The gold standard for COVID-19 diagnosis, reverse transcription polymerase chain reaction (RT-PCR), suffers from relatively low sensitivity (60-71%), leading to numerous false negatives. Chest computed tomography (CT) scans offer a potential complementary modality, revealing radiological features indicative of COVID-19. However, the accuracy of CT-based diagnosis heavily relies on radiologists' expertise, and significant inter-reader variability exists. Developing an accurate and automated CT-based diagnostic tool faces three key data-related challenges: (1) Incompleteness of available high-quality training data; (2) Isolation of data across multiple centers due to privacy concerns, preventing collaborative model improvement; and (3) Heterogeneity in CT acquisition protocols across institutions, making data integration challenging. Furthermore, the consistency of COVID-19 radiological patterns across diverse geographic regions and demographics remains an open question. These challenges necessitate a novel approach that balances the need for large-scale data with stringent privacy requirements. Federated learning, which allows distributed model training without direct data sharing, presents a promising solution to overcome these challenges.

Existing literature highlights the limitations of RT-PCR in accurately detecting COVID-19 infections. Several studies have explored the use of chest CT scans for COVID-19 diagnosis, reporting varying sensitivity levels (0.56 to 0.98). However, these studies also point to the significant inter-observer variability among radiologists in interpreting CT findings, underlining the need for automated diagnostic tools. The challenges in building robust and generalized AI models for CT-based COVID-19 diagnosis, including incomplete, isolated, and heterogeneous data, are discussed in various publications. Previous work has investigated the use of federated learning in other contexts, demonstrating its potential for privacy-preserving machine learning. However, its application to large-scale medical image analysis, particularly for a globally impactful disease like COVID-19, is relatively unexplored.

The Unified CT-COVID AI Diagnostic Initiative (UCADI) was launched to address the challenges of developing a robust AI model for COVID-19 diagnosis using chest CT scans. UCADI involved 23 hospitals (5 in China, 18 in the UK), contributing a total of 9,573 CT scans from 3,336 patients. The data from the three branches of Wuhan Tongji Hospital Group in China formed one data set, while the National COVID-19 Chest Imaging Database (NCCID) from the UK formed another. A three-dimensional convolutional neural network (3D-DenseNet) model was developed for COVID-19 identification, with a four-class classification task (COVID-19, other viral pneumonia, bacterial pneumonia, healthy) also investigated. Data preprocessing included adaptive sampling of 16 CT slices per case, standardization, and trilinear interpolation for resizing and computational efficiency. Model training utilized a weighted cross-entropy loss function due to data imbalance and an SGD optimizer with cosine annealing for learning rate scheduling. The Chinese CT scans were non-contrast, while approximately 80% of the UK scans were contrast-enhanced. To address this heterogeneity, CycleGAN, an image-to-image translation method, was used to convert contrast-enhanced CTs into non-contrast versions for augmentation during training. The performance of the 3D-DenseNet model was compared against two other 3D CNN baseline models: 3D-ResNet and 3D-Xception. Grad-CAM was employed for model interpretability, visualizing the regions of the CT scans that were most influential in the model's classification decisions. Federated learning, using the FedAvg algorithm with Learning with Errors (LWE) encryption for privacy preservation, was implemented to collaboratively train the model across participating institutions without direct data sharing. The performance of the federated model was compared to the locally trained models on the Chinese and UK datasets, also comparing with a panel of six expert radiologists. Experiments were conducted to analyze the trade-off between model performance and communication cost in the federated training process by varying the frequency of model parameter synchronization.

Locally trained 3D-DenseNet models achieved varying performances, with the Chinese model showing higher sensitivity and specificity for COVID-19 identification than the UK model (due to data heterogeneity). The application of CycleGAN improved the performance of the UK model trained on non-contrast CTs. The federated learning model substantially outperformed all locally trained models, demonstrating enhanced generalization. On the Chinese test set (1,076 CTs), the federated model achieved a sensitivity/specificity/AUC of 0.973/0.951/0.980, respectively. On the UK test set, it achieved 0.730/0.942/0.894. Grad-CAM analysis revealed that the model's predictions were based on relevant radiological features, correlating well with radiologists' annotations. The federated model showed comparable performance to a panel of six expert radiologists in differentiating COVID-19 from other pneumonia types, with the radiologists' consensus reaching a sensitivity of 0.90 and a specificity of 0.96. Analysis of the trade-off between communication frequency and model performance indicated that more frequent synchronization improved performance with only a modest increase in training time. The federated model also demonstrated superior performance on hold-out test sets from Wuhan Tianyou Hospital and Wuhan Union Hospital.

The study's findings strongly support the use of federated learning for developing AI models for COVID-19 diagnosis from chest CT scans. The superior performance of the federated model compared to locally trained models highlights the benefit of collaborative data utilization, addressing the limitations imposed by data isolation and heterogeneity. The comparable performance to expert radiologists shows that this approach can support clinical decision-making. The use of LWE encryption ensures privacy and facilitates multinational collaboration, overcoming legal and ethical barriers to data sharing. The model's interpretability, as demonstrated by Grad-CAM, enhances trust and facilitates clinical integration. This work offers a scalable and privacy-preserving framework for building robust AI diagnostic tools in the healthcare sector, particularly relevant for rapidly emerging infectious diseases.

This study introduces UCADI, a novel framework for collaborative AI development in medical imaging, successfully addressing the challenges of data scarcity, isolation, and heterogeneity in the context of COVID-19 diagnosis. The federated learning approach demonstrates superior performance to local models and achieves results comparable to human experts, while effectively preserving patient data privacy. Future work could focus on incorporating additional imaging modalities, improving model efficiency, and expanding the framework to other diseases and clinical settings. The open-sourced framework encourages wider participation and continuous model improvement.

The study acknowledges potential bias in comparing the AI model with radiologists due to geographical limitations in data access for cross-continental comparison of radiologists. The study also notes that occasional interruptions in the federated learning process occurred due to unstable internet connections and that further improvements in the computational efficiency of the 3D CNN model are possible. The uncertainty and heterogeneity in medical data remain inherent challenges requiring continuous investigation and refinement of the methodologies.