TLDR: FedGIN is a Federated Learning framework that uses a Global Intensity Non-linear (GIN) augmentation module to enable accurate organ segmentation from multi-modal medical images (CT and MRI) without sharing raw patient data. It addresses challenges like data scarcity, domain shift, and privacy, showing significant performance improvements, especially for complex organs, and achieving near-centralized performance in decentralized settings.
Medical image segmentation is a vital component in modern AI-assisted diagnostics, surgical planning, and treatment monitoring. Developing accurate and robust models for this purpose is crucial, especially when dealing with diverse imaging modalities like CT and MRI. However, real-world deployment of such AI models faces significant hurdles, including limited access to large, diverse datasets, variations in image characteristics across different modalities (known as domain shift), and strict privacy regulations that prevent direct sharing of patient data.
To overcome these challenges, researchers have introduced FedGIN, a novel Federated Learning (FL) framework designed to enable multi-modal organ segmentation without the need to share sensitive raw patient data. Federated Learning is a distributed machine learning approach where models are trained locally on decentralized datasets, and only model parameters (not the data itself) are shared with a central server for aggregation. This approach inherently addresses privacy concerns.
The core innovation within FedGIN is its integration of a lightweight Global Intensity Non-linear (GIN) augmentation module. This module plays a crucial role in harmonizing the intensity distributions specific to each imaging modality during the local training phase. By applying randomized, anatomy-preserving intensity transformations, GIN encourages the model to learn features that are invariant to the specific modality, thus improving its ability to generalize across different types of medical images.
The FedGIN workflow begins with a central server broadcasting a global model to all participating clients. Each client, which might have either CT, MRI, or both types of unpaired data, then applies the GIN augmentation to its local data and trains a U-Net model – a popular architecture for image segmentation. After local training, only the updated model weights are sent back to the central server. The server then aggregates these updates using a method called FedAvg to refine the global model, and this cycle repeats over multiple communication rounds, continuously improving the model’s ability to generalize across modalities and institutions.
The effectiveness of FedGIN was rigorously evaluated using two public 3D medical imaging datasets: TotalSegmentator for training and validation, and AMOS2020 for testing. The experiments focused on segmenting five abdominal organs: liver, kidneys, spleen, pancreas, and gall bladder. Two main scenarios were explored: a limited dataset scenario and a complete dataset scenario.
In the limited dataset scenario, where the model was initially trained on MRI data and then progressively exposed to CT data, FedGIN demonstrated significant performance improvements. It achieved a 12–18% improvement in 3D Dice scores on MRI test cases compared to FL without GIN, and consistently outperformed local baseline models. This was particularly evident for low-contrast, complex organs like the spleen, gallbladder, and pancreas, where adding CT data through FedGIN substantially enhanced MRI segmentation performance.
For organs that are anatomically simpler or have more homogeneous intensity distributions, such as the liver and kidneys, the performance gain from adding CT data and GIN was more modest, as MRI-only models already performed very well. Importantly, standard Federated Learning without GIN augmentation consistently underperformed, highlighting the critical role of GIN in stabilizing performance in cross-domain federated settings.
In the complete dataset scenario, where both MRI and CT data were fully utilized across all clients, FedGIN showcased near-centralized performance. It achieved a remarkable 30% Dice score improvement over the MRI-only baseline and a 10% improvement over the CT-only baseline. This demonstrates FedGIN’s strong cross-modality generalization capabilities even under privacy constraints, effectively matching the performance of models trained with centralized access to all data.
These findings underscore that multimodal training, especially when harmonized by GIN, consistently outperforms single-modality baseline models. FedGIN’s ability to closely match centralized GIN performance validates its robustness and scalability for decentralized training environments. This framework offers a scalable and privacy-preserving alternative to traditional centralized approaches, particularly for challenging segmentation tasks in low-contrast and structurally variable organs like the pancreas, spleen, and gallbladder.
Also Read:
- Advancing Medical Image Segmentation with M3HL: A Semi-Supervised Method
- DS2Net: Enhancing Medical Image Segmentation with Combined Detail and Semantic Understanding
In conclusion, FedGIN presents a significant advancement in federated multimodal organ segmentation. By integrating unpaired CT and MRI data from multiple clients and leveraging GIN augmentation to mitigate domain shifts, it consistently enhances segmentation performance. This work paves the way for scalable, privacy-preserving, and modality-agnostic learning strategies crucial for real-world clinical applications. For more details, you can refer to the full research paper here.
