AI Outperforms Radiologists in Quantifying Early COVID-19 Lung Lesions, Study Finds
In the early months of the global pandemic, clinicians faced an unprecedented diagnostic dilemma: how to rapidly and accurately identify patients infected with SARS-CoV-2 when traditional testing methods were slow, inconsistent, or unavailable. While reverse transcription–polymerase chain reaction (RT-PCR) remained the gold standard for confirming infection, its limitations—ranging from false negatives to logistical bottlenecks—prompted a surge in the use of chest computed tomography (CT) as a frontline screening tool. Yet interpreting the subtle, often atypical imaging patterns of early-stage COVID-19 placed immense cognitive strain on radiologists already overwhelmed by surging caseloads.
Enter artificial intelligence. A new study published in the Journal of Hubei University of Medicine demonstrates that AI-powered analysis of initial CT scans not only matches but significantly surpasses human performance in quantifying key pathological features of COVID-19 pneumonia. Conducted by a team led by Shuming Liu and Guangbin Chen at the Radiology Imaging Center of Shiyan Renmin Hospital—a teaching affiliate of Hubei University of Medicine—the research offers compelling evidence that AI can transform CT from a qualitative imaging modality into a precise, data-driven diagnostic and prognostic instrument.
The study retrospectively analyzed 66 patients with laboratory-confirmed SARS-CoV-2 infection who underwent their first chest CT scan within seven days of a positive nucleic acid test. All patients met the diagnostic criteria outlined in China’s “Diagnosis and Treatment Protocol for Novel Coronavirus Pneumonia (Trial Version 7).” Crucially, individuals with pre-existing lung conditions—such as tuberculosis, pulmonary congestion, tumors, or severe emphysema—that could confound image interpretation or interfere with AI segmentation were excluded, ensuring a cleaner dataset focused solely on acute viral pathology.
Using a GE LightSpeed 16-slice multidetector CT scanner, the team acquired high-resolution volumetric scans under standardized protocols: 120 kV tube voltage, 190 mA current, 5.0 mm slice thickness, and lung window settings (width 1200 HU, level –500 HU). These DICOM images were then processed through the Yitu Lung Intelligent AI-assisted diagnostic system, developed by Shanghai-based Yitu Technology—a company that rapidly pivoted its medical imaging platform to address the pandemic.
The AI system automatically segmented and classified lung lesions into two primary categories: ground-glass opacities (GGOs) and consolidations. GGOs appear as hazy, increased lung density that does not obscure underlying bronchial or vascular structures—a hallmark of early interstitial inflammation. Consolidations, by contrast, represent denser, more opaque regions where alveolar airspaces are filled with fluid, cells, or debris, often signaling disease progression. The software generated color-coded 3D reconstructions: blue for GGOs, orange for consolidations, and purple for mixed patterns. More importantly, it output precise volumetric measurements in cubic centimeters for each lesion type in both the left and right lungs.
Three experienced radiologists independently reviewed all scans, blinded to the AI results. In cases of disagreement, a senior chief radiologist provided final adjudication. The human assessments were then compared against the AI’s quantitative outputs using rigorous statistical methods, including paired t-tests and chi-square analyses, with significance set at P < 0.05.
The findings were striking. The AI system successfully quantified total lesion volume in all 66 cases (100% success rate), whereas human readers were unable to provide any volumetric data—a task considered impractical in routine clinical practice due to its time-consuming and subjective nature. Moreover, the AI correctly identified the specific lung lobes involved in 65 out of 66 cases (98.5%), compared to only 42 cases (63.6%) for human readers. The difference was statistically significant (P < 0.001), underscoring AI’s superior spatial mapping capability.
Perhaps the most clinically relevant discovery concerned laterality. The study revealed that right lung involvement was consistently and significantly greater than left lung involvement across all metrics. The mean GGO volume in the right lung was 79.03 ± 116.11 cm³ versus 49.80 ± 86.69 cm³ in the left (P < 0.05). Similarly, consolidation volume was 20.41 ± 40.87 cm³ on the right versus 11.74 ± 27.00 cm³ on the left (P < 0.05). Total lesion volume followed the same pattern: 99.54 ± 152.64 cm³ versus 61.70 ± 110.36 cm³ (P < 0.05). This asymmetry aligns with prior clinical observations but had not been rigorously quantified until now.
The researchers hypothesize two anatomical explanations for this right-sided predominance. First, the right main bronchus is shorter, wider, and more vertically oriented than the left, potentially facilitating greater deposition of inhaled viral particles in the right lung—particularly the lower lobe. Second, the right hemidiaphragm is elevated due to the underlying liver, which may alter regional ventilation-perfusion dynamics and increase susceptibility to inflammatory infiltration.
Beyond volume and location, the AI system generated density histograms—graphs plotting the frequency of pixels (y-axis) against CT attenuation values in Hounsfield Units (x-axis). These histograms provided a novel, objective biomarker of disease severity. In mild cases, the histogram was dominated by blue (GGO) pixels clustered in the –600 to –300 HU range, characteristic of interstitial edema and partial alveolar filling. In critical cases, a pronounced shift toward higher attenuation values (–200 to +100 HU) reflected extensive consolidation, with orange and purple pixels dominating the histogram. This quantitative shift correlated strongly with clinical severity grading based on national guidelines, achieving an 85.35% concordance rate with assessments by two independent pulmonologists.
This capability to objectively stratify disease severity has profound implications for triage and resource allocation. During pandemic surges, when ICU beds and ventilators are scarce, an AI tool that can rapidly flag high-risk patients based on CT-derived biomarkers could save lives by ensuring timely escalation of care. Conversely, it could reassure clinicians managing mild cases in outpatient or isolation settings, reducing unnecessary hospitalizations.
The study also highlights AI’s potential for longitudinal monitoring. By comparing serial CT scans, the system can track changes in lesion volume and density over time, offering a dynamic view of disease evolution or treatment response. One illustrative case showed a 56-year-old patient whose follow-up scan revealed marked reduction in both GGO and consolidation volumes—a visual and quantitative confirmation of clinical improvement that might be missed or underestimated in subjective human review.
Critically, the research addresses a key limitation of RT-PCR: its imperfect sensitivity. In the early pandemic, false-negative rates of up to 30% were reported, leading to delayed diagnosis and inadvertent community transmission. Chest CT, even when interpreted by humans, demonstrated higher sensitivity in symptomatic patients. But AI-enhanced CT goes further—it doesn’t just detect abnormalities; it characterizes them with reproducible, operator-independent metrics. This is especially valuable for asymptomatic or paucisymptomatic carriers, who may present with subtle imaging findings easily overlooked during visual inspection.
The advantages extend beyond accuracy to efficiency. In high-volume settings like emergency departments or fever clinics, AI can process a full chest CT in seconds, flagging suspicious cases for urgent review while clearing low-risk scans. This not only accelerates diagnosis but also reduces radiologist burnout—a growing concern during prolonged public health crises.
Of course, AI is not a replacement for clinicians. The study emphasizes that AI serves as a decision-support tool, augmenting—not supplanting—human expertise. Final diagnosis still requires integration of imaging findings with clinical history, exposure risk, laboratory data, and epidemiological context. Moreover, the algorithm was trained and validated on a specific population from a single center in central China. Its generalizability to other ethnicities, healthcare systems, or viral variants remains to be tested in multi-center, prospective trials.
Nonetheless, this work represents a significant step toward precision radiology in infectious disease. By converting subjective visual impressions into objective, quantifiable data, AI enables a more granular understanding of disease pathophysiology. The observed right-lung predominance, for instance, may inform future research into regional immune responses or biomechanical stress in the lungs during viral infection.
As the world prepares for future pandemics—or potential resurgences of SARS-CoV-2—tools like this Yitu-powered platform could become indispensable components of public health infrastructure. They offer a scalable, rapid, and accurate means of screening, triaging, and monitoring large populations, particularly in resource-limited settings where access to molecular testing is constrained.
The implications also extend beyond COVID-19. The same AI architecture could be adapted to other forms of pneumonia, interstitial lung diseases, or even early-stage lung cancer, where precise lesion quantification is equally critical. The density histogram, in particular, emerges as a versatile biomarker that could standardize severity assessment across a spectrum of pulmonary conditions.
In conclusion, the study by Liu, Zeng, Ao, Xie, He, and Chen demonstrates that AI-driven CT analysis provides superior quantification, localization, and severity stratification of early COVID-19 lung lesions compared to conventional radiological interpretation. By delivering objective, reproducible metrics at scale, this technology enhances diagnostic confidence, optimizes clinical workflows, and ultimately supports better patient outcomes during infectious disease outbreaks.
Authors: Shuming Liu¹,²,³, Zhaojun Zeng²,³, Feng Ao²,³, Xingwu Xie²,³, Huan He²,³, Guangbin Chen¹,²,³
Affiliations:
¹Postgraduate Training Base, Renmin Hospital, Jinzhou Medical University
²Department of Radiography, Renmin Hospital, Hubei University of Medicine
³Institute of Radiological Imaging, Renmin Hospital, Hubei University of Medicine, Shiyan, Hubei 442000, China
Published in: Journal of Hubei University of Medicine, 2021, 40(3): 273–276, 280
DOI: 10.13819/j.issn.2096-708X.2021.03.012