Evaluation of deep learning-based methods for automatic detection and segmentation of brain metastases in T1-contrast MRI for stereotactic radiosurgery.

backgroundBrain metastases (BMs) are manually contoured during stereotactic radiosurgery (SRS) treatment planning, which is both time-consuming and potentially inconsistent. To address these challenges, researchers have been actively developing deep learning-based approaches for the detection and segmentation of BMs. However, a comprehensive comparative analysis of deep learning models across different frameworks remains largely absent in the current literature. This study aimed to evaluate and compare deep learning models based on different frameworks for the detection and segmentation of BMs in T1-contrast MRI. MATERIALS AND

methodsEight deep learning models, based on CNN, Transformer, or Mamba architectures, were trained and validated 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. A BM was considered detected if the ground-truth contour overlapped with a predicted structure. Sensitivity at both the patient-level (proportion of patients with at least one lesion detected) and lesion-level (proportion of ground-truth lesions detected) were used to evaluate BM detection. Segmentation performance was assessed using several metrics: dice similarity coefficient (DSC), positive predictive value (PPV), surface DSC (sDSC), and Hausdorff distance 95% (HD95). The performance across different BM diameters was also evaluated.

resultsAmong the eight deep learning models, 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, 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 and 6 mm, and 0.885 for BMs between 6 and 9 mm. Moreover, U-Mamba (Enc) demonstrated significantly higher lesion-level segmentation performance, with DSC value of 0.632 ± 0.224. In terms of tumor boundary segmentation, nnU-Netv2 achieved the best performance, with Surface DSC and HD95 values of 0.877 ± 0.149 and 1.770 ± 1.458 mm.

conclusionThe 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.