A federated MRI method called MOSAIC reported higher tumor-segmentation scores than weakly supervised comparison methods. On the FeTS2022 dataset, its average Dice score was 0.84 and its IoU score was 0.79, compared with about 0.73 and 0.67 for CAM baselines. Its HD95 boundary-distance value was 32.6 millimeters, versus 42–43 millimeters for those baselines.
The findings come from an arXiv preprint, version 1 dated 20 August 2026. The study asks whether binary tumor segmentation can be performed from image-level labels when institutions have different MRI modalities missing.
Built for uneven MRI data
MOSAIC uses two federated phases. First, image-level labels are used to generate class-activation-map (CAM) pseudo-masks, or initial estimated tumor maps. A dedicated segmentation stage then refines them. Both phases use client-specific modality alignment and spectral prototype alignment.
The evaluation used three publicly available, multi-institutional brain-tumor MRI datasets: FeTS2022, BraTS-MEN and BraTS-SSA. The reported configuration included 160 BraTS-MEN subjects across four clients and 60 BraTS-SSA subjects across three clients. Researchers compared Dice, IoU and 95th-percentile Hausdorff distance, or HD95, for each client and in macro-averaged results.
Results varied by comparison
On BraTS-MEN, MOSAIC averaged 0.81 on Dice and 0.79 on IoU. Its Dice score was 0.03 higher than AME-CAM and 0.02 higher than FedDM. The reported comparisons met the p<0.01 threshold, except for HD95 against FedDM.
On BraTS-SSA, MOSAIC averaged 0.78 Dice, 0.73 IoU and 51.6 millimeters in average boundary error. It was 0.08 Dice above ScoreCAM and 0.09 above FedDM, while a fully supervised reference was reported at roughly 0.80–0.83 Dice.
The comparison with a centralized image-level reference on FeTS2022 was mixed. MOSAIC had Dice of 0.84 versus 0.82 and a 0.03 IoU advantage, but the centralized reference had the lower HD95 value: 29.9 millimeters versus 32.6 millimeters.
What the extra stages added
The second phase was associated with average Dice gains of 0.04 on FeTS2022, 0.02 on BraTS-MEN and 0.03 on BraTS-SSA, from Phase 1 averages of 0.80, 0.79 and 0.75. Reported client-level changes included 0.77 to 0.86 and 0.66 to 0.75.
A patient-aligned feature analysis also found lower distribution-distance values in the spectral-prototype-alignment condition. At the alignment output, MMD2 was 1.10 versus 1.39 and the Fréchet distance was 167.4 versus 346.8. At the bottleneck, the corresponding figures were 0.23 versus 0.37 and 28.9 versus 43.4.
When a new client was added, its Dice values ranged from 0.80 to 0.82, compared with 0.82 to 0.86 for jointly trained base clients—a gap of 0.01 to 0.04 Dice. Under test-time augmentation, average background entropy was 0.0047 on FeTS2022, 0.0033 on BraTS-MEN and 0.0039 on BraTS-SSA, while the reported Dice variability stayed below 0.035. These figures are stability diagnostics, not validated clinical calibration estimates.
Why the findings are preliminary
The evidence is limited to retrospective computational experiments on public MRI benchmarks. The method handles binary tumor segmentation in a slice-based 2D/2.5D setup; the study did not test fully volumetric 3D performance, genuinely noisy report-derived image-level labels, or combinations such as MRI with PET, CT or clinical time-series.
No prospective clinical deployment, reader study or patient-outcome evaluation was reported. The authors also note that performance depends on the quality of CAM pseudo-labels and that the tested public datasets may not represent every institution or acquisition protocol. Independent testing on external multi-institutional data will be needed to assess how well the method generalizes.
Paper data and sources
Original title: MOSAIC: Modality-agnostic Spectral Alignment for Federated Image-level Weakly Supervised Tumor Segmentation under Client-specific Missing Modalities
Authors: Tarun Kumar Garg, Vaanathi Sundaresan
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-20
DOI: Not available
Original paper · Full text