Although machine learning (ML) has shown promise across disciplines, out-of-sample generalizability is concerning. This is currently addressed by sharing multi-site data, but such centralization is challenging/infeasible to scale due to various limitations. Federated ML (FL) provides an alternative paradigm for accurate and generalizable ML, by only sharing numerical model updates. Here we present the largest FL study to-date, involving data from 71 sites across 6 continents, to generate an automatic tumor boundary detector for the rare disease of glioblastoma, reporting the largest such dataset in the literature (n = 6, 314). We demonstrate a 33% delineation improvement for the surgically targetable tumor, and 23% for the complete tumor extent, over a publicly trained model. We anticipate our study to: 1) enable more healthcare studies informed by large diverse data, ensuring meaningful results for rare diseases and underrepresented populations, 2) facilitate further analyses for glioblastoma by releasing our consensus model, and 3) demonstrate the FL effectiveness at such scale and task-complexity as a paradigm shift for multi-site collaborations, alleviating the need for data-sharing.

The multiple-drug resistance (MDR) transporter P-glycoprotein (P-gp) is highly expressed at the human blood–brain barrier (BBB). P-gp actively effluxes a wide variety of drugs from the central nervous system, including anticancer drugs. We have previously demonstrated P-gp activity at the human BBB using PET of ¹¹C-verapamil distribution into the brain in the absence and presence of the P-gp inhibitor cyclosporine-A (CsA). Here we extend the initial noncompartmental analysis of these data and apply compartmental modeling to these human verapamil imaging studies. Methods: Healthy volunteers were injected with ¹⁵O-water to assess blood flow, followed by ¹¹C-verapamil to assess BBB P-gp activity. Arterial blood samples and PET images were obtained at frequent intervals for 5 and 45 min, respectively, after injection. After a 60-min infusion of CsA (intravenously, 2.5 mg/kg/h) to inhibit P-gp, a second set of water and verapamil PET studies was conducted, followed by ¹¹C-CO imaging to measure regional blood volume. Blood flow was estimated using dynamic ¹⁵O-water data and a flow-dispersion model. Dynamic ¹¹C-verapamil data were assessed by a 2-tissue-compartment (2C) model of delivery and retention and a 1-tissue-compartment model using the first 10 min of data (1C₁₀). Results: The 2C model was able to fit the full dataset both before and during P-pg inhibition. CsA modulation of P-gp increased blood–brain transfer (K₁) of verapamil into the brain by 73% (range, 30%−118%; n = 12). This increase was significantly greater than changes in blood flow (13%; range, 12%−49%; n = 12, P < 0.001). Estimates of K₁ from the 1C₁₀ model correlated to estimates from the 2C model (r = 0.99, n = 12), indicating that a short study could effectively estimate P-gp activity. Conclusion: ¹¹C-verapamil and compartmental analysis can estimate P-gp activity at the BBB by imaging before and during P-gp inhibition by CsA, indicated by a change in verapamil transport (K₁). Inhibition of P-gp unmasks verapamil trapping in brain tissue that requires a 2C model for long imaging times; however, transport can be effectively measured using a short scan time with a 1C₁₀ model, avoiding complications with labeled metabolites and tracer retention.