Privacy-aware deep vein thrombosis segmentation using a multi-model federated learning framework with the federated...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Science

Background:

• Deep Vein Thrombosis (DVT) diagnosis relies on precise segmentation of Computer Tomography (CT) scans.
• Existing segmentation methods may face challenges with data privacy and distributed datasets.
• Federated Learning (FedL) offers a privacy-preserving approach for training models on decentralized data.

Purpose of the Study:

• To develop and evaluate an efficient Federated Learning (FedL) architecture for DVT segmentation using CT images.
• To enhance segmentation accuracy and efficiency while maintaining data privacy and security.
• To assess the scalability and stability of the FedL framework across varying dataset sizes and model complexities.

Main Methods:

• Proposed an efficient FedL architecture utilizing the Federated Averaging (FedAvg) algorithm.
• Trained seven distinct local models on non-independent and identically distributed (Non-IID) CT images across three phases with increasing client numbers and model diversity (CNN, Sequential, Semantic, U-Net, VGG Net-19, Modified U-Net, Modified-Net).
• Aggregated local model weights to progressively improve a global model, evaluated on datasets of 1000, 2000, and 3000 samples.

Main Results:

• Significant performance gains observed with increasing dataset size, including higher Accuracy and F1-score, and decreased Tversky Loss.
• Consistent improvement across all phases, with validation loss reducing from 0.910 to 0.061.
• Framework demonstrated scalability with increased communication costs (14 MB to 3279 MB) and training time (7.67 s to 18,702 s), while preserving differential privacy and improving client heterogeneity.

Conclusions:

• The proposed FedL framework effectively enhances DVT segmentation accuracy and efficiency on CT images.
• The architecture demonstrates robust scalability, stability, and differential privacy preservation in heterogeneous environments.
• Federated Learning is a viable approach for privacy-preserving medical image analysis in distributed settings.