Source-linked AI summary
Large-Scale Screening of COVID-19 from Community Acquired Pneumonia using Infection Size-Aware Classification
Feng Shi, Liming Xia, Fei Shan, Dijia Wu, Ying Wei, Huan Yuan, Huiting Jiang, Yaozong Gao, He Sui, Dinggang Shen
TL;DR
The study targets rapid COVID-19 screening from CAP, where RT-PCR can have limited detection and infection-size distributions are imbalanced. It combines automated CT segmentation, location-specific features, and infection-size-aware random forests, achieving strong overall discrimination under five-fold cross-validation. Performance difficulties remain for very small and very large infections, and the study is limited to baseline CT findings without clinical characteristics or pneumonia subtype severity.
Problem
COVID-19 must be screened from CAP, but RT-PCR has a reported detection rate around 30–60% and imbalanced infection-size distributions can weaken classification of medium-sized infections.
Method
The framework automatically segments infected lesions and lung fields, extracts location-specific features, and uses iSARF to route patients by infection size before group-specific random-forest classification.
Results
The proposed method achieved sensitivity 0.907, specificity 0.833, and accuracy 0.879 under five-fold cross-validation, with an overall highest AUC of 0.942.
Takeaways & Limitations
The framework could assist clinical decision making as a CT-based approach for screening COVID-19 from CAP.
Takeaways & Limitations
The study used baseline CT findings without pneumonia-related clinical characteristics, follow-up scans, symptom severity, or differential pneumonia subtypes.
Abstract
from arXiv · showhide
The worldwide spread of coronavirus disease (COVID-19) has become a threatening risk for global public health. It is of great importance to rapidly and accurately screen patients with COVID-19 from community acquired pneumonia (CAP). In this study, a total of 1658 patients with COVID-19 and 1027 patients of CAP underwent thin-section CT. All images were preprocessed to obtain the segmentations of both infections and lung fields, which were used to extract location-specific features. An infection Size Aware Random Forest method (iSARF) was proposed, in which subjects were automated categorized into groups with different ranges of infected lesion sizes, followed by random forests in each group for classification. Experimental results show that the proposed method yielded sensitivity of 0.907, specificity of 0.833, and accuracy of 0.879 under five-fold cross-validation. Large performance margins against comparison methods were achieved especially for the cases with infection size in the medium range, from 0.01% to 10%. The further inclusion of Radiomics features show slightly improvement. It is anticipated that our proposed framework could assist clinical decision making.
I. INTRODUCTION
The paper addresses the need to screen COVID-19 from CAP using CT-based machine learning, motivated by limitations of RT-PCR and imbalanced infection-size distributions. It proposes size-aware classification with location-specific features and evaluates the approach on a large, multi-center dataset.
- COVID-19 screening from CAP is important for patient triage, treatment protocol design, and follow-up evaluation.
- RT-PCR has a reported detection rate of around 30–60%, so repeated testing may be needed and negative results cannot rule out infection.
- Automated machine-learning-assisted COVID-19 screening remains limited compared with studies describing typical CT radiographic signs.
- Imbalanced infection-size distributions may produce high overall performance while yielding low sensitivity and specificity for medium-sized infections.
- The proposed approach uses infection-size-aware evaluation, location-specific features tailored to radiographic appearance, and a large multi-center dataset spanning patients aged 12 to 98 years.
A. Participants and image acquisition
The study retrospectively assembled thin-section CT data from 2685 participants with confirmed COVID-19 or CAP across three hospitals. Multiple scanner models and standardized acquisition parameters were represented.
- 2685 participants were included: 1658 confirmed COVID-19 cases and 1027 CAP cases.
- The dataset combined patients from Tongji Hospital, Shanghai Public Health Clinical Center, and China-Japan Union Hospital.
- All patients underwent chest CT with thin sections, using reconstructed slice thicknesses ranging from 0.625 to 2 mm.
- CT acquisition used scanners from UIH, GE, Toshiba, Siemens, and Hitachi, with images de-identified before analysis.
B. Preprocess of Lung Infections and Fields
The preprocessing pipeline automatically segments infected lung regions and lung fields from thin-section CT images. These masks support location-specific handcrafted features describing infection distribution across anatomical regions.
- Infected lung regions and bilateral lung fields were automatically segmented to support handcrafted feature extraction.
- VB-Net, a V-Net variant with bottleneck layers, was used to segment infected lesions and lung fields.
- The segmentation software achieved a Dice similarity coefficient of 92% between automated and manual infection segmentations.
- Figure 1 illustrates original lung images alongside infected-lesion and lung-field segmentations for three CAP and three COVID-19 patients.
C. Infection Size Distribution and Proposed Size-Aware
The method defines infection size relative to total lung volume and uses size-aware grouping to address polarized COVID-19 and CAP distributions. iSARF routes cases through size groups before group-specific random-forest classification.
- Infection Size Distribution: Infection size is defined as infected-region volume divided by whole segmented-lung volume.
- Infection Size Distribution: COVID-19 cases comprise 96% (538 vs. 20) of patients in the >10% infection-size group, whereas CAP comprises 91% (309 vs. 30) in the <0.01% group.
- Infection Size Distribution: Exponential infection-size intervals were used to separate patients into groups because infection size does not vary linearly with disease progress.
- Proposed Size-Aware Method: iSARF uses infection size as the only feature in a three-level decision tree that assigns cases to four size groups.
- Proposed Size-Aware Method: A separate random forest is trained for each size group, and test cases are routed to the corresponding forest for final diagnosis.
D. Extraction of Location-Specific Features
The framework extracts handcrafted CT features from infection regions and anatomically defined lung fields, covering volume, lesion number, histogram, and surface characteristics. These features encode regional distribution, bilateral asymmetry, lesion morphology, and distances to lung boundaries.
- Volume features: Volume features quantify infection and lung volumes or percentages overall, by lobe, and by pulmonary segment.
- Infected lesion number: Lesion-number features capture multifocal involvement across bilateral lungs, lobes, and pulmonary segments.
- Histogram distribution: Histogram features characterize infection distributions associated with differing pneumonia manifestations.
- 96 handcrafted features comprise 26 volume, 31 number, 32 histogram, and 7 surface features.
- Surface area: Surface features measure infection-to-lung-boundary distances across five ranges, including 3, 6, 9, 12, and 15 voxels.
E. Feature Selection and Prediction
The study selects interpretable clinico-radiological features before training multiple classifiers and evaluates them with ROC-based metrics. Performance is also examined according to infected lesion size.
- Feature selection: LASSO selects an optimal subset of clinico-radiological features for classification while providing variable importance and interpretability.
- Prediction: Selected features are supplied to logistic regression, SVM, neural network, and iSARF models for disease diagnosis.
- Evaluation: ROC analysis evaluates sensitivity, specificity, and AUC, while results are additionally investigated across infected-lesion-size groups.
III. RESULTS
Results evaluate 2685 subjects using five-fold comparisons and analyses stratified by infection size. The proposed method performs strongly overall, while small lesions remain difficult to detect and performance varies across size groups.
- Dataset: 2685 subjects comprise 1658 COVID-19 patients and 1027 CAP patients.
- Overall performance: The proposed iSARF achieved sensitivity 0.907, specificity 0.833, accuracy 0.879, and the highest AUC of 0.942 under five-fold cross-validation.
- Size-stratified performance: All methods showed low sensitivity for infections smaller than 0.01% and low specificity for infections larger than 10%.
- Size-stratified performance: Accuracy increased from 0.753 to 0.819 and 0.907 across the 0.01–0.1%, 0.1–1%, and 1–10% groups.
- Selected features: The most frequently selected features were mainly segment-level volume and number features, broad histogram bins, and short-range surface features at 3 and 6 voxels.
IV. DISCUSSION
The study presents a size-aware, location-specific machine-learning framework for screening COVID-19 from CAP using CT, with promising performance and potential clinical deployment. The authors also identify dataset and evaluation limitations, including missing clinical characteristics and reliance on baseline scans.
- Performance: 90.7% sensitivity, 83.3% specificity, and 87.9% accuracy were achieved for COVID-19 screening from CAP using CT-based machine learning.These results were obtained in the study’s five-fold experiments.
- Infection-size awareness: Infection-size imbalance between COVID-19 and CAP patients motivated a framework that separates cohorts into size groups before classification.Five-fold experiments produced consistent thresholds around 0.01%, 0.3%, and 7% when separating the dataset into four groups.
- Feature analysis: Radiomics features alone performed worse than handcrafted features, while combining both feature types slightly improved accuracy.Radiomics alone was 2.4% lower in accuracy, whereas the combined approach improved accuracy by 0.6%.
- Limitations: The study lacked pneumonia-related clinical characteristics, follow-up CT findings, symptom severity, and differential diagnoses of pneumonia subtypes.The authors identify these omissions as areas requiring further investigation.
- Clinical implementation: The method was integrated into uCloud and made available to more than 20 clinical facilities in China.Related research portals were deployed in over 50 hospitals, including more than 20 in Wuhan.
- Conclusion: The authors conclude that infection-size bias should be considered in both method development and result evaluation.They describe CT imaging combined with machine learning as a potentially efficient tool for COVID-19 screening.
Supplementary Material for “Large-Scale Screening of COVID-19 from
The supplementary material evaluates the proposed method across decision-tree size groups and feature sources. It compares handcrafted, Radiomics, and combined features using ROC curves and overall and size-group performance metrics.
- Decision-tree analysis: The proposed method uses three-level decision trees in each fold to split subjects according to infected-lesion size.The infected-lesion size was the only feature used to determine split thresholds.
- Feature sources: 93 Radiomics features were automatically extracted from infected lesions, including first-order intensity statistics and texture features.The texture features included gray-level co-occurrence, run-length, size-zone, and neighborhood graytone difference matrices.
- Feature comparison: 2% higher AUC was obtained with handcrafted features than with Radiomics features alone, while combining them increased AUC by 0.4%.The combined method had 0.4% lower sensitivity, 2.1% higher specificity, and 0.6% higher accuracy than handcrafted features alone.
- Size-group analysis: Around 2% accuracy improvement from combining features was observed in the 0.01~0.1% and 1~10% infection-size groups.The improvement was reported in the size-group analysis.
- Supplementary evaluation: The supplementary figures report mean ROC curves across five folds, overall performance, and performance after dividing cases into five infection-size groups.Reported metrics include sensitivity, specificity, accuracy, and area under the curve.