Latest observed volumetric MRI response and lesion level local control after gamma knife radiosurgery for breast cancer brain metastases

Background Serial magnetic resonance imaging (MRI) reveals heterogeneous lesion responses after stereotactic radiosurgery beyond binary local-control status. We characterized latest-observed volumetric response categories after Gamma Knife radiosurgery (GKRS) for breast cancer brain metastases and examined their concordance with 12-month local control (LC 12 m). Methods This retrospective study included 80 patients, 182 lesions, and 819 MRI records. LC 12 m incorporated earlier local failure or assessment within a 10–14-month window. Four response categories compared the latest eligible whole-tumor (WT) volume with baseline. Baseline imaging and dose correlates of failure were evaluated using variational Bayesian mixed-effects logistic regression and patient-clustered generalized estimating equations (GEE). Early-landmark analyses explored associations with subsequent failure. Results LC 12 m was evaluable in 105/182 lesions (70 controlled, 35 failed). Among 154 response-classifiable lesions, 24 (15.6%) showed complete response, 56 (36.4%) partial regression, 48 (31.2%) stable volume, and 26 (16.9%) enlargement. Enlargement was strongly associated with failure (GEE odds ratio 16.797, 95% confidence interval 3.398–83.033). Contemporaneous concordance measures were 54.3% sensitivity, 97.1% specificity, 90.5% positive predictive value, and 81.0% negative predictive value; 97/105 category assessments occurred within the endpoint window. Early-landmark analyses were inconclusive. In 102 complete cases, variational Bayesian odds ratios were 1.888 [95% credible interval (CrI): 1.079–3.302] per unit increase in ln[1 + WT (cm 3 )], 3.071 (1.520–6.203) for edema, and 0.760 (0.697–0.829) per 10-Gy increase in BED 10 . GEE estimates were directionally consistent but did not reach statistical significance. Same-model, same-prior Markov chain Monte Carlo analysis supported the inverse association between BED 10 and local failure, whereas evidence for the WT and edema associations remained inconclusive. Conclusions Serial volumetric MRI and latest-observed response categories complement binary local-control assessment by describing heterogeneous lesion-level responses after GKRS. Enlargement showed strong contemporaneous concordance with LC 12 m, with overlapping definitions and assessment times limiting claims of independent prediction. Baseline imaging and BED 10 associations were model-dependent. External validation is needed to establish the framework's added clinical value.