Federated Learning for Multi-Class 3D Cardiac Segmentation Using Convolutional and Transformer Models

Senol Piskin1, MUSAB AHMED KHAN UMAIR2, Irem Yilmaz3, Furkan Enes Beskardes1, Ömer Faruk Şahin4, Ahmed Boray5, Femilda Josephin Joseph Shobana Bai6, Doruk Gunveren6
1Istinye University, XRLab, Istanbul, Turkey, 2Istinye univeristy, 3Yildiz Technical University, 4University of Potsdam, 5Istinye university, XRLab, Istanbul turkey, 6Istinye University

Abstract

Aims: This study aimed to evaluate the performance of convolutional and transformer-based deep learning architectures for multi-class 3D cardiac segmentation under federated learning (FL) settings, using a unique pediatric cardiac imaging cohort. The investigation focused on modality-dependent behavior across CT and MRI, with particular attention to the pulmonary artery. By leveraging FL, this work addresses the scarcity and sensitivity of pediatric data across institutions while integrating XAI techniques. Methods: Three models, MedNeXt, UXNet, and UNETR, were trained for 3D cardiac segmentation in a federated learning framework with three clinical silos using FedAvg over 500 communication rounds and two local epochs per round. Modality-specific experiments were conducted on a pediatric CT dataset (48 patients; 38/5/5 train/validation/test split) and a pediatric MRI dataset (60 patients; 50/5/5 split). Segmentation targeted three cardiac structures: the aorta, left ventricle (LV), and pulmonary artery (PA). Performance was evaluated using the Dice similarity coefficient. A softmax-averaging ensemble strategy was assessed. Post hoc explainability analysis was performed using Grad-CAM and Integrated Gradients. Results: On MRI, convolutional architectures achieved superior performance, with UXNet and MedNeXt reaching Dice scores of 0.874 and 0.873, respectively, outperforming UNETR (0.848). The ensemble achieved the highest MRI performance with a Dice score of 0.887. On CT, the transformer-based UNETR outperformed convolutional models, achieving a Dice score of 0.707 compared to 0.668 (MedNeXt) and 0.638 (UXNet), while the ensemble improved performance to 0.727. Across both modalities, the pulmonary artery was the most challenging structure, with lower Dice scores (MRI: up to 0.819; CT: up to 0.523). Explainability analysis showed localized attention maps on MRI and diffuse activations in UNETR. Conclusions: Federated learning enables privacy-preserving multi-center pediatric cardiac segmentation without centralized data sharing. Performance is modality-dependent: convolutional architectures perform better on MRI, whereas transformer-based models show advantages on CT. Ensemble learning improves accuracy across both modalities.