PaperlayerView exact revision on arXiv

Source-linked AI summary

The Federated Tumor Segmentation and Classification (FeTS++) Challenge

Wenyi Tang, Carla Pitarch-Abaigar, Siddhesh Thakur, Verena Chung, Konstantinos Koukoutegos, Alexandros Karargyris, Holger R. Roth, Spyridon Bakas

arXiv:2610.18771v1cs.CV

TL;DR

Centralized collection constrains the size and diversity of multi-institutional datasets, while FL allows institutions to train shared models without exchanging patient data. FeTS++ establishes a federated benchmark spanning tumor segmentation and classification, but this manuscript describes the challenge design and reports no results.

  • Problem

    Centralized data collection is limited by privacy, ownership, and legal considerations, while accurate tumor sub-region delineation remains time-consuming and subject to inter-rater variability.

    Source: §1 · p. 2

  • Method

    FeTS++ benchmarks federated segmentation and classification on a multi-institutional glioma and meningioma cohort, with locked model architectures and an NVFlare evaluation protocol.

    Source: §1 · p. 3, §4 · p. 10

  • Results

    This manuscript reports the challenge design rather than results; results and a complete meta-analysis are deferred until after the challenge.

    Source: §4 · p. 10

  • Takeaways & Limitations

    The shared, locked protocol isolates the contribution of federated methods from model architecture across segmentation and classification tasks.

    Source: §4 · p. 10

    The manuscript deliberately reports no results; results and a complete meta-analysis will follow after challenge completion.

    Source: §4 · p. 10

Abstract

from arXiv · show

International challenges have become the standard for developing and benchmarking biomedical image analysis methods using centralized multi-institutional data. Federated learning (FL) describes a paradigm-shift on multi-institutional collaborations, where models are trained across institutions while their data are always retained locally. The Federated Tumor Segmentation (FeTS) challenge, first held in 2021, was the first challenge dedicated to FL in medical imaging and editions in 2022 and 2024 pushed the frontiers of real-world method validation. FeTS++ 2027 returns after a two-year gap with a substantially larger and more diverse cohort, and with the two "pluses" that name the challenge: an official classification task alongside segmentation, and a move of the infrastructure to NVIDIA FLARE (NVFlare). Built on the clinically acquired, multi-institutional multi-parametric MRI data of the BraTS 2025 Lighthouse challenge, FeTS++ spans gliomas and meningiomas across patient cohorts from institutions in North America, Europe, Asia, and Africa, and focuses on ranking submitted methodological approaches according to their model utility and an efficiency-aware score. Task 1 asks participants to find the best strategy for aggregating knowledge from per-institution segmentation models, while Task 2 asks them to distinguish glioma from meningioma under the same federated protocol, in which the entity label is heavily confounded with the site of origin.

1 Introduction

FeTS++ addresses constraints on centralized multi-institutional data sharing with a federated brain-tumor benchmark spanning segmentation and classification. Its design brings gliomas and meningiomas into a larger cohort and evaluates methods under a shared protocol.

1 Introduction

  • Manual delineation of tumor sub-regions remains time-consuming and subject to inter-rater variability despite mpMRI’s role in diagnosis, treatment planning, and response assessment.The passage describes mpMRI as typically comprising T1, T1-Gd, T2, and T2-FLAIR sequences.Source: §1 · p. 2
  • Centralized BraTS data collection is constrained by privacy regulations, data ownership, and legal considerations that limit dataset size and diversity.FL enables institutions to train a shared model without exchanging patient data, keeping data at local sites while communicating model updates.Source: §1 · p. 2
  • FeTS++ 2027 covers federated brain-tumor segmentation and classification across gliomas and meningiomas, using a clinically acquired cohort of over 3,500 patients.Participants can change aggregation strategies and client training behavior, while model architecture is fixed; evaluation uses hidden data in a controlled NVFlare simulator and an efficiency-aware PermRanker score.Source: §1 · p. 3
  • The challenge adds classification alongside segmentation and moves its infrastructure to NVIDIA FLARE (NVFlare).It builds on the FeTS series, which benchmarked FL aggregation and segmentation algorithms on medical imaging distributed across international institutions.Source: §1 · pp. 2–3
  • The challenge brings together more than 3,500 patients across gliomas, meningiomas, and multiple populations, including a Sub-Saharan Africa cohort.The challenge combines the two official tasks under one federated protocol and locks model architecture so gains are attributable to the FL aggregation method.Source: §1 · p. 3

2.1 Data

FeTS++ combines three clinically acquired MRI cohorts, preprocesses scans through registration and brain extraction, and uses expert-reviewed annotations with a held-out test set.

2.1.1 Data Sources

  • The dataset combines FeTS 2024 glioma cases, meningioma challenge cases, and a Sub-Saharan Africa cohort, with four co-registered MRI sequences per case.Training cases include expert reference sub-region labels; the data were publicly released with a de-identified institution identifier for each case.Source: §2.1.1 · p. 3

2.1.2 Data Pre-processing

  • All four MRI sequences are registered to the SRI-24 atlas and resampled to 1 mm^3 before manually checked brain extraction.N4 bias-field correction is used temporarily to estimate transformation matrices and then discarded because it otherwise obliterates tumor signal.Source: §2.1.2 · p. 4

2.1.3 Annotation Protocol

  • Glioma annotations were produced by clinical neuroradiologists and reviewed by experienced board-certified approvers, while meningioma annotations used refined nnU-Net pre-segmentations.Sub-Saharan African cases were refined by local radiology residents and radiologists before final approval by an expert BraTS neuroradiologist; annotations were not aggregated.Source: §2.1.3 · p. 4
  • Approximately 80% of collaborator data is retained for federated training outside the challenge, with the remainder set aside to test submitted algorithms.Testing datasets are not shared with participants or the public.Source: §2.1.3 · p. 5

2.2 Code Framework

The challenge uses NVFlare and MONAI in a simulator-first environment, with a locked editing surface for participant methods. Its ranking accounts for both task performance and convergence efficiency.

2.2.1 Software Stack

  • NVFlare coordinates server-client model distribution and update collection, while its modular design lets participants investigate aggregation strategies and local training behavior.MONAI supplies image transforms, sliding-window inference, the metric, and a baseline model wrapper; public and hidden evaluation share one frozen evaluator and code path.Source: §2.2.1 · p. 5

2.2.2 Locked Editing Policy

  • Participants may edit local training and aggregation, while the model architecture, assigned institutional data partitions, and permitted FLModel communications remain constrained.The simulator’s round sends the global model to clients, returns local updates, and aggregates accepted updates; the FLModel payload is the only information crossing the link.Source: §2.2.2 · pp. 5–6, Figure 1 · p. 6

2.2.3 Efficiency

  • Efficiency is measured as area under the performance-versus-time or round curve, normalized relative to the best-performing method and weighted at 20% of the final ranking.This rewards how quickly a method reaches its best accuracy rather than only its final score.Source: §2.2.3 · pp. 6–7

3.1 Task 1: Federated Tumor Segmentation

Task 1 isolates federated-learning method choices for aggregating institution-trained segmentation models under a fixed architecture, data pipeline, and evaluation protocol. Its deliberately imbalanced cohort and multi-region scoring define the challenge setting.

3.1.1 Task Description

  • Task 1 targets aggregation, client selection, per-round local training, compression, and communication efficiency rather than redesigning the segmentation network.The architecture, data pipeline, splits, initial checkpoint, and federated lifecycle are fixed, so measured differences reflect FL methods rather than model capacity or data access.Source: §3.1.1 · p. 7

3.1.2 Model Definition and Data

  • The 26-site cohort is deliberately imbalanced, with 4–542 cases per site and 19 sites contributing fewer than 50 cases.The baseline is a fixed MONAI SegRes-Net with four input channels, three output channels, an initial filter count of 8, and dropout probability 0.2.Source: §3.1.2 · p. 7, Figure 2 · p. 8

3.1.3 Performance Evaluation

  • Submissions are scored on enhancing tumor, tumor core, and whole tumor regions using Dice similarity coefficient and Normalized Surface Distance.The three regions comprise label 3; labels 1 and 3; and labels 1, 2, and 3, respectively. DSC measures predicted-mask overlap with ground truth, while NSD measures boundary agreement within a distance tolerance.Source: §3.1.3 · pp. 7–8

3.2 Task 2: Federated Tumor Classification

Task 2 evaluates federated classification of glioma versus meningioma using multi-institutional mpMRI without segmentation masks. The entity label is strongly associated with site, and scoring combines class-imbalance-insensitive metrics with an efficiency term.

3.2.1 Task Description

  • Task 2 classifies cases as glioma or meningioma under the shared FL protocol, using multi-institutional mpMRI without segmentation masks.Correct entity identification is described as the first clinical decision conditioning the treatment pathway.Source: §3.2.1 · p. 8

3.2.2 Model Definition and Data

  • The entity label is heavily confounded with site: 18 of 26 training sites are entity-pure, and the mixed sites are heavily unbalanced.The challenge evaluates behavior when label distributions differ across clients, making weighting schemes, client selection, and per-entity calibration relevant.Source: §3.2.2 · p. 9, §3.2.1 · p. 9

3.2.3 Performance Evaluation

  • Balanced accuracy and Matthews correlation coefficient (MCC) are ranked across methods, with higher values preferred and ties handled by rank averaging.Both metrics are described as insensitive to cohort class imbalance; balanced accuracy is mean per-class recall, while MCC measures chance-corrected prediction-label agreement.Source: §3.2.3 · p. 9
  • The final score is S_i = 0.4 P_BA + 0.4 P_MCC + 0.2 E_i, combining normalized metric ranks with normalized convergence score.Accuracy, AUC-ROC, F1, sensitivity, and specificity are reported as diagnostics.Source: §3.2.3 · p. 10

4 Discussion and Conclusion

FeTS++ 2027 provides a locked-protocol benchmark spanning segmentation and classification on a deliberately unbalanced international cohort. The manuscript reports challenge design rather than results, with meta-analysis planned after completion.

4 Discussion and Conclusion

  • The benchmark spans segmentation and classification across an international multi-institutional cohort that is deliberately realistic rather than balanced.The shared, locked protocol permits only aggregation and client-training logic to change, isolating FL-method contributions from architecture contributions.Source: §4 · p. 10
  • Task 2 tests whether methods that stabilize segmentation transfer to classification when entity labels are confounded with site and client case mix differs from the global average.The challenge also asks whether the two tasks require different methodological approaches.Source: §4 · p. 10
  • This manuscript reports no challenge results; results and a complete meta-analysis are planned for a joint publication after challenge completion.Source: §4 · p. 10

Related papers

Loading 2610.18771v1…