CT-based radiomics combined with signs: a valuable tool to help radiologist discriminate COVID-19 and influenza pneumonia

Yilong Huang, Zhenguang Zhang, Siyun Liu, Xiang Li, Yunhui Yang, Jiyao Ma, Zhipeng Li, Jialong Zhou, Yuanming Jiang, Bo HeBMC Medical Imaging 2021 · DOI

TL;DR

This study introduces a novel approach combining CT-based radiomic features and traditional radiological signs to improve the differentiation between COVID-19 and influenza pneumonia. The combined model significantly outperforms individual models, offering a more accurate and potentially faster diagnostic tool for radiologists.

What Problem Does This Paper Address?

The COVID-19 pandemic presented a significant challenge in distinguishing it from other viral pneumonias, particularly influenza pneumonia, due to overlapping clinical symptoms and radiographic findings. While RT-PCR is the gold standard for diagnosis, its sensitivity can be low, leading to false negatives. Computed Tomography (CT) plays a crucial role in diagnosis, but subjective interpretation by radiologists can be influenced by experience. This study aims to develop an objective, quantitative tool to aid radiologists in differentiating these two conditions, thereby improving early diagnosis and patient management.

Methods

The researchers employed a multi-faceted approach:

  • CT Sign Analysis: Three experienced radiologists independently analyzed CT images to identify and quantify specific pneumonia signs. This involved a predefined list of 16 signs, including ground-glass opacities (GGO), consolidation, and interstitial thickening.
  • Radiomics Feature Extraction: A large set of radiomic features (1316 in total) were extracted from segmented pneumonia lesions in CT images. These features capture quantitative information about the image texture and intensity, going beyond what the human eye can easily perceive.
  • Feature Selection: To identify the most predictive features, the study utilized several statistical methods:
  • Chi-square test/Fisher exact test: To identify significant differences in CT signs between COVID-19 and influenza groups.
  • Mann-Whitney U test: To select radiomic features with significant differences.
  • mRMR (minimum redundancy maximum relevance): To reduce redundancy among radiomic features.
  • LASSO logistic regression: A regularization technique used for feature selection and model building, particularly effective for high-dimensional data.
  • Backward stepwise logistic regression with minimum AIC: To further refine the selected features and construct the final predictive models.
  • Model Development: Three models were built:
  • CT Signs Model: Based on selected independent CT signs.
  • Radiomics Model: Based on selected radiomic features.
  • Combined Model: Integrating both selected CT signs and radiomic features.
  • Validation: The models were internally validated using 1000-times bootstrap resampling to assess their reliability and generalizability. Diagnostic performance was evaluated using Receiver Operating Characteristic (ROC) analysis, calculating Area Under the Curve (AUC), sensitivity, specificity, and accuracy.

Datasets

The study retrospectively collected data from 154 patients diagnosed with viral pneumonia across 15 hospitals. This cohort comprised:

  • 89 cases of COVID-19 pneumonia
  • 65 cases of influenza pneumonia

Patients were included if they had confirmed viral pneumonia (identified by RT-PCR assays for various viruses) and underwent chest CT examination within 48 hours of admission, with symptoms onset within 7 days. Patients with multiple viral infections, bacterial or fungal infections, or incomplete data were excluded.

Key Insights and Ideas

  • Synergy of Radiomics and Signs: The most striking insight is that combining traditional CT signs with quantitative radiomic features significantly enhances diagnostic performance. This suggests that neither approach alone captures the full picture, and their integration offers a more robust predictive capability.
  • Beyond Subjective Interpretation: Radiomics offers a way to move beyond the subjective interpretation of CT images by radiologists, providing objective, quantitative data that can complement visual assessment.
  • Early Differentiation is Crucial: The study highlights the importance of early differentiation between COVID-19 and influenza pneumonia, as timely diagnosis is critical for prognosis and management.
  • Specific CT Signs as Predictors: Certain CT signs, such as peripheral lesion distribution, GGO, intralobular interstitial thickening, and halo sign, were found to be more common in COVID-19, aligning with previous observations.

Key Results

The study demonstrated the following performance metrics:

  • Radiomics Model: AUC of 0.888 (Sensitivity: 86.5%, Specificity: 78.4%, Accuracy: 83.1%)
  • CT Signs Model: AUC of 0.906 (Sensitivity: 86.5%, Specificity: 81.5%, Accuracy: 84.4%)
  • Combined Model: AUC of 0.959 (Sensitivity: 89.9%, Specificity: 90.7%, Accuracy: 90.3%)

The combined model showed a statistically significant improvement in AUC compared to either the radiomics model (P=0.002) or the CT signs model (P=0.004) alone. Calibration curves and decision curve analysis further supported the superior clinical utility and reliability of the combined model.

Limitations and Future Work

The authors acknowledge several limitations:

  • Retrospective Study Design: This inherently carries the risk of selection bias.
  • Limited Sample Size for Other Pneumonias: The study focused on COVID-19 and influenza, and did not compare characteristics with other viral pneumonias due to smaller sample sizes.
  • Host Factors: CT data alone cannot fully distinguish viral pneumonias, as host factors significantly influence the lung's response to viruses.
  • Need for Clinical Data: The authors suggest that incorporating more clinical features and laboratory examination data could further improve the predictive model's accuracy.

Future work could involve larger, prospective studies to validate these findings and explore the integration of clinical and laboratory data with CT-based radiomics for even more comprehensive diagnostic tools.

Summary updated

BibTeX
@article{huang2021ct,
  author = {Yilong Huang and Zhenguang Zhang and Siyun Liu and Xiang Li and Yunhui Yang and Jiyao Ma and Zhipeng Li and Jialong Zhou and Yuanming Jiang and Bo He},
  title = {{CT-based radiomics combined with signs: a valuable tool to help radiologist discriminate COVID-19 and influenza pneumonia}},
  journal = {BMC Medical Imaging},
  year = {2021},
  doi = {10.1186/s12880-021-00564-w},
}
Figure