TL;DR
This paper introduces Federated Voxel Scene Graph Generation (Fed-VSGG) to improve the detection and understanding of intracranial hemorrhages (ICH) by training models across multiple hospitals without sharing sensitive patient data. Their federated approach significantly enhances the model's ability to recall clinically relevant relationships between ICH and surrounding brain structures, outperforming traditional centralized methods, especially in the face of diverse data distributions.
What Problem Does This Paper Address?
Intracranial Hemorrhage (ICH) is a life-threatening condition with diverse manifestations that vary significantly across different clinical centers. While deep learning models have shown promise in detecting ICH, they often struggle to generalize to new datasets due to these variations and privacy concerns that limit data sharing. Traditional methods trained on single datasets fail to capture the full spectrum of ICH presentations, leading to suboptimal performance when applied to data from different sources. This gap hinders the development of robust AI tools that can reliably assist clinicians in diagnosing and managing ICH patients across diverse healthcare settings.
Methods
The authors propose Federated Voxel Scene Graph Generation (Fed-VSGG), a novel approach that combines Federated Learning (FedL) with Voxel Scene Graph Generation (V-SGG). The core idea is to train models collaboratively across multiple "clients" (hospitals) without centralizing their data.
The Fed-VSGG pipeline consists of two main stages:
- Federated Object Detection: In this initial phase, a 3D Retina-UNet architecture is trained to detect key anatomical structures: the bleeding itself, the ventricle system, and the midline. This stage utilizes a novel VC-IoU loss for improved bounding box regression.
- Federated Relation Prediction: Once the object detector is robust, the focus shifts to predicting relationships between these detected objects. This involves training federated variants of V-SGG methods (Fed-MOTIF and Fed-IMP) that learn a global distribution of relation classes, making them less susceptible to local data biases. The training process involves clients computing local relation statistics, which are then aggregated by a central server.
This federated approach allows models to learn from the collective experience of multiple institutions while preserving patient privacy.
Datasets
The study utilizes four datasets for training and evaluation:
- INSTANCE2022: A publicly available dataset for ICH segmentation.
- BHSD: Another public dataset for ICH segmentation.
- CQ500: A public dataset used for ICH detection.
- Private Cohort (PC): A newly curated dataset from Germany.
The authors performed data curation to exclude images with artifacts or unclear ICH manifestations. They also modified existing ICH masks from INSTANCE2022 and BHSD to improve annotation quality and reduce bias.
Key Insights and Ideas
- Federated Learning is Crucial for Generalization: The study strongly demonstrates that FedL is essential for building models that generalize well across diverse ICH datasets. Centralized models trained on single datasets perform poorly on unseen data, highlighting the limitations of traditional approaches.
- Structured Representations Matter: Beyond simple detection, modeling the relationships between ICH and other brain structures (like the ventricle system and midline) provides a richer, clinically relevant understanding of the patient's condition.
- Privacy-Preserving Collaboration is Key: The ability to leverage data from multiple institutions without compromising patient privacy is a significant advantage of FedL, especially in the medical domain.
- Frequency Bias for Relations: Learning a global frequency bias for relation classes helps to mitigate local distribution shifts and improve the robustness of relation prediction.
Key Results
- Object Detection: Federated models (FedAvg, FedSGD) achieved comparable or superior object detection performance compared to centralized "all seen" models, with detection rates for ventricles and midline reaching over 90%.
- Relation Prediction: Fed-VSGG models demonstrated a significant improvement in recalling clinically relevant relations, recalling up to 20% more relations compared to models trained on a single centralized dataset.
- Qualitative Improvements: Figure 3 shows that the Fed-MOTIF model, trained on all datasets, provided more precise segmentation of bleedings and correctly identified ventricle system asymmetry and a small subarachnoidal bleeding, which the single-dataset trained model missed or misclassified.
- Robustness: Federated models consistently performed well across different datasets, bridging the domain gap that severely affected centralized models.
Limitations and Future Work
- Annotation Bias: While efforts were made to reduce annotation bias, the authors acknowledge that some residual bias might exist. They plan to release their annotation tools and data for further community efforts.
- Data Curation: The exclusion of certain images during data curation might have removed potentially valuable information.
- Future Work: The authors suggest that their federated approach opens avenues for developing new methods that can leverage this finer, structured information to regularize across clients more effectively. They also highlight the potential for FedL to improve clinical decision-making by providing more robust and generalizable insights into ICH.
Summary updated