2564 - Impact of MRI Protocols and Tumor Volume on the Performance of AI-Driven Segmentation in GK-Radiosurgery
Presenter(s)
S. Oh1, R. Ljungqvist2, B. Valcu2, A. Heermann1, K. Renick3, H. Lee1, T. Mitchell1, S. Ferguson1, S. M. Goddu1, N. Knutson1, E. Filiput4, J. Dowling5, J. Huang1, and T. Kim1; 1WashU Medicine, Department of Radiation Oncology, St. Louis, MO, 2Brainlab Inc., Munich, Germany, 3Washington University in St. Louis School of Medicine, Department of Radiation Oncology, St. Louis, MO, 4BJC Healthcare, Department of Radiation Oncology, St. Louis, MO, 5WashU Medicine, Department of Neurosurgery, Saint Louis, MO
Purpose/Objective(s): AI algorithms for tumor detection are highly sensitive to image quality, which varies based on imaging protocols and tumor volumes—a critical factor in Gamma Knife Radiosurgery (GKRS). This retrospective study evaluates an FDA-approved AI algorithm across different MRI protocols and volumes to assess its performance and reliability in the GKRS workflow.
Materials/Methods: Contrast-enhanced T1-weighted MRIs from 84 patients, who received GKRS and had no prior radiation therapy history, were retrospectively evaluated for this study. The MRIs were processed through an AI-based tumor detection algorithm. 1) To assess the performance dependence on MRI protocols, patients were divided into two groups based on the MRI protocol used for defining tumors; MPRAGE (Magnetization-Prepared Rapid Gradient-Echo) and SPGR (Spoiled Gradient Recalled). Each group consisted of 42 patients. 2) To evaluate the performance based on tumor volume, clinically treated tumors were categories by their reported volumes. In this study, clinically treated tumors were considered as ground truth in this study.
Results: A total of 233 brain metastases were identified and treated among the 84 patients (107 from MPRAGE and 126 from SPGR). The AI algorithm achieved a sensitivity of 0.798 and a positive predictive value (PPV) of 0.921 from all data. 1) When comparing MRI protocols, the sensitivity/PPV were 0.720/0.963 for MPRAGE and 0.865/0.893 for SPGR. 2) Regarding volume dependency, for tumors >0.05cc, the AI’s sensitivity exceeded 0.900 in all datasets. For tumors >0.4cc, the sensitivity was 1.00 from SPGR vs 0.881 from MPRAGE.
Conclusion: The AI-based tumors detection algorithm shows promising results, particularly for larger tumors, although some variations are noted depending on the imaging protocol used. The SPGR protocol demonstrated higher sensitivity than MPRAGE protocol. Further studies are needed to validate these findings and improve the algorithm's performance for smaller tumors.
Table 1. Information on Considered Tumor Volumes per MR Imaging Protocol. The mean ± standard deviation of the tumor volumes, with median volume in parentheses, are summarized for all considered tumors, AI tool detected tumors, and undetected tumors. There was no statistically significant difference in all considered tumor volumes between two protocols.| MPRAGE | SPGR | ||
| All considered tumors (True) | 107 | 126 | |
| 0.82±1.51 (0.18) | 0.98±2.46 (0.13) | ||
| AI detected tumors (True positive) | 77 | 109 | |
| 1.08±1.70 (0.88) | 1.13±2.61 (0.21) | ||
| AI undetected tumors (False Negative) | 30 | 17 | |
| 0.15±0.29 (0.02) | 0.05±0.09 (0.02) |