Abstract
Early detection of COVID-19 based on chest CT enables timely treatment of patients and helps control the spread of the disease. We proposed an artificial intelligence (AI) system for rapid COVID-19 detection and performed extensive statistical analysis of CTs of COVID-19 based on the AI system. We developed and evaluated our system on a large dataset with more than 10 thousand CT volumes from COVID-19, influenza-A/B, non-viral community acquired pneumonia (CAP) and non-pneumonia subjects. In such a difficult multi-class diagnosis task, our deep convolutional neural network-based system is able to achieve an area under the receiver operating characteristic curve (AUC) of 97.81% for multi-way classification on test cohort of 3,199 scans, AUC of 92.99% and 93.25% on two publicly available datasets, CC-CCII and MosMedData respectively. In a reader study involving five radiologists, the AI system outperforms all of radiologists in more challenging tasks at a speed of two orders of magnitude above them. Diagnosis performance of chest x-ray (CXR) is compared to that of CT. Detailed interpretation of deep network is also performed to relate system outputs with CT presentations.
Introduction
Early in the COVID-19 pandemic, the "gold standard" RT-PCR testing faced significant hurdles, including low sensitivity in patients with low viral loads and severe supply shortages. While chest CT scans offer a high-sensitivity alternative for detecting lung lesions, they present a different bottleneck: a single CT volume contains hundreds of slices, making manual review by radiologists time-consuming. Furthermore, COVID-19 shares many visual characteristics with other respiratory illnesses, such as influenza and non-viral community-acquired pneumonia (CAP). Distinguishing between these "look-alike" diseases is a high-stakes challenge that requires significant expertise and time—resources that are often stretched thin during an outbreak. This study presents a deep learning-based AI system capable of diagnosing COVID-19 from chest CT scans with human-surpassing accuracy and remarkable speed.
Methods and Materials
The researchers developed a multi-stage AI pipeline designed for both accuracy and interpretability. The process begins with lung segmentation using a U-Net architecture to isolate lung regions and remove background noise. Following this, individual 2D slices are classified using a ResNet-152 backbone into four categories: COVID-19, Influenza, CAP, or non-pneumonia. A custom fusion module then aggregates scores from all slices in a volume to produce a final patient-level diagnosis.
To ensure generalizability, the system was trained and validated on a massive, diverse dataset:
- Primary Dataset: 11,356 CT scans from 9,025 subjects across three centers in Wuhan, China.
- Public Benchmarks: Validation against the CC-CCII and MosMedData (Russian) datasets.
- Interpretability: The team employed Guided Grad-CAM to visualize AI focus areas and used radiomics to correlate mathematical findings with clinical phenotypes like "ground-glass opacities."
Results
The AI system achieved an Area Under the Curve (AUC) of 97.81% for multi-way classification on the main test cohort. It maintained high performance on external datasets, with AUCs of 92.99% (CC-CCII) and 93.25% (MosMedData). For COVID-19 specifically, the system reached 87.03% sensitivity and 96.60% specificity.
In a head-to-head reader study, the AI outperformed five board-certified radiologists in the most difficult tasks, such as distinguishing COVID-19 from Influenza. Crucially, the AI processed a scan in roughly 2.73 seconds, compared to an average of 6.5 minutes for a radiologist. This represents a 140x speedup, which is transformative for triage in overwhelmed hospitals. Additionally, the study quantitatively confirmed that CT is significantly more effective than chest X-ray (CXR) for COVID-19 diagnosis.
Discussion and Conclusion
By utilizing radiomics, the authors demonstrated that the AI focuses on specific features like "pleural parallel signs" and "vascular thickening," aligning with the pathological understanding of COVID-19. While the system shows immense promise, future work should focus on increasing subtype diversity in training to further reduce false positives and integrating non-imaging clinical data (e.g., symptoms, blood work) to provide a more holistic severity score. The current system provides a robust, rapid tool for clinical triage, significantly reducing the burden on radiologists during pandemic surges.
Summary updated