1154 - Advancing SRS Workflow: A Comparative Evaluation of Innovative Deep Learning Architectures for Automated Brain Metastasis Detection and Segmentation in T1-MRI
Presenter(s)
S. Huang1, Z. Xu2, Y. Yang2, G. Wang2, Y. Xue3, X. Zhang2, Y. Dong2, L. Xu2, Q. Wang2, W. Wang2, and Z. Yuan4; 1Tianjin Medical University Cancer Institute & Hospital, Tianjin, China, 2National Clinical Research Center for Cancer, Tianjin’s Clinical Research Center for Cancer, Tianjin Medical University Cancer Institute & Hospital, Tianjin, China, 3National Clinical Research Center for Cancer, Tianjin’s Clinical Research Center for Cancer, Tianjin Medical University Cancer Institute & Hospital, Tianjin, Tianjin, China, 4Department of Radiation Oncology, Tianjin Medical University Cancer Institute and Hospital, National Clinical Research Center for Cancer and Tianjin Key Laboratory of Cancer Prevention and Therapy, Tianjin, China
Purpose/Objective(s):
Manual contouring of brain metastases (BMs) in stereotactic radiosurgery (SRS) planning remains a critical bottleneck, with inter-observer variability exceeding 30% in clinical practice. This first-of-its-kind study systematically evaluates state-of-the-art deep learning architectures to establish an AI-powered solution for simultaneous BM detection and segmentation in T1-MRI.Materials/Methods:
We developed a comprehensive benchmark framework comparing eight deep learning models, based on CNN, Transformer, or Mamba architectures for the task of detecting and segmenting brain metastatic lesions in T1-contrast MRI. A total of 934 patients were included, with 667 cases from publicly available datasets and 267 cases from our institution, designated for training and testing, respectively. Data were retrospectively collected and organized at our institution, and GTV defined as the total BM tumor volume delineated by the physician at the time of stereotactic radiosurgery (SRS). Additionally, labels in the publicly available dataset were modified under clinician guidance to create a BM GTV that met clinical criteria to improve ground-truth accuracy. Sensitivity at both the patient and lesion levels was used to evaluate BM detection. Segmentation performance was assessed using several metrics: Dice Similarity Coefficient (DSC), Jaccard Coefficient (JC), Positive Predictive Value (PPV), Surface DSC (sDSC), Hausdorff Distance 95% (HD95), and Average Surface Distance (ASD). The performance across different BM diameters was also evaluated.Results: Among the eight deep learning models evaluated, the U-Mamba (Bot) achieved a lesion-level sensitivity of 0.796 (95% CI: 0.779-0.812) for all sizes of BM, which was significantly higher than that of the other models (pairwise McNemar test: p < 0.05), with a false positive rate of 2.46 ± 4.96 per patient. Further stratification by metastasis diameter, the sensitivity was 0.505 for BMs < 3 mm, 0.797 for BMs between 3 mm and 6 mm, and 0.885 for BMs between 6 mm and 9 mm. Moreover, U-Mamba (Enc) demonstrated significantly higher lesion-level segmentation performance (pairwise Wilcoxon signed rank test: p <0.05), with DSC and JC values of 0.632 ± 0.224 and 0.499 ± 0.230, respectively. In terms of tumor boundary segmentation, nnU-Netv2 achieved the best performance, with Surface DSC, HD95, and ASD values of 0.877 ± 0.149, 1.770 ± 1.458 mm, and 0.457 ± 0.369, respectively.
Conclusion: The nnU-Netv2 allows precise segmentation of lesion areas in T1-contrast MRI, while U-Mamba provide effective detection of brain metastasis, potentially aiding in treatment planning for SRS.