## Introduction

FL has tremendous potential in healthcare22,23, particularly towards addressing health disparities, under-served populations, and “rare” diseases24, by enabling ML models to gain knowledge from ample and diverse data that would otherwise not be available. With that in mind, here we focus on the “rare” disease of glioblastoma, and particularly on the detection of its extent using multi-parametric magnetic resonance imaging (mpMRI) scans25. While glioblastoma is the most common malignant primary brain tumor26,27,28, it is still classified as a “rare” disease, as its incidence rate (i.e., 3/100,000 people) is substantially lower than the rare disease definition rate (i.e., <10/100,000 people)24. This means that single sites cannot collect large and diverse datasets to train robust and generalizable ML models, necessitating collaboration between geographically distinct sites. Despite extensive efforts to improve the prognosis of glioblastoma patients with intense multimodal therapy, their median overall survival is only 14.6 months after standard-of-care treatment, and 4 months without treatment29. Although the subtyping of glioblastoma has been improved30 and the standard-of-care treatment options have expanded during the last 20 years, there have been no substantial improvements in overall survival31. This reflects the major obstacle in treating these tumors which is their intrinsic heterogeneity26,28, and the need for analyses of larger and more diverse data toward a better understanding of the disease. In terms of radiologic appearance, glioblastomas comprise of three main sub-compartments, defined as (i) the “enhancing tumor” (ET), representing the vascular blood-brain barrier breakdown within the tumor, (ii) the “tumor core” (TC), which includes the ET and the necrotic (NCR) part, and represents the surgically relevant part of the tumor, and (iii) the “whole tumor” (WT), which is defined by the union of the TC and the peritumoral edematous/infiltrated tissue (ED) and represents the complete tumor extent relevant to radiotherapy (Fig. 1b). Detecting these sub-compartment boundaries, therefore, defines a multi-parametric multi-class learning problem and is a critical first step towards further quantifying and assessing this heterogeneous rare disease and ultimately influencing clinical decision-making.

Co-authors in this study have previously introduced FL in healthcare in a simulated setting15 and further conducted a thorough quantitative performance evaluation of different FL workflows11 (refer to supplementary figures for illustration) for the same use-case as the present study, i.e., detecting the boundaries of glioblastoma sub-compartments. Findings from these studies supported the superiority of the FL workflow used in the present study (i.e., based on an aggregation server10,14), which had almost identical performance to CL, for this use-case. Another study32 has explored the first real-world federation for a breast cancer classification task using 5 sites, and another16 used electronic medical records along with x-ray images from 20 sites to train a classifier to output a label corresponding to future oxygen requirement for COVID-19 patients.

This study describes the largest to-date global FL effort to develop an accurate and generalizable ML model for detecting glioblastoma sub-compartment boundaries, based on data from 6314 glioblastoma patients from 71 geographically distinct sites, across six continents (Fig. 1a). Notably, this describes the largest and most diverse dataset of glioblastoma patients ever considered in the literature. It was the use of FL that successfully enabled our ML model to gain knowledge from such an unprecedented dataset. The extended global footprint and the task complexity are what sets this study apart from current literature, since it dealt with a multi-parametric multi-class problem with reference standards that require expert clinicians following an involved manual annotation protocol, rather than simply recording a categorical entry from medical records16,32. Moreover, varying characteristics of the mpMRI data due to scanner hardware and acquisition protocol differences33,34 were handled at each collaborating site via established harmonized preprocessing pipelines35,36,37,38,39.

The scientific contributions of this manuscript can be summarized by (i) the insights garnered during this work that can pave the way for more successful FL studies of increased scale and task complexity, (ii) making a potential impact for the treatment of the rare disease of glioblastoma by publicly releasing clinically deployable trained consensus models, and most importantly, iii) demonstrating the effectiveness of FL at such scale and task complexity as a paradigm shift redefining multi-site collaborations, while alleviating the need for data sharing.

## Results

The complete federation followed a staged approach, starting from a “public initial model” (trained on data of 231 cases from 16 sites), followed by a “preliminary consensus model” (involving data of 2471 cases from 35 sites), to conclude on the “final consensus model” (developed on data of 6314 cases from 71 sites). To quantitatively evaluate the performance of the trained models, 20% of the total cases contributed by each participating site were excluded from the model training process and used as “local validation data”. To further evaluate the generalizability of the models in unseen data, 6 sites were not involved in any of the training stages to represent an unseen “out-of-sample” data population of 590 cases. To facilitate further evaluation without burdening the collaborating sites, a subset (n = 332) of these cases was aggregated to serve as a “centralized out-of-sample” dataset. The training was initiated from a pre-trained model (i.e., our public initial model) rather than a random initialization point, in order to have faster convergence of the model performance40,41. Model performance was quantitatively evaluated here using the Dice similarity coefficient (DSC), which assesses the spatial agreement between the model’s prediction and the reference standard for each of the three tumor sub-compartments (ET, TC, WT).

### Increased data can improve performance

When the federation began, the public initial model was evaluated against the local validation data of all sites, resulting in an average (across all cases of all sites) DSC per sub-compartment, of DSCET = 0.63, DSCTC = 0.62, DSCWT = 0.75. To summarize the model performance with a single collective score, we then calculate the average DSC (across all 3 tumor sub-compartments per case, and then across all cases of all sites) as equal to 0.66. Following model training across all sites, the final consensus model garnered significant performance improvements against the collaborators’ local validation data of 27% (pET < 1 × 10−36), 33% (pTC < 1 × 10−59), and 16% (pWT < 1 × 10−21), for ET, TC, and WT, respectively (Fig. 1c). To further evaluate the potential generalizability improvements of the final consensus model on unseen data, we compared it with the public initial model against the complete out-of-sample data and noted significant performance improvements of 15% (pET < 1 × 10−5), 27% (pTC < 1 × 10−16), and 16% (pWT < 1 × 10−7), for ET, TC, and WT, respectively (Fig. 1d). Notably, the only difference between the public initial model and the final consensus model, was that the latter gained knowledge during training from increased datasets contributed by the complete set of collaborators. The conclusion of this finding reinforces the importance of using large and diverse data for generalizable models to ultimately drive patient care.

### Data size alone may not predict success

This is initially observed in our federated setting, where the comparative evaluation of the public initial model, the preliminary consensus model, and the final consensus model, against the centralized out-of-sample data, indicated performance improvements not directly related to the amount of data used for training. Specifically, we noted major significant (p < 7 × 10−18, Wilcoxon signed-rank test) performance improvements between the public initial model and the preliminary consensus model, as opposed to the insignificant (p > 0.067, Wilcoxon signed-rank test) ones between the preliminary and the final consensus model, as quantified in the centralized out-of-sample data for all sub-compartments and their average (Fig. 2).

We further expanded this analysis to assess this observation in a non-federated configuration, where we selected the largest collaborating sites (comprehensive cancer centers contributing > 200 cases, and familiar with computational analyses), and coordinated independent model training for each, starting from the public initial model and using only their local training data. The findings of this evaluation indicate that the final consensus model performance is always superior or insignificantly different (pAverage = 0.1, pET = 0.5, pTC = 0.2, pWT = 0.06, Wilcoxon signed-rank test) to the ensemble of the local models of these four largest contributing collaborators, for all tumor sub-compartments (Fig. 2). This finding highlights that even large sites can benefit from collaboration.

### FL is robust to data quality issues

Data quality issues relating to erroneous reference annotations (with potential negative downstream effects on output predictions) were identified by monitoring the global consensus model performance during training. However, only data quality issues that largely affected the global validation score could be identified and corrected during training. Those with more subtle effects in the global validation score were only identified after the completion of the model training by looking for relatively low local validation scores of the consensus model across collaborating sites. An example of such a quality issue with erroneous reference labels (from Site 48) is shown in Fig. 3c. Looking closer, local validation scores at Site 48 (Fig. 3b) are significantly different (pET < 3 × 10−12, pTC < 3 × 10−12, pWT < 3 × 10−12, Wilcoxon signed-rank test) than the average scores across the federation (Fig. 3a). Significant differences were calculated by sample pairs for each federated round, where a sample pair consists of the mean validation score over samples for Site 48 paired with those across all sites. These local validation scores (Fig. 3b) indicate that the model is not gaining knowledge from these local data, and their comparison with the average scores across the federation (Fig. 3a) indicates that the global consensus model performance is not adversely affected. This finding supports the importance of robustness at a global scale.

### FL benefits the more challenging tasks

The complexity of boundary detection drops when moving from smaller to larger sub-compartments, i.e., from ET to TC, and then to WT35,36,37,38. This is further confirmed here, as evidenced by the model’s relative performance indicated by the local validation curves and their underlying associated areas in Fig. 3.a. Since the current clinically actionable sub-compartments are TC (i.e., considered for surgery) and WT (i.e., considered for radiotherapy)42, performance improvements of their boundary detection may contribute to the model’s clinical impact and relevance.

Our findings indicate that the benefits of FL are more pronounced for the more challenging sub-compartments, i.e., larger performance improvements for ET and TC compared to WT (Fig. 1c). Notably, the largest and most significant improvement (33%, p < 7 × 10−60) is noted for the TC sub-compartment, which is surgically actionable and not a trivial sub-compartment to delineate accurately43,44. This finding of FL benefiting the more challenging tasks rather than boosting performance on the relatively easier task (e.g., thresholding the abnormal T2-FLAIR signal for the WT sub-compartment) by gaining access to larger amounts of good quality data holds a lot of promise for FL in healthcare.

### Optimal model selection is non-trivial

Using the performance of the global consensus model during training across all local validation cases, two distinct model configurations were explored for selecting the final consensus model. Analyzing the sequence of consensus models produced during each federated round, we selected four different models: the singlet, for which the average DSC across all sub-compartments scored high, and three independent models, each of which yielded high DSC scores for each tumor sub-compartment, i.e., ET, TC, WT. We defined the collection of these three independent consensus models as a triplet.

To identify the best model, 5 singlets and 5 triplets were selected based on their relative performance on all local validation cases and evaluated against the centralized out-of-sample data. Only small differences are observed between the singlet and triplet models, and these differences diminish as the sub-compartment size increases. Comparing the means of singlet and triplet, the larger (and only significant) performance improvement difference compared to the public initial model is noted for the ET sub-compartment (improved by < 3%, pET = 0.02), followed by TC (improved by < 1.4%, pTC = 0.09), and then lastly WT (improved by < 1.1%, pWT = 0.2) (Tables S1 and S2). However, the decision of using a singlet or a triplet model should also rely on computational cost considerations, as triplets will be three times more expensive than singlets during model inference.

## Discussion

In this study, we have described the largest real-world FL effort to-date utilizing data of 6314 glioblastoma patients from 71 geographically unique sites spread across 6 continents, to develop an accurate and generalizable ML model for detecting glioblastoma sub-compartment boundaries. Notably, this extensive global footprint of the collaborating sites in this study also yields the largest dataset ever reported in the literature assessing this rare disease. It is the use of FL that successfully enabled (i) access to such an unprecedented dataset of the most common and fatal adult brain tumor, and (ii) meaningful ML training to ensure the generalizability of models across out-of-sample data. In comparison with the limited existing real-world FL studies16,32, our use-case is larger in scale and substantially more complex, since it (1) addresses a multi-parametric multi-class problem, with reference standards that require expert collaborating clinicians to follow an involved manual annotation protocol, rather than simply recording a categorical entry from medical records, and (2) requires the data to be preprocessed in a harmonized manner to account for differences in MRI acquisition. Since glioblastoma boundary detection is critical for treatment planning and the requisite first step for further quantitative analyses, the models generated during this study have the potential to make a far-reaching clinical impact.

The large and diverse data that FL enabled, led to the final consensus model garnering significant performance improvements over the public initial model against both the collaborators’ local validation data and the complete out-of-sample data. The improved result is a clear indication of the benefit that can be afforded through access to more data. However, increasing the data size for model training without considerations relating to data quality, reference labels, and potential site bias (e.g., scanner acquisition protocols, demographics, or sociocultural considerations, such as more advanced presentation of disease at diagnosis in low-income regions45) might not always improve results. Literature also indicates an ML performance stagnation effect, where each added case contributes less to the model performance as the number of cases increase46. This is in line with our finding in the federated setting (Fig. 2), where performance improvements across the public initial model, the preliminary consensus model, and the final consensus model, were not directly/linearly related to the amount of data used for training. This happened even though the final consensus model was trained on over twice the number of cases (and included 2 of the largest contributing sites—Sites 1 and 4) when compared to the preliminary consensus model. Further noting that the preliminary federation model was already within the intra- and inter-rater variability range for this use-case (20% and 28%, respectively)47, any further improvements for the full federation consensus model would be expected to be minimal35,36,37,38.

To further assess these considerations, we coordinated independent model training for the four largest collaborating sites (i.e., >200 cases) by starting from the same public initial model and using only their local training data. The ensemble of these four largest site local models did not show significant performance differences to the final consensus model for any tumor sub-compartment, yet the final consensus model showed superior performance indicating that even sites with large datasets can benefit from collaboration. The underlying assumption for these results is that since each of these collaborators initiated their training from the public initial model (which included diverse data from 16 sites), their independent models and their ensemble could have inherited some of the initial model’s data diversity, which could justify the observed insignificant differences (Fig. 2 and Supplementary Fig. 3). Though these findings are an indication that the inclusion of more data alone may not lead to better performance, it is worth noting that these four largest sites used for the independent model training represent comprehensive cancer centers (compared to hospitals in community settings) with affiliated sophisticated labs focusing on brain tumor research, and hence were familiar with the intricacies of computational analyses. Further considering the aforementioned ML performance stagnation effect, we note the need for generalizable solutions to quantify the contribution of collaborating sites to the final consensus model performance, such that future FL studies are able to formally assess both the quantity and the quality of the contributed data needed by the collaborating sites and decide on their potential inclusion on use-inspired studies.

As noted in our results, due to the lack of such generalizable solutions, we were only able to identify quality issues after the model training. Specifically, we hypothesize that although Site 48 had data quality issues, its effect on the consensus model performance was not significant due to its relatively small dataset (n = 46) when compared to the other collaborating sites. The curves of Fig. 3a indicate that the global consensus model continues to consistently gain knowledge from the federation as a whole during training, highlighting robustness to such data quality issues. It remains unknown, however, how much better the consensus model would have performed if sites with problematic data were excluded or if these specific problematic data at Site 48 were excluded or corrected. These findings are aligned with literature observations (on the same use-case)48, where a DL model49 trained on 641 glioblastoma cases from 8 sites produced higher quality predictions on average than those created as reference standard labels by radiology expert operators. Quality was judged by 20 board-certified neuroradiologists, in a blinded side-by-side comparison of 100 sequestered unseen cases, and concluded that perfect or near-perfect reference labels may not be required to produce high-quality prediction systems. In other words, DL models may learn to see past imperfect reference training labels. These findings provide the impetus for further experimentation as they have implications for future FL studies. Future research is needed to automatically detect anomalies in the consensus model performance during training, particularly associated with contributions from individual sites.

There are a number of practical considerations that need to be taken into account to set up a multi-national real-world federation, starting with a substantial amount of coordination between each participating site. As this study is the first at this scale and task complexity, we have compiled a set of governance insights from our experience that can serve as considerations for future successful FL studies. These insights differ from previous literature that describes studies that were smaller in scale and involved simpler tasks16,32. By “governance” of the federation we refer both to the accurate definition of the problem statement (including reference labels and harmonization considerations accounting for inter-site variability), and the coordination with the collaborating sites for eligibility and compliance with the problem statement definition, as well as security and technical considerations. For future efforts aiming to conduct studies of a similar global scale, it would be beneficial to identify a solution for governance prior to initiating the study itself.

The coordination began with engaging the security teams of collaborating sites and providing them access to the source code of the platform developed to facilitate this study. These security discussions highlighted the benefit of the platform being open-source, making security code reviews easier. Resource gathering was then carried out by identifying technical leads and assessing computational resources at each site. With the technical leads, we then proceeded to test the complete workflow to further identify gaps in the requirements, such as network configurations and hardware requirements. We then proceeded with data curation and preprocessing, and finally connected individual sites to the aggregation server to initiate their participation.

In general, debugging issues with the inputted local data and annotations is more difficult during FL due to the level of coordination and/or privacy issues involved, since the data are always retained at the collaborating site. We gained substantial experience during this effort that went into further development of use-inspired but generalizable data sanity-checking functionality in the tools we developed, towards facilitating further multi-site collaborations.

Upon conclusion of the study, sites participating in the model training process were given a survey to fill in regarding various aspects of their experience. According to the provided feedback, 96% of the sites found the comprehensive documentation on preprocessing and data curation essential and thought that lack of such documentation could have resulted in inconsistent annotations. Additionally, 92% found the documentation relating to establishing secure connectivity to the aggregation server easy to follow and essential to expedite reviews by the related groups. Furthermore, 84% of the sites appreciated the user-friendly interface of the provided tool and its associated complete functionality (beyond its FL backend), and indicated their intention to use it and recommend it for projects and data analysis pipelines beyond the scope of this study. To generate the reference standard labels for their local data, 86% of the collaborating sites indicated that they used either the FeTS Tool39 (i.e., the tool developed for this study), CaPTk51, or ITK-SNAP52, whereas the remaining 14% used either 3D-Slicer53, the BraTS toolkit54, or something else. In terms of hardware requirements at each site, 88% used a dedicated workstation for their local workload, and the remaining 12% used either a containerized form of the FeTS tool or a virtual machine.

Although data are always retained within the acquiring site during FL (and hence FL is defined as private-by-design), different security and privacy threats remain55,56,57. These threats include attempted extraction of training data information from intermediate and final models, model theft, and submission of poison model updates with the goal of introducing unwanted model behavior (including incentivizing the model to memorize more information about the training data in support of subsequent extraction, i.e., leakage). A number of technologies can be used to mitigate security and privacy concerns during FL55,56,57. Homomorphic encryption58, secure multiparty compute59, and trusted execution environments (TEEs)60,61 allow for collaborative computations to be performed with untrusted parties while maintaining confidentiality of the inputs to the computation. Differentially private training algorithms62,63,64 allow for mitigation of information leakage from both the collaborator model updates and the global consensus aggregated models. Finally, assurance that remote computations are executed with integrity can be designed for with the use of hardware-based trust provided by TEEs, as well as with some software-based integrity checking65. Each of these technologies comes with its own benefits in terms of security and/or privacy, as well as costs and limitations, such as increased computational complexity, associated hardware requirements and/or reduced quality of computational output (such as the reduction of model utility that can be associated with differentially private model training). Further experimentation needs to be done in order to best inform prospective federations as to which technologies to use towards addressing their specific concerns within the context of the collaborator infrastructure and trust levels, depending on the use-case, the extent of the collaborating network, and the level of trust within the involved parties. Our study was based on a collaborative network of trusted sites, where authentication was based on personal communication across collaborating sites and the combination of TLS and TEEs were considered sufficient.

Although our study has the potential to become the baseline upon which future ML research studies will be done, there is no automated mechanism to assess inputted data quality from collaborators, which could result in models trained using sub-optimal data. Additionally, we used a single off-the-shelf neural network architecture for training, but it has been shown that model ensembles perform better for the task at hand35,36,37,38, and it remains to be explored how such a strategy could be explored in a federated study. Moreover, the instantiation of the federation involved a significant amount of coordination between each site and considering the limited real-world FL studies at the time, there were no tools available to automate such coordination and orchestration. These involved (i) getting interviewed by information security officers of collaborating sites, (ii) ensuring that the harmonized preprocessing pipeline was used effectively, (iii) clear communication of the annotation protocol, and iv) testing the network communication between the aggregator and each site. This amount of effort, if not aided by automated tools, will continue to be a huge roadblock for FL studies, and dedicated coordination and orchestration resources are required to conduct this in a reproducible and scalable manner.

We have demonstrated the utility of an FL workflow to develop an accurate and generalizable ML model for detecting glioblastoma sub-compartment boundaries, a finding which is of particular relevance for neurosurgical and radiotherapy planning in patients with this disease. This study is meant to be used as an example for future FL studies between collaborators with an inherent amount of trust that can result in clinically deployable ML models. Further research is required to assess privacy concerns in a detailed manner63,64 and to apply FL to different tasks and data types66,67,68,69. Building on this study, a continuous FL consortium would enable downstream quantitative analyses with implications for both routine practice and clinical trials, and most importantly, increase access to high-quality precision care worldwide. Furthermore, the lessons learned from this study with such a global footprint are invaluable and can be applied to a broad array of clinical scenarios with the potential for great impact on rare diseases and underrepresented populations.

## Methods

The study and results presented in this manuscript comply with all relevant ethical regulations and follow appropriate ethical standards in conducting research and writing the manuscript, following all applicable laws and regulations regarding the treatment of human subjects. Use of the private retrospective data collection of each collaborating site has been approved by their respective institutional review board, where informed consent from all participants was also obtained and stored.

### Data

The data considered in this study described patient populations with adult-type diffuse glioma30, and specifically displaying the radiological features of glioblastoma, scanned with mpMRI to characterize the anatomical tissue structure25. Each case is specifically described by (i) native T1-weighted (T1), (ii) Gadolinium-enhanced T1-weighted (T1Gd), (iii) T2-weighted (T2), and (iv) T2-weighted-Fluid-Attenuated-Inversion-Recovery (T2-FLAIR) MRI scans. Cases with any of these sequences missing were not included in the study. Note that no inclusion/exclusion criterion applied relating to the type of acquisition (i.e., both 2D axial and 3D acquisitions were included, with a preference for 3D if available), or the exact type of sequence (e.g., MP-RAGE vs. SPGR). The only exclusion criterion was for T1-FLAIR scans that were intentionally excluded to avoid mixing varying tissue appearance due to the type of sequence, across native T1-weighted scans.

The publicly available data from the International Brain Tumor Segmentation (BraTS) 2020 challenge35,36,37, was used to train the public initial model of this study. The BraTS challenge35,36,37,38, seeking methodological advancements in the domain of neuro-oncology, has been providing the community with (i) the largest publicly available and manually-curated mpMRI dataset of diffuse glioma patients (an example of which is illustrated in Fig. 1b), and (ii) a harmonized preprocessing pipeline51,70,71 to handle differences in inter-site acquisition protocols. The public initial model was used to initialize the FL training, instead of a randomly generated initialization, as starting from a pre-trained model leads to faster convergence41. The complete BraTS 2020 dataset originally included cases from sites that also participated in this study as independent collaborators. To avoid any potential data leakage, we reduced the size of the complete BraTS dataset by removing cases acquired by these specific sites, resulting in a dataset of 231 cases from 16 international sites, with varying contributing cases across sites (Fig. 1e). The exact site IDs that construct the data of the public initial model are: 47, 51, 55, 57, 58, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, and 71. Subsequently, the resulting dataset was split at a 4:1 ratio between cases for training (n = 185) and validation (n = 46).

The eligibility of collaborating sites to participate in the federation was determined based on data availability, and approval by their respective institutional review board. 55 sites participated as independent collaborators in the study defining a dataset of 6083 cases. The MRI scanners used for data acquisition were from multiple vendors (i.e., Siemens, GE, Philips, Hitachi, Toshiba), with magnetic field strength ranging from 1T to 3T. The data from all 55 collaborating sites followed a male:female ratio of 1.47:1 with ages ranging between 7 and 94 years.

From all 55 collaborating sites, 49 were chosen to be part of the training phase, and 6 sites were categorized as “out-of-sample”, i.e., none of these were part of the training stage. These specific 6 out-of-sample sites (Site IDs: 8, 11, 19, 20, 21, 43) were allocated based on their availability, i.e., they have indicated expected delayed participation rendering them optimal for model generalizability validation. One of these 6 out-of-sample sites (Site 11) contributed aggregated a priori data from a multi-site randomized clinical trial for newly diagnosed glioblastoma (ClinicalTrials.gov Identifier: NCT00884741, RTOG082572,73, ACRIN668674,75), with inherent diversity benefiting the intended generalizability validation purpose. The American College of Radiology (ACR - Site 11) serves as the custodian of this trial’s imaging data on behalf of ECOG-ACRIN, which made the data available for this study. Following screening for the availability of the four required mpMRI scans with sufficient signal-to-noise ratio judged by visual observation, a subset of 362 cases from the original trial data were included in this study. The out-of-sample data totaled 590 cases intentionally held out of the federation, with the intention of validating the consensus model in completely unseen cases. To facilitate further such generalizability evaluation without burdening the collaborating sites, a subset consisting of 332 cases (including the multi-site clinical data provided by ACR) from this out-of-sample data was aggregated, to serve as the “centralized out-of-sample” dataset. Furthermore, the 49 sites participating in the training phase define a collective dataset of 5493 cases. The exact 49 site IDs are: 1, 2, 3, 4, 5, 6, 7, 9, 10, 12, 13, 14, 15, 16, 17, 18, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 44, 45, 46, 48, 49, 50, 52, 53, 54, 56, 59, 60. These cases were automatically split at each site following a 4:1 ratio between cases for training and local validation. During the federated training phase, the data used for the public initial model were also included as a dataset from a separate node, such that the contribution of sites providing the publicly available data is not forgotten within the global consensus model. This results in the final consensus model being developed based on data from 71 sites over a total dataset of 6314 cases. Collective demographic information of the included population is provided in Table S3.

### Harmonized data preprocessing

Once each collaborating site identified its local data, they were asked to use the preprocessing functionality of the software platform we provided. This functionality follows the harmonized data preprocessing protocol defined by the BraTS challenge35,36,37,38, as described below. This would allow accounting for inter-site acquisition protocol variations, e.g., 3D vs. 2D axial plane acquisitions.

#### File-type conversion/patient de-identification

The respective mpMRI scans (i.e., T1, T1Gd, T2, T2-FLAIR) of every case are downloaded onto a local machine in the Digital Imaging and Communications in Medicine (DICOM) format76,77,78 and converted to the Neuroimaging Informatics Technology Initiative (NIfTI) file format79 to ensure easier parsing of the volumetric scans during the computational process. The conversion of DICOM to NIfTI files has the benefit of eliminating all patient-identifiable metadata from the header portion of the DICOM format80,81.

#### Rigid registration

Once the scans are converted to the NIfTI format, each volume is registered to a common anatomical space, namely the SRI24 atlas82, to ensure a cohesive data shape ([240, 240, 155]) and an isotropic voxel resolution (1 mm3), thereby facilitating in the tandem analysis of the mpMRI scans. One of the most common types of MRI noise is based on the inhomogeneity of the magnetic field83. It has been previously36 shown that the use of non-parametric, non-uniform intensity normalization to correct for these bias fields84,85 obliterates the MRI signal relating to the regions of abnormal T2-FLAIR signal. Here, we have taken advantage of this adverse effect and used the bias field-corrected scans to generate a more optimal rigid registration solution across the mpMRI sequences. The bias field-corrected images are registered to the T1Gd image, and the T1Gd image is rigidly registered to the SRI24 atlas, resulting in two sets of transformation matrices per MRI sequence. These matrices are then aggregated into a single matrix defining the transformation of each MRI sequence from its original space to the atlas. We then apply this single aggregated matrix to the NIfTI scans prior to the application of the bias field correction to maximize the fidelity of the finally registered images.

#### Brain extraction

This process focuses on generating a brain mask to remove all non-brain tissue from the image (including neck, fat, eyeballs, and skull), to enable further computational analyses while avoiding any potential face reconstruction/recognition86. For this step we utilized the Brain Mask Generator (BrainMaGe)87, which has been explicitly developed to address brain scans in presence of diffuse glioma and considers brain shape as a prior, hence being agnostic to the sequence/modality input.

#### Generation of automated baseline delineations of tumor sub-compartment boundaries

We provided the ability to the collaborating sites to generate automated delineations of the tumor sub-compartments from three popular methods from the BraTS challenge, using models trained using the challenge’s training data: (i) DeepMedic49, (ii) DeepScan88, and (iii) nnU-Net89. Along with segmentations from each method, label fusion strategies were also employed to provide a reasonable approximation to the reference labels that should be manually refined and approved by expert neuroradiologists to create the final reference labels. The label fusion approaches considered were i) standard voting90, (ii) Simultaneous Truth And Performance Level Estimation (STAPLE)91,92, iii) majority voting93, and iv) Selective and Iterative Method for Performance Level Estimation (SIMPLE)94.

#### Manual refinements towards reference standard labels

It was communicated to all participating sites to leverage the annotations generated using the automated mechanism as a baseline on which manual refinements were needed by neuroradiology experts, following a consistently communicated annotation protocol. The reference annotations comprised the Gd-enhancing tumor (ET—label ‘4’), the peritumoral edematous/invaded tissue (ED—label ‘2’), and the necrotic tumor core (NCR—label ‘1’). ET is generally considered the most active portion of the tumor, described by areas with both visually avid, as well as faintly avid, enhancement on the T1Gd scan. NCR is the necrotic part of the tumor, the appearance of which is hypointense on the T1Gd scan. ED is the peritumoral edematous and infiltrated tissue, defined by the abnormal hyperintense signal envelope on the T2-FLAIR scans, which includes the infiltrative non-enhancing tumor, as well as vasogenic edema in the peritumoral region35,36,37,38 (an illustration can be seen in Fig. 1b).

#### Data splits

Once the data were preprocessed, training and validation cohorts were created randomly in a 4:1 ratio, and the splits were preserved during the entire duration of the FL training to prevent data leakage. The performance of every model was compared against the local validation data cohort on every federated round.

We leveraged the data loading and processing pipeline from the Generally Nuanced Deep Learning Framework (GaNDLF)95, to enable experimentation with various data augmentation techniques. Immediately after data loading, we removed the all-zero axial, coronal, and sagittal planes from the image, and performed a z-score normalization of the non-zero image intensities96. Each tumor sub-compartment of the reference label is first split into an individual channel and then passed to the neural network for processing. We extracted a single random patch per mpMRI volume set during every federated round. The patch size was kept constant at [128, 128, 128] to ensure that the trained model can fit the memory of the baseline hardware requirement of each collaborator, i.e., a discrete graphics processing unit with a minimum of 11 GB dedicated memory. For data augmentation, we added random noise augmentation (μ = 0.0, σ = 0.1) with a probability of p = 0.2, random rotations (90° and 180°, with the axis of rotation being uniformly selected in each case from the set of coronal, sagittal, and axial planes) each with a probability of p = 0.5, and a random flip augmentation with a probability of p = 1.0 with equal likelihood of flips across the sagittal, coronal, and axial planes.

### The neural network architecture

The trained model to delineate the different tumor sub-compartments was based on the popular 3D U-Net with residual connections (3D-ResUNet)97,98,99,100,101, an illustration of which can be seen in the Supplementary Fig. 1. The network had 30 base filters, with a learning rate of lr = 5 × 10−5 optimized using the Adam optimizer102. For the loss function used in training, we used the generalized DSC score103,104 (represented mathematically in Eq. (1)) on the absolute complement of each tumor sub-compartment independently. Such mirrored DSC loss has been shown to capture variations in smaller regions better89. No penalties were used in the loss function, due to our use of ‘mirrored’ DSC loss105,106,107. The final layer of the model was a sigmoid layer, providing three channel outputs for each voxel in the input volume, one output channel per tumor sub-compartment. While the generalized DSC score was calculated using a binarized version of the output (check sigmoid value against the threshold 0.5) for the final prediction, we used the floating point DSC108 during the training process.

$${\rm {DSC}}=\frac{2{\left|{\rm {RL}}\odot {\rm {PM}}\right|}_{1}}{{\left|{\rm {RL}}\right|}_{1}+{\left|{\rm {PM}}\right|}_{1}}$$
(1)

where RL serves as the reference label, PM is the predicted mask,  is the Hadamard product109 (i.e., component-wise multiplication), and $${\left|x\right|}_{1}$$ is the L1-norm110, i.e., the sum of the absolute values of all components).

### The Federation

We followed a staged approach for the training of the global consensus model, starting from a preliminary smaller federation across a subset (n = 35) of the participating sites to evaluate the complete process and resolve any initial network issues. Note that 16 of these 35 sites were used to train the public initial model, and used in the preliminary federation as an aggregated dataset. The exact 19 site IDs that participated in the training phase of the preliminary federation, as independent sites are: 2, 3, 9, 14, 22, 23, 24, 27, 28, 29, 31, 33, 36, 37, 41, 46, 53, 54, and 59. The total data held by this smaller federation represented approximately 42% (n = 2471) of the data used in the full federation. We also trained individual models (initialized using the public initial model) using centralized training at all sites holding >200 training cases, and performed a comparative evaluation of the consensus model with an ensemble of these “single site models”. The per voxel sigmoid outputs of the ensemble were computed as the average of such outputs over the individual single-site models. As with all other models in this study, binary predictions were computed by comparing these sigmoid outputs to a threshold value of 0.5. The single-site model ensemble utilized (via the data at the single site) approximately 33% of the total data across the federation.

### Model runtime in low-resource settings

Clinical environments typically have constrained computational resources, such as the availability of specialized hardware (e.g., DL acceleration cards) and increased memory, which affect the runtime performance of DL inference workloads. Thus, taking into consideration the potential deployment of the final consensus model in such low-resource settings, we decided to proceed with a single 3D-ResUNet, rather than an ensemble of multiple models. This decision ensured a reduced computational burden when compared with running multiple models, which is typically done in academic research projects35,36,37,38.

To further facilitate use in low-resource environments, we have provided a post-training run-time optimized114 version of the final consensus model. Graph level optimizations (i.e., operators fusion) were initially applied, followed by optimizations for low precision inference, i.e., converting the floating point single precision model to a fixed precision 8-bit integer model (a process known as “quantization”115). In particular, we used accuracy-aware quantization116, where model layers were iteratively scaled to a lower precision format. These optimizations yielded run-time performance benefits, such as lower inference latency (a platform-dependent 4.48 × average speedup and 2.29 × reduced memory requirement when compared with the original consensus model) and higher throughput (equal to the 4.48 × speedup improvement since the batch size used is equal to 1), while the trade-off was an insignificant (pAverage < 7 × 10−5) drop in the average DSC.

#### Clinically-deployable consensus models

To further encourage the reproducibility of our study, and considering enhancing the potential impact for the study of the rare disease of glioblastoma, we publicly released the trained models of this study. We specifically released the final singlet and triplet consensus models, including the complete source code used in the project. Taking into consideration the potential deployment of these models in clinical settings, we refrained from training an ensemble of models (as typically done in academic research projects35,36,37,38), due to the additional computational burden of running multiple models. Furthermore, to facilitate use in low-resource environments, we also provide a post-training run-time optimized114 version of the final consensus model that obviates the need for any specialized hardware (such as DL acceleration cards) and performs insignificantly different from the final consensus model when evaluated against the centralized out-of-sample data.

### Reporting summary

Further information on research design is available in the Nature Research Reporting Summary linked to this article.