Purpose: To automatically track control and progression of brain metastases (BrM) following stereotactic radiosurgery (SRS); and to provide a platform for studying BrM management with SRS and systemic therapies.
Methods: 299 patients with 527 BrM of diameter 10+ mm were retrospectively identified for this study. They were imaged with T1 MR post-Gd prior to SRS treatment for BrM, then imaged again at least once within 1 year of SRS. The target SRS dose varied between 18 and 30 Gy in 1-5 fractions. Median time interval between SRS and first follow-up was 61 days, with interquartile range of [48, 75] days . A 3D convolutional neural network was developed in-house to segment BrM gross tumor volumes (GTVs) on T1 MR post-Gd images with 100% lesion sensitivity above 10 mm diameter, and average Hausdorff distances of 2 mm. Pre- and post-treatment images were rigidly registered using an automatic multi-step process that considers the whole head, brain, and lateral ventricles. Then, the deep learning model was applied to automatically contour BrM GTVs on all images. Based on changes in longest diameter post-SRS, BrM were classified on follow-up imaging as: size increased (+20% change or more), size stable, size decreased (-30% change or more), and unappreciable.
Results: The registration method was robust, with error of approximately 1 mm . Relative size changes of 527 BrM were assessed on first available follow-up imaging after SRS, with 53 lesions increased in size, 83 stable, 193 decreased, and 84 unappreciable. Changes in 208 BrM (126 patients) were also analyzed after 180 days post-SRS, with 26 lesions increased in size, 58 stable, 59 decreased, and 65 unappreciable.
Conclusion: This translational study demonstrates a proof-of-concept for automatic longitudinal tracking of BrM. Future work can build on this to study BrM dose response and interactions between systemic therapies and SRS.