ESTRO 2025 - Abstract Book

S2516

Physics - Autosegmentation

ESTRO 2025

3219

Digital Poster Evaluating the Performance of a Novel Federated AI Learning Platform in Auto-detection and Segmentation of Intracranial Meningiomas Eyub Y AKDEMIR 1 , Wan-Yuo Guo 2 , Selin Gurdikyan 1 , Chun-Hsien Yu 3 , Kai-Wei Yu 2 , Kuei-Hong Kuo 4 , Robert H Press 1 , Matthew D Hall 1 , D Jay Wieczorek 1 , Yongsook C Lee 1 , Ranjini Tolakanahalli 1 , Alonso N Gutierrez 1 , Michael W McDermott 5 , Minesh P Mehta 1 , Rupesh Kotecha 1 1 Radiation oncology, Miami Cancer Institute, Miami, USA. 2 Radiology, Taipei Veterans General Hospital, Taipei, Taiwan. 3 AI labs, Taiwan AI labs, Taipei, Taiwan. 4 Radiology, Far Eastern Memorial Hospital, Taipei, Taiwan. 5 Neurosurgery, Miami Neuroscience Institute, Miami, USA Purpose/Objective: The adoption of artificial intelligence (AI) in stereotactic radiosurgery (SRS) treatment planning is rapidly evolving. This validation study evaluates the capabilities of a cutting-edge federated learning platform designed to automate the detection and segmentation of meningiomas, assessed through detailed quantitative and qualitative analyses. Material/Methods: The AI algorithm, developed from a multicenter dataset encompassing 23 institutions, was evaluated for its ability to assist the detection and segmentation of intracranial meningiomas. Anonymized, high-resolution treatment planning images were uploaded to a secure, cloud-based system for external validation. Tumors identified by the AI algorithm, including both true and false positives, were compared to ground truth. Sensitivity and positive predictive value (PPV) were calculated. Quantitative performance was assessed using Dice Similarity Coefficient (DSC) values (mean, median, standard deviation [SD], and interquartile range [IQR]), while qualitative accuracy was rated on a four-tier scale by an independent physician: no revision, minor, moderate, or major revisions required. Results: The validation cohort included 40 patients treated with SRS for 47 meningiomas, visualized using high-resolution 3T MPRAGE post-contrast imaging. The median lesion volume was 3.08 cc (IQR: 0.9–5.6 cc). The AI algorithm achieved a mean sensitivity of 72.3% and a PPV of 53.1% by detecting 34 out of 47 meningiomas with 30 false positive tumors. The algorithm attained a mean DSC of 66.1% (SD: ±29.7) and a median DSC of 78.4% (IQR: 53.9–89.5%). Most AI contours (70.6%, 24/34) were fully encapsulated by the external boundaries of the ground truth contours. When stratified by lesion size, larger lesions (≥ 3.5 cc, n=14) demonstrated significantly higher DSC values compared to smaller lesions (< 3.5 cc, n=20): 88.5% (IQR: 78.8–92.1%) vs. 64.0% (IQR: 24.8–80.8%), respectively (p = 0.002). However, independent physician evaluations indicated that 76.5% of AI-generated contours required moderate to major revisions. Conclusion: This AI-based algorithm, built on a federated learning platform, shows modest sensitivity and PPV for detecting intracranial meningiomas, particularly for larger lesions (≥ 3.5 cc). Despite its potential, the majority of AI-generated contours still required significant manual adjustments, emphasizing the need for further refinement. These results underscore the promise of AI integration into SRS workflows while highlighting the importance of continued development to enhance precision and reduce contour revision workload.

Keywords: artificial intelligence, meningioma, SRS