Abstract
Artificial intelligence (AI) holds significant potential for streamlining chest computed tomography (CT) diagnosis of COVID-19. However, the development of robust models is often hindered by the deficiency, isolation, and heterogeneity of data across diverse institutions. These factors lead to a lack of generalization, preventing clinical application. To address this, we propose the Unified CT-COVID AI Diagnostic Initiative (UCADI), a decentralized federated learning framework. In this architecture, the AI model is distributed to host institutions for local training and inference without sharing individual patient data. Our results demonstrate that while a model trained on a single site's data fails to generalize to external institutions (dropping from 97.5% to 72% sensitivity), the federated UCADI model restores performance, achieving 98% sensitivity on external datasets. This framework provides a scalable, privacy-preserving solution for global medical AI collaboration.
Introduction
The rapid spread of COVID-19 necessitated efficient diagnostic tools. While chest CT scans are a primary diagnostic modality, the manual interpretation of these scans is time-consuming and subject to radiologist fatigue. AI models offer a solution, yet they frequently encounter the "data silo" problem. Privacy regulations and the sheer volume of medical imaging data make it difficult to centralize datasets for training. Furthermore, models trained at a single site often suffer from a "generalization gap"—a sharp decline in accuracy when applied to data from different scanners or patient populations.
This paper introduces UCADI (Unified CT-COVID AI Diagnostic Initiative), a federated learning-based engine designed to facilitate collaborative AI development without compromising patient privacy.
The UCADI Framework and Methodology
The core of UCADI is a decentralized architecture that allows multiple institutions to contribute to a global model while keeping their raw data behind local firewalls.
Federated Learning (FL)
Unlike traditional centralized training, UCADI utilizes Federated Learning. In this setup, the AI model is sent to each participating hospital. Training occurs locally on the hospital’s own hardware. Only the model gradients—the mathematical updates learned from the data—are transmitted back to a central server.
Privacy and Encryption
To ensure that sensitive information cannot be reconstructed from these gradients, the framework implements Learning with Errors (LWE)-based additively homomorphic encryption. This ensures that the central server only sees encrypted updates, which are then aggregated to refine the "Global Model."
Model Architecture
The initiative utilizes a 3D-Densenet architecture. This convolutional neural network (CNN) is specifically optimized for volumetric data, allowing the model to capture spatial features across the entire 3D stack of a CT scan, rather than analyzing individual 2D slices in isolation.
Experimental Setup and Datasets
The study utilized a robust clinical dataset to train and validate the framework:
- Primary Training Set: 5,732 CT images from 1,276 individuals across three branches of Tongji Hospital in Wuhan. This dataset included COVID-19 cases, other viral pneumonias (e.g., Influenza A, RSV), bacterial pneumonia, and healthy controls.
- External Validation Set: 507 COVID-19 cases from Wuhan Union Hospital (WU) were used to test the model's ability to generalize to a new clinical environment with different imaging protocols.
Results and Discussion
The Generalization Cliff
Initial experiments highlighted the critical need for federated approaches. An AI model trained exclusively on Tongji Hospital data performed exceptionally well on its internal test set (97.5% sensitivity). However, when tested on the Wuhan Union Hospital dataset, sensitivity plummeted to 72%. This 25.5% drop illustrates the "generalization cliff" caused by data heterogeneity between institutions.
Bridging the Gap with UCADI
By implementing the federated model—which integrated insights from both Tongji and Wuhan Union without transferring raw data—the sensitivity on the external Wuhan Union test cases rose to 98%. This performance was achieved while maintaining high accuracy on the original Tongji data, proving that collaborative training significantly enhances model robustness.
AI vs. Human Performance
In a comparative study, the AI model achieved a sensitivity of 97.5%, significantly outperforming a group of six experienced radiologists who averaged 79% sensitivity. This suggests that the UCADI engine can serve as a vital "second opinion," particularly in high-pressure clinical settings where the risk of false negatives is high.
Multi-Class and Severity Assessment
The model demonstrated versatility beyond binary diagnosis. It successfully distinguished COVID-19 from other viral pneumonias (AUC 0.88) and bacterial pneumonia (AUC 0.91). Additionally, the model was able to categorize the severity of COVID-19 infections into three levels with high accuracy (AUC 0.92–0.95), providing clinicians with actionable prognostic information.
Conclusion
The UCADI framework demonstrates that federated learning can effectively overcome the barriers of data isolation and heterogeneity in medical AI. By achieving near-perfect sensitivity on external datasets without compromising patient privacy, UCADI provides a blueprint for global collaborative diagnostics. Future work will focus on optimizing hyperparameter tuning for local iterations and addressing non-IID (independent and identically distributed) data challenges through dynamic weighted averaging. The UCADI platform remains open for global participation, aiming to create a validated, world-class tool for pandemic response and beyond.
Summary updated