Study sample

We analyzed 99 patients from the cohort studied by Rodrigues et al.⁷ for whom MRI and BM contouring data were available. These 99 patients were randomly selected from the initial cohort to minimize selection bias. For the initial cohort, there were implicit inclusion and exclusion criteria reflective of clinical practice due to the retrospective design of the study by Rodrigues et al.⁷. Specifically, patients with highly symptomatic BMs or poor prognosis would be excluded from receiving SRS. For the retrospective study itself, the inclusion criteria consisted of patients with up to three newly diagnosed, radiologically confirmed BMs that were treated using SRS between 2003 and 2011. Patients with previous surgical resection, prior radiation therapy to the brain, or recurrent BMs were excluded. Due to the data requirements for MRI analysis and a radiologically confirmed endpoint, patients were also excluded if they did not have a pre-treatment or follow-up MRI.

The SRS treatments were performed using either the Novalis or Novalis TX linac with a Gill–Thomas–Cosman frame or BrainLAB frameless mask system for immobilization (BrainLAB, Feldkirchen, Germany). BMs < 7.5 cc generally received the most aggressive prescription (21 Gy in one fraction), while larger BMs would receive less aggressive prescriptions. This dose prescription guideline was developed by the centre at which the data were collected, as unified prescription protocols were not yet widely available. 127 BMs in total were treated with SRS. Four of these BMs were excluded as they consisted of a single voxel due to partial volume effects during application of the BM contouring data to the MRI voxel grid, and many radiomic features (e.g. all second-order texture features) would not be computable as they require more than one voxel. Outcome prediction and analysis was performed per BM instead of per patient, giving a total of n = 123 BMs that were individually analyzed. Table ​Table11 shows the clinical features of the patients and BMs in this study.

Clinical feature distributions for number of BMs, BMs progressing post-SRS, and patients (where applicable) for the study sample.

The “Neurological Symptoms Steroid Response” feature qualitatively scores the improvement of neurological symptoms after the administration of steroids, based on the previously reported methodology of Lagerwaard et al.²⁴. p-values provided for statistical comparisons between BMs that progressed and did not progress post-SRS. The Wilcoxon rank sum test was used for continuous features (age and GTV volume). The Chi-squared test was used for the remaining categorical features. NSCLC non-small cell lung cancer, ECOG Eastern Cooperative Oncology Group, WHO World Health Organization.

Each BM’s region-of-interest was defined as the gross tumour volume (GTV) manually contoured during SRS treatment planning by an experienced radiation oncologist, which was based on the outer border of the enhancing region in high-resolution T1w-CE MRI. Non-enhancing regions within the border were included in the contour. T1w-CE MRI was acquired with different MR scanner models, voxel resolutions, and acquisition orientations. Five scanner models from Siemens and General Electric (Erlangen, Germany; Chicago, USA) and eight scan configurations were used (Table ​(Table22).

Number of patients and BMs scanned by each of the MR scanner model and acquisition parameters configurations.

Chi-squared test for progression yielded p = 0.226 across all scanner models and p = 0.069 across further investigated Vision, Avanto, and Expert scanners.