S1579
Physics - Autosegmentation
ESTRO 2026
Michael W. McDermott 2,3 , Minesh P. Mehta 1,2 , Rupesh Kotecha 1,2 1 Department of Radiation Oncology, Miami Cancer Institute, Baptist Health South Florida, Miami, USA. 2 Herbert Wertheim College of Medicine, Florida International University, Miami, USA. 3 Department of Neurosurgery, Miami Neuroscience Institute, Baptist Health South Florida, Miami, USA Purpose/Objective: With the expanding role of artificial intelligence (AI) in radiation oncology, its application to stereotactic radiosurgery (SRS) planning has become increasingly prominent. This study quantitatively evaluates an FDA- cleared AI algorithm for automated detection and segmentation of brain metastases in a unique cohort of patients imaged with both MPRAGE and 3D TSE sequences. Material/Methods: Dedicated treatment planning images from a cohort of patients with intact brain metastases were uploaded to a system installed in a dedicated workstation for this evaluation. For each patient, lesions detected by the AI-based algorithm were classified as true or false positives (FP) and compared against the ground truth (GT), serving as the reference for calculating sensitivity and positive predictive value (PPV). To evaluate contouring accuracy, we employed the volumetric Dice Similarity Coefficient (DSC). The distribution of above metrics was reported using mean and standard deviation (SD), along with median and interquartile range (IQR). Results: This validation dataset was comprised of 50 patients with 236 brain metastases which were detected using 3T post-contrast MPRAGE and 3D TSE image sets followed by a three-physician independent peer review process. In the overall cohort, median effective lesion size (maximum diameter) and volume were 0.46 cm (IQR: 0.29-0.83 cm) and 0.05 cc (IQR: 0.01-0.30 cc), respectively. At the patient level, the AI algorithm demonstrated a mean ± SD and median sensitivity for lesion detection of 74.4% ±31.6 and 96.7% (IQR: 50.0- 100.0%) with PPV of 94.1% ±13.9 and 100.0% (100.0- 100.0%), respectively. At the lesion level, the AI algorithm correctly identified 129 (54.6%) and falsely identified 11(4.6%) lesions. Among the 129 detected lesions, the AI algorithm achieved mean ± SD and median DSC of 80.6% ±13.4 % and 84.3% (IQR: 77.8- 88.6%). When stratified by effective lesion diameter, median DSCs were notably improved for lesions ≥ 0.5 cm in effective diameter (n=82) compared to < 0.5 cm (n=47) (63.6% [IQR: 83.2-89.6%] vs. 36.4% [IQR: 67.5- 81.4%], p<0.001), respectively. When using a DSC threshold of ≥ 0.85, 64.6% of the lesions with effective diameters ≥ 0.5 cm achieved this criterion, as compared to 10.6% for those lesions with effective
Figure 1. Comparison of expert annotations and automatically predicted masks. Conclusion: The proposed solution demonstrates superior performance over current SOTA benchmarks, yielding robust segmentation across diverse brain tumor entities (GLI, MEN, and PED) in MRI. This high-accuracy delineation confirms the model's potential as an effective tool for automatic contouring in treatment planning systems. Nevertheless, precise automatic delineation of tiny metastatic lesions remains an open challenge References: [1] Aboian M, Anazodo U, Baid U, Bakas S, Calabrese E, Marco Conte G, et al. MICCAI 2025 Lighthouse Challenge: Brain Tumor Segmentation Cluster of Challenges (BraTS)[2] Kato S, Hotta K. Adaptive t-vMF dice loss: An effective expansion of dice loss for medical image segmentation. Comput Biol Med 2024;168:107695. https://doi.org/10.1016/J.COMPBIOMED.2023.107695.[ 3] Isensee F, Wald T, Ulrich C, Baumgartner M, Roy S, Maier-Hein K, et al. nnU-Net Revisited: A Call for Rigorous Validation in 3D Medical Image Segmentation 2024. Keywords: brain tumor segmentation
Digital Poster 3104
Evaluation of an FDA-cleared AI Platform for Automated Brain Metastases Identification and Segmentation Naseem Ud Din 1 , Eyub Y. Akdemir 1 , Robert Herrera 1 , Matthew Hall 1,2 , Djay J. Wieczorek 1,2 , Yongsook C. Lee 1,2 , Ranjini Tolakanahalli 1,2 , Alonso N. Gutierrez 1,2 ,
Made with FlippingBook - Share PDF online