Abstract
Background: The use of CT imaging enhanced by artificial intelligence to effectively diagnose COVID-19, instead of or in addition to reverse transcription-polymerase chain reaction (RT-PCR), can improve widespread COVID-19 detection and resource allocation. Methods: 904 axial lung window CT slices from 338 patients in 17 countries were collected and labeled. The data included 606 images from COVID-19 positive patients, 224 images of other pulmonary diseases, and 74 images of normal patients. We developed an object detection model to identify ground-glass opacities (GGOs) for COVID-19, GGOs for non-COVID-19 diseases, and features inconsistent with COVID-19. These features were processed by an interpretable decision tree model. Results: On an independent test of 219 images, the model achieved 96.80% accuracy and 98.33% sensitivity. High performance was also maintained for asymptomatic patients (97.06% accuracy). Conclusion: We present an interpretable AI tool that differentiates COVID-19 GGOs from similar pathologies, suggesting that GGOs can be disease-specific.
Introduction
Early in the COVID-19 pandemic, the gold standard for diagnosis was the RT-PCR test. However, RT-PCR faced significant hurdles, including limited availability, slow turnaround times, and a false-negative rate ranging from 10% to 30%. While Computed Tomography (CT) scans offered a more sensitive alternative, they presented a different challenge: the radiological presentation of COVID-19 often overlaps significantly with other viral pneumonias.
Furthermore, most AI tools developed to bridge this gap have been "black boxes." While these models might output a diagnosis, they fail to explain the underlying reasoning. For clinicians, a diagnosis without visual evidence or a clear logic path is difficult to trust. This research addresses two major hurdles: improving the specificity of CT-based diagnosis and making the AI’s decision-making process transparent and interpretable.
Methodology
The researchers developed a two-stage "transparent" pipeline that combines object detection with a logical decision framework:
- Object Detection (YOLOv3): Instead of analyzing the image as a whole, the model uses the YOLOv3 (You Only Look Once) architecture to identify and localize specific radiological features. The model was trained to recognize three distinct categories:
- COVID-19 GGOs: Ground-glass opacities specifically associated with COVID-19.
- Non-COVID-19 GGOs: Opacities associated with other lung diseases.
- Disqualifying (DQ) Features: Radiological signs inconsistent with COVID-19, such as cavitations or pleural effusions.
- Interpretable Decision Tree: The features detected by YOLOv3 are fed into a decision tree. This model follows a human-verifiable logic path—for example, if COVID-specific GGOs are present and disqualifying features are absent, the model suggests a positive COVID-19 diagnosis.
To optimize performance with a specialized medical dataset, the team utilized transfer learning, initializing the model with weights from the COCO dataset.
Data and Validation
The study utilized a diverse, global dataset to ensure the model could generalize across different populations and hospital protocols. Data consisted of 904 axial lung window CT slices from 338 patients across 17 countries. The composition included 606 images from RT-PCR-confirmed COVID-19 patients, 224 images representing various other pulmonary diseases, and 74 images from healthy patients.
To ensure high-quality training data, regions were labeled by two upper-level radiology residents and confirmed by a board-certified radiologist with 10 years of experience. This rigorous labeling process allowed the model to learn the subtle phenotypical nuances of different lung pathologies.
Results and Discussion
The model demonstrated exceptional performance across several metrics on an independent test set of 219 images:
- Accuracy: 96.80% (95% CI, 96.75 to 96.86)
- Sensitivity: 98.33% (95% CI, 98.29 to 98.40)
- Specificity: 94.95% (95% CI, 94.84 to 95.05)
- AUC-ROC: 0.9664
Notably, the model maintained a 97.06% accuracy when tested on 34 images from asymptomatic patients. This suggests that AI-enhanced CT could be a powerful tool for identifying "silent" spreaders who show no clinical symptoms but have incidental CT findings. Additionally, the system is highly efficient, processing an image in less than 20 milliseconds on a standard Tesla P100 GPU.
The Specificity of GGOs
A key insight from this study is that ground-glass opacities, conventionally viewed as non-specific markers of inflammation, may actually be disease-specific. The AI's ability to differentiate between COVID-19 GGOs and non-COVID-19 GGOs suggests there are subtle physical characteristics in these opacities that the human eye may struggle to categorize consistently, but which a trained model can identify.
Conclusion and Future Work
This interpretable AI tool provides a "second pair of eyes" for radiologists, offering both a diagnosis and the visual evidence required to support it. By generating tight bounding boxes around relevant features, the model moves away from the "black box" paradigm toward a collaborative clinical tool.
Future work will focus on expanding the dataset to include rarer pulmonary conditions and exploring multimodal integration. Combining AI-driven CT analysis with Electronic Health Records (EHR) could further refine diagnostic accuracy and provide a more robust paradigm for managing respiratory pandemics.
Summary updated