ESTRO 2025 - Abstract Book

S2515

Physics - Autosegmentation

ESTRO 2025

3214

Digital Poster Evaluating the Performance of a Novel Federated AI Learning Platform in Auto-detection and Segmentation of Brain Metastases Eyub Y AKDEMIR 1 , Wan-Yuo Guo 2 , Selin Gurdikyan 1 , Yi-Chin E Tu 3 , 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: Integration of AI algorithms into stereotactic radiosurgery (SRS) treatment planning is gaining momentum. This study assesses the performance of a novel federated learning platform to facilitate auto-detection and segmentation of brain metastases through quantitative and qualitative measures. Material/Methods: We assessed the performance of an AI-based algorithm developed on a 23-center data consortium to auto-detect and segment intact brain metastases (BM). Dedicated treatment planning images were anonymized and uploaded to a secure, cloud-based platform to perform this external validation. True and false positive lesions identified by the AI-based algorithm were compared to ground truth lesions independently for each patient, verified independently by a three-physician review, to calculate sensitivity and positive predictive value (PPV). To quantify the accuracy of AI contours, detected lesions were analyzed using both quantitative methods, volumetric Dice Similarity Coefficient (DSC), in mean and median values with standard deviation (SD) and interquartile range (IQR) and qualitative clinical assessments for each lesion using the following scale: no revision, minor, moderate, or major revisions. Results: The validation cohort consisted of 236 lesions treated with SRS in 50 patients, visualized with 3T MPRAGE post contrast images. Median lesion sizes and volumes were 0.78 cm (IQR: 0.49-1.17 cm) and 0.25 cc (IQR: 0.06-0.84 cc), respectively. The AI-based algorithm demonstrated a mean and median sensitivity for lesion detection of 69.2% (SD: ±33.3) and 80.0% (IQR: 40.9-100.0%) with PPV values of 78.2% (SD: ±29.8) and 100.0% (50.0-100.0%) at the patient level. At the lesion level, the algorithm achieved a mean and median DSC of 60.4% (SD: ±29.2) and 72.3% (IQR: 33.3 85.7%) over the 123 lesions, which were detected and segmented by the AI algorithm. Moreover, AI-based segmentations were consistently within the boundaries of ground truth contours, 98.4% (121/123). When stratified by diameter size, the DSCs were improved for lesions ≥ 1 cm in maximum diameter (n=40) compared to < 1 cm (n=83) (84.5% [IQR: 74.0-88.9%] vs. 56.0% [IQR: 21.7-80.0%], p<0.001). respectively. Based on independent physician evaluation, 58.7% of the AI-based contours required moderate to major revision. Conclusion: This novel AI-based algorithm based on a federated learning platform demonstrates promising sensitivity and positive predictive value in detecting BM, especially in larger lesions (≥ 1 cm). Physician assessment indicated that the majority of AI-generated contours required moderate to major revisions. These findings highlight the potential of AI integration into SRS workflows, but critically underscore the need for refinement to improve reliability and reduce contour revision workload.

Keywords: artificial intelligence, brain metastases, SRS