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.

Magnetic resonance (MR) nano-theranostic hyperthermia uses magnetic nanoparticles to target and accumulate at the lesions and generate heat to kill lesion cells directly through hyperthermia or indirectly through thermal activation and control releasing of drugs. Preclinical and translational applications of MR nano-theranostic hyperthermia are currently limited by a few major theoretical difficulties and experimental challenges in in vivo conditions. For example, conventional models for estimating the heat generated and the optimal magnetic nanoparticle sizes for hyperthermia do not accurately reproduce reported in vivo experimental results. In this work, a revised cluster-based model was proposed to predict the specific loss power (SLP) by explicitly considering magnetic nanoparticle aggregation in in vivo conditions. By comparing with the reported experimental results of magnetite Fe3O4 and cobalt ferrite CoFe2O4 magnetic nanoparticles, it is shown that the revised cluster-based model provides a more accurate prediction of the experimental values than the conventional models that assume magnetic nanoparticles act as single units. It also provides a clear physical picture: the aggregation of magnetic nanoparticles increases the cluster magnetic anisotropy while reducing both the cluster domain magnetization and the average magnetic moment, which, in turn, shift the predicted SLP toward a smaller magnetic nanoparticle diameter with lower peak values. As a result, the heating efficiency and the SLP values are decreased. The improvement in the prediction accuracy in in vivo conditions is particularly pronounced when the magnetic nanoparticle diameter is in the range of ~10–20 nm. This happens to be an important size range for MR cancer nano-theranostics, as it exhibits the highest efficacy against both primary and metastatic tumors in vivo. Our studies show that a relatively 20%–25% smaller magnetic nanoparticle diameter should be chosen to reach the maximal heating efficiency in comparison with the optimal size predicted by previous models.

No abstract

Previous data suggest that apparent diffusion coefficient (ADC) imaging phenotypes predict survival response to anti-VEGF monotherapy in glioblastoma. However, the mechanism by which imaging may predict clinical response is unknown. We hypothesize that decorin (DCN), a proteoglycan implicated in the modulation of the extracellular microenvironment and sequestration of pro-angiogenic signaling, may connect ADC phenotypes to survival benefit to anti-VEGF therapy. Patients undergoing resection for glioblastoma as well as patients included in The Cancer Genome Atlas (TCGA) and IVY Glioblastoma Atlas Project (IVY GAP) databases had pre-operative imaging analyzed to calculate pre-operative ADCL values, the average ADC in the lower distribution using a double Gaussian mixed model. ADCL values were correlated to available RNA expression from these databases as well as from RNA sequencing from patient derived mouse orthotopic xenograft samples. Targeted biopsies were selected based on ADC values and prospectively collected during resection. Surgical specimens were used to evaluate for DCN RNA and protein expression by ADC value. The IVY Glioblastoma Atlas Project Database was used to evaluate DCN localization and relationship with VEGF pathway via in situ hybridization maps and RNA sequencing data. In a cohort of 35 patients with pre-operative ADC imaging and surgical specimens, DCN RNA expression levels were significantly larger in high ADCL tumors (41.6 vs. 1.5; P = 0.0081). In a cohort of 17 patients with prospectively targeted biopsies there was a positive linear correlation between ADCL levels and DCN protein expression between tumors (Pearson R2 = 0.3977; P = 0.0066) and when evaluating different targets within the same tumor (Pearson R2 = 0.3068; P = 0.0139). In situ hybridization data localized DCN expression to areas of microvascular proliferation and immunohistochemical studies localized DCN protein expression to the tunica adventitia of blood vessels within the tumor. DCN expression positively correlated with VEGFR1 & 2 expression and localized to similar areas of tumor. Increased ADCL on diffusion MR imaging is associated with high DCN expression as well as increased survival with anti-VEGF therapy in glioblastoma. DCN may play an important role linking the imaging features on diffusion MR and anti-VEGF treatment efficacy. DCN may serve as a target for further investigation and modulation of anti-angiogenic therapy in GBM.

Purpose To evaluate the influence of phosphate on amine, amide, and hydroxyl CEST contrast using Bloch‐McConnell simulations applied to physical phantom data. Methods Phantom solutions of 4 representative metabolites with exchangeable protons—glycine (α‐amine protons), Cr (η‐amine protons), egg white protein (amide protons), and glucose (hydroxyl protons)—were prepared at different pH levels (5.6 to 8.9) and phosphate concentrations (5 to 80 mM). CEST images of the phantom were collected with CEST‐EPI sequence at 3 tesla. The CEST data were then fitted to full Bloch‐McConnell equation simulations to estimate the exchange rate constants. With the fitted parameters, simulations were performed to evaluate the intracellular and extracellular contributions of CEST signals in normal brain tissue and brain tumors, as well as in dynamic glucose‐enhanced experiments. Results The exchange rates of α‐amine and hydroxyl protons were found to be highly dependent on both pH and phosphate concentrations, whereas the exchange rates of η‐amine and amide protons were pH‐dependent, albeit not catalyzed by phosphate. With phosphate being predominantly intracellular, CEST contrast of α‐amine exhibited a higher sensitivity to changes in the extracellular microenvironment. Simulations of dynamic glucose‐enhanced signals demonstrated that the contrast between normal and tumor tissue was mostly due to the extracellular CEST effect. Conclusion The proton exchange rates in some metabolites can be greatly catalyzed by the presence of phosphate at physiological concentrations, which substantially alters the CEST contrast. Catalytic agents should be considered as confounding factors in future CEST‐MRI research. This new dimension may also benefit the development of novel phosphate‐sensitive imaging methods.