2065 - Back to the Past: Challenges and Pitfalls of Manual vs. Automatic MR-CT Registration in Stereotactic Radiosurgery for Brain Metastases
Presenter(s)
V. Faccenda1, D. Panizza1,2, R. R. Colciago2, S. A. Mascellani2, B. Bordigoni1, S. Carminati1, C. Ingraito1, S. Arcangeli2, and E. De Ponti1; 1Fondazione IRCCS San Gerardo dei Tintori - Medical Physics, Monza, Italy, 2University of Milan Bicocca - School of Medicine and Surgery, Milan, Italy
Purpose/Objective(s): To investigate the accuracy and quantify uncertainties of manual vs. automatic MR-CT registration in stereotactic radiosurgery (SRS) for brain metastases (BM).
Materials/Methods: Twenty MR-CT rigid registrations were independently performed twice by three expert operators. A reference fusion was derived by averaging the translational and rotational degrees of freedom across all operators. Thirty-nine BM volumes contoured on the MR images were mapped onto the corresponding CT scans based on each fusion to define a mean reference barycenter. To evaluate inter- and intra-operator variability, the mean absolute errors (MAE) of the deviations and barycenter shifts were calculated. The same metrics were also used to compare the reference with the outputs of automatic fusions performed by commercially available software using three different algorithms (standard, box-based, and contour-based) and with the registrations performed by four non-experts.
Results: Inter-operator variability was adequate to establish the mean reference (agreement within 0.4° and 0.3mm), although intra-operator differences of up to 1.0° and 1.1mm, resulting in a mean barycenter displacement of 0.8 mm (range:0.1–2.4), were observed. The contour-based algorithm (focused on the brain structure) exhibited the closest agreement with manual values, with mean translation MAE of 0.2mm (P=0.466) and rotational MAE of 0.2° (P=0.731). The mean BM centroid shift was 0.7 mm (P=0.089), with no statistically significant differences observed between brain structures delineated manually or automatically (P=0.395). Non-expert registrations exhibited greater discrepancies, with mean translational MAE, rotational MAE, and barycenter shift of 1.0mm, 1.0°, and 2.0mm, respectively.
Conclusion: Inter- and intra-operator variability, along with operator experience, must be accounted for when planning high precision SRS treatments. Even experts showed barycenter deviations exceeding 1 mm, prompting reconsideration of a “fusion gold-standard” and the use of sub-millimeter margins. Automated algorithms, especially contour-based approaches, showed promise in reducing dependence on operator expertise and improving efficiency in a clinical workflow.