Source-linked AI summary
Federated Learning for Multi-Center Imaging Diagnostics: A Study in Cardiovascular Disease
Akis Linardos, Kaisar Kushibar, Sean Walsh, Polyxeni Gkontra, Karim Lekadir
TL;DR
Single-center medical imaging diagnosis offers limited evidence about generalization across heterogeneous institutions, while diagnosis-focused federated studies remain limited. This paper evaluates federated CMR diagnosis across four centers using HCM classification and shape priors; it reports performance comparable to centralized data sharing, alongside increased robustness and sensitivity to domain shift.
Problem
Single-center CMR diagnosis studies provide limited evidence about generalization to heterogeneous, unseen centers, and privacy-preserving federated learning for diagnosis has been less explored than for segmentation.
Source: §1 · pp. 1–2Method
The study compares federated learning and centralized data sharing for HCM-versus-normal CMR diagnosis across four centers, testing mask-derived shape priors and different evaluation setups.
Source: §1 · pp. 2–3, §2.6 · p. 6Results
With 180 subjects from four centers, federated CMR diagnosis achieved promising performance comparable to collaborative data sharing, with increased robustness and greater sensitivity to domain shift effects.
Source: §4 · p. 9Takeaways & Limitations
The authors highlight that evaluation should account for both in-site and out-of-site performance when assessing federated CMR diagnosis.
Source: §4 · pp. 8–9The study is limited to binary classification and clinician-provided segmentation masks, and the authors call for larger, more diverse datasets and automated federated segmentation.
Source: §4 · p. 10
Abstract
from arXiv · showhide
Deep learning models can enable accurate and efficient disease diagnosis, but have thus far been hampered by the data scarcity present in the medical world. Automated diagnosis studies have been constrained by underpowered single-center datasets, and although some results have shown promise, their generalizability to other institutions remains questionable as the data heterogeneity between institutions is not taken into account. By allowing models to be trained in a distributed manner that preserves patients' privacy, federated learning promises to alleviate these issues, by enabling diligent multi-center studies. We present the first federated learning study on the modality of cardiovascular magnetic resonance (CMR) and use four centers derived from subsets of the M\&M and ACDC datasets, focusing on the diagnosis of hypertrophic cardiomyopathy (HCM). We adapt a 3D-CNN network pretrained on action recognition and explore two different ways of incorporating shape prior information to the model, and four different data augmentation set-ups, systematically analyzing their impact on the different collaborative learning choices. We show that despite the small size of data (180 subjects derived from four centers), the privacy preserving federated learning achieves promising results that are competitive with traditional centralized learning. We further find that federatively trained models exhibit increased robustness and are more sensitive to domain shift effects.
1 Introduction
The introduction frames multi-center CMR diagnosis as a privacy-preserving learning challenge under heterogeneous data and limited cross-center evidence. It presents a four-center HCM study comparing federated learning with centralized data sharing and examining priors, augmentation, and in-site versus out-of-site evaluation.
1 Introduction
- Federated learning transfers model parameters rather than patient data, addressing privacy concerns associated with sending data to a central server.The introduction contrasts FL with centralized data sharing, where data is transferred and training occurs at the server.Source: §1 · p. 1, Figure 1 · p. 3
- Single-center diagnosis studies leave generalization to unseen centers uncertain, while heterogeneous scanners, protocols, and demographics make multi-center data non-IID.Prior cardiac MRI diagnosis studies used the single-center ACDC dataset and did not study privacy preservation.Source: §1 · pp. 1–2
- The study presents the first federated CMR diagnosis study, comparing federated learning with centralized data sharing across four centers for HCM classification.The study focuses on binary classification between normal and HCM subjects because labels overlap poorly across centers.Source: §1 · p. 2
- The study examines ground-truth-mask-derived shape priors and diverse augmentations, including an equal-voting federated variant that is beneficial in some cases.It uses a 3D-CNN pretrained on action recognition and repeats experiments with different weight initializations.Source: §1 · pp. 2–3
- Two repeated cross-validation setups estimate performance on partitions of known centers and on entirely unseen centers, showing that the two differ substantially.The authors highlight the importance of both on-site and out-of-site performance evaluation.Source: §1 · p. 3
2.1 Dataset
The study combines selected M&M and ACDC data into a four-center dataset for binary HCM-versus-normal diagnosis, and illustrates the shape-prior input variants.
2.1 Dataset
- Three M&M centers and a subset of ACDC form the four-center dataset used for HCM-versus-normal classification.The selected labels were chosen from a subset with the most balanced HCM and NOR representation.Source: §2.1 · p. 3
- Severe label imbalance across M&M centers makes multi-label federated training difficult, so the experiments restrict classification to HCM versus NOR.Local models can overfit to the labels present at individual centers.Source: §2.1 · p. 3
- The illustrated inputs are baseline one-channel MRI, MRI multiplied by a segmentation mask, and three channels for the right ventricle, myocardium, and left ventricle.Source: Figure 2 · p. 4
2.2 Data Preprocessing
The preprocessing pipeline standardizes MRI volumes, uses clinician-provided masks to introduce heart-shape priors, and harmonizes intensities across centers with privacy-preserving aggregate histograms.
2.2 Data Preprocessing
- Volumes are resampled to 1×1×1 mm3 and cropped to a 150×150×10-voxel window centered on the segmentation-mask bounding box.N4 bias field correction is applied first to remove low-frequency intensity non-uniformity.Source: §2.2 · p. 4
2.2.1 Induced priors
- The model receives either replicated single-channel MRI, mask-multiplied MRI, or three channels for the left ventricle, right ventricle, and myocardium.End-Diastole and End-Systole volumes are separate samples, but always remain in the same data split.Source: §2.2.1 · p. 4
2.2.2 Data Harmonization
- Nyúl histogram matching aligns image intensities to the average histogram of training-center images, followed by rescaling to [0, 1].The aggregate histogram avoids access to individual subject data; masked experiments match using values within segmentation masks, and the centers included can vary by evaluation set-up.Source: §2.2.2 · p. 4
2.3 Data Augmentation
The study compares four augmentation set-ups to increase the limited training data and assess how induced domain shifts affect centralized and federated learning.
Experimental scope
- The four set-ups are compared in both federated and centralized learning to study how dataset shift affects the frameworks.The motivation is limited M&M sample size and the possibility that augmentation-induced domain shifts improve generalization to unseen datasets.Source: §2.3 · p. 4
Augmentation set-ups
- Basic augmentation uses rotations and axial-view flips; shape augmentation adds elastic deformation, while shape-and-intensity augmentation also adds MRI artifacts, Gaussian noise, and gamma changes.The techniques are illustrated on a single CMR slice.Source: §2.3 · pp. 4–5, Figure 3 · p. 5
Application and rationale
- Augmentations are sampled from their pools and applied with 50% probability, excluding the no-augmentation set-up.The selected transformations are intended to reflect clinical variation, including poor acquisitions, while avoiding extreme elastic deformation.Source: §2.3 · p. 5
2.4 3D-CNN Model
The model adapts a Kinetics-400-pretrained 3D ResNet18 for binary HCM-versus-NOR classification, balancing transfer-learning bias against overfitting through fine-tuning.
Model and initialization
- The 3D ResNet18 is initialized with Kinetics-400 action-recognition weights and adapted to classify HCM versus NOR.The authors motivate transfer learning by the scarcity of medical-imaging data and freeze initial layers in a preliminary configuration, training only the 512-parameter final linear layer.Source: §2.4 · p. 5
Fine-tuning and stopping
- A learning rate of 1e-5 for pretrained layers performed best in preliminary experiments, while the new linear layer used 0.01.Training ran for at most 100 epochs with early stopping after 10 stagnant validation epochs; stopping occurred around epochs 20–30.Source: §2.4 · p. 5
2.5 Federated Learning
Federated training distributes a shared model to centers, performs local updates, and aggregates them; the study also examines equal center voting against sample-size weighting.
Federated training procedure
- The global model is initialized identically across four centers, trained locally for seven batches per epoch, and aggregated after each epoch.Batch sizes differ so centers parse their entire datasets in the same number of iterations.Source: §2.5 · p. 5
Cross-validation set-ups
- Collaborative Cross Validation assigns each center 20% of its data for testing across five folds, whereas Leave Center Out tests one center at a time.The remaining data are used for training and validation in each set-up.Source: Figure 4 · p. 6
Aggregation choices
- FederatedAveraging aggregates center models in proportion to sample size, which can bias learning toward larger centers under center-level domain shifts and data imbalance.In this study, Sagrada Familia has three times the data of SantPau; equal voting across centers is tested as an alternative and is beneficial in some cases.Source: §2.5 · p. 6
2.6 Evaluation Procedure
The study evaluates CDS and federated learning with matched data splits, using both coordinated cross-validation and leave-center-out validation to measure performance on held-out data and unseen centers. Five repeated initializations and histogram standardization further shape the evaluation.
2.6 Evaluation Procedure
- Four-fold LCO-CV holds out a different center in each fold, estimating generalization to centers unseen during training.In both LCO-CV and CCV, CDS and FL use the same data split for comparable folds.Source: §2.6 · p. 6
- Five-fold CCV aggregates test data from all centers, with the same fold splits used for CDS and FL.Each fold reserves 20% for testing; of the remaining data, 90% is used for training and 10% for validation.Source: §2.6 · p. 6
- After cross-validation, AUC is calculated from predictions on each fold’s test set.Table 2 reports AUC results for CDS and FL across curation types.Source: §2.6 · p. 6, Table 2 · p. 7
- Five different network initializations are averaged to estimate mean AUC and framework robustness.The same folds are used for each experiment’s CDS and FL comparison.Source: §2.6 · p. 6
- In LCO-CV, histogram matching excludes the held-out center from the reference, unlike CCV’s reference averaged across all centers.The Figure 5 histograms illustrate center intensities before and after standardization.Source: §2.6 · p. 7, Figure 5 · p. 7
3 Results
Shape priors improve performance, while augmentation effects depend on the evaluation split: CDS benefits under LCO-CV, whereas performance drops under CCV. Federated models show consistent results across initializations, with center-specific exceptions in ACDC and Sagrada Familia.
3 Results
- Both CDS and FL improve with induced shape priors, peaking with the per-structure split; their performance difference narrows after constraining the solution space.FL has lower standard deviations than CDS across these curation types.Source: §3 · p. 7
- Under CCV, introducing augmentation lowers performance across collaborative learning frameworks, as training-only augmentation creates domain shift between training and test data.The reported results indicate that this shift outweighs the benefit of additional data.Source: §3 · p. 7
- Under LCO-CV, CDS improves consistently with additional augmentations, while FL benefits from rotation and flipping but declines with shape and intensity augmentation.With shape and intensity augmentation, CDS surpasses FL; FL-trained models are more consistent across initializations than CDS models.Source: §3 · p. 7
- In ACDC, CDS outperforms FL in most cases, with a 0.3 AUC margin when intensity augmentations are introduced.The comparison is reported for the LCO-CV center-level analysis.Source: §3 · p. 7
- With maximum augmentation, FL-EV outperforms FL by 0.05 on Sagrada Familia, the largest center in the cohort.Sagrada Familia is the only center reported to benefit substantially from FL-EV.Source: §3 · p. 7
4 Discussion
Federated CMR diagnosis achieved promising performance comparable to collaborative data sharing on 180 subjects from four centers. Results varied with evaluation setting and augmentation, while the authors identify binary classification and clinician-supplied masks as scope limits.
4 Discussion
- FL outperformed CDS in most experiments, but complex shape and intensity augmentation reversed that pattern as FL became more sensitive to domain shifts.FL first benefited from rotation and flipping; CDS performance worsened with augmentation in CCV but improved as more augmentation was applied in LCO-CV.Source: §4 · pp. 8–9
- LCO-CV AUC was lower by approximately 0.1; under shape and intensity augmentation, FL scored 0.472 on ACDC versus about 0.85 to 0.89 for M&M centers, while CCV ACDC scored 0.785.The authors suggest more similar M&M acquisition protocols may explain the bias against ACDC.Source: §4 · p. 8, Table 4 · p. 9, Table 3 · p. 9
- Equal voting among centers sometimes improved and sometimes worsened performance; Sagrada Familia gained 0.05 AUC under shape and intensity augmentation.The authors propose exploring voting schemes between equal votes and sample-size weighting.Source: §4 · p. 9
- With 180 subjects from four centers, federated CMR diagnosis achieved promising performance comparable to collaborative data sharing.The authors also report increased robustness and greater sensitivity to domain-shift effects in federatively trained models.Source: §4 · p. 9
- The study was limited to binary classification and clinician-supplied segmentation masks, and the authors call for larger datasets with more diverse centers.They note that the current pipeline still relies on expert support for segmentations.Source: §4 · p. 10
Additional information
Sean Walsh disclosed financial interests and relationships; the other authors declared no competing interests.
Additional information
- Sean Walsh reported grants and sponsored research agreements in medical imaging, artificial intelligence, data science, oncology, and respiratory medicine.The disclosure covers interests within and outside the submitted work.Source: Additional information · p. 13
- Sean Walsh disclosed a leadership position and shares in Oncoradiomics SA, plus co-invented patents submitted on the company's behalf.Source: Additional information · p. 13
- The remaining authors declared no competing interests.Source: Additional information · p. 13