commercial AI synthetic CT performed convincingly for bulk soft tissue and global geometry, but those are not the areas most likely to break an MR-only workflow. Bone, air interfaces, small OARs, implants and even DICOM interoperability need explicit local QA, and this study should not be interpreted as a standalone dosimetric commissioning study.