TL;DR
This study introduces a novel radiomics approach using CT scans of the adrenal glands and surrounding fat to predict COVID-19 severity. By developing an automated segmentation tool and a radiomics nomogram, the researchers demonstrate that these imaging features can help allocate healthcare resources more effectively, particularly in situations of high demand.
What Problem Does This Paper Address?
The COVID-19 pandemic highlighted the critical need for accurate and early prediction of disease severity to ensure appropriate patient care and resource allocation. While chest CT scans are standard for diagnosing COVID-19, and some studies have explored radiomics for predicting progression, the role of endocrine organs like the adrenal glands and their surrounding fat in disease exacerbation remained largely unexplored. The authors hypothesize that these organs, involved in immune and inflammatory responses, might show detectable changes related to COVID-19 severity. Furthermore, manual segmentation of these structures for radiomics analysis is time-consuming and prone to errors, necessitating an automated solution.
Methods
The study employed a multi-stage approach:
- Automated Segmentation: A cascaded V-net network was developed to automatically segment the adrenal glands from CT images. This involved a coarse segmentation network for initial localization and a fine segmentation network for precise delineation. Periadrenal fat was then obtained by inflating the segmented adrenal gland region.
- Radiomics Feature Extraction: Radiomics features were extracted from the segmented regions of interest (ROIs) using the PyRadiomics package. This process involved resampling to isotropic voxels and extracting a total of 2,264 features.
- Feature Selection: A two-step process was used to select the most informative features. First, a univariate K-best analysis identified significant features, followed by a LASSO (Least Absolute Shrinkage and Selection Operator) algorithm to further reduce dimensionality and prevent overfitting.
- Model Development:
- Radiomics Models: Three radiomics models were built: an Adrenal Gland model (AM), a Periadrenal Fat model (PM), and a Fusion model (FM) combining features from both. Logistic regression was used for model construction.
- Clinical Model (CM): A model was built using seven selected independent predictive clinical indicators.
- Radiomics Nomogram (RN): This model integrated the radiomics score from the FM with the seven clinical indicators to provide a more comprehensive prediction.
- Model Evaluation: The performance of all models was evaluated using Area Under the Curve (AUC) from Receiver Operating Characteristic (ROC) curves, sensitivity, specificity, and accuracy. DeLong's test was used to compare AUCs between models. Calibration curves and the Hosmer-Lemeshow test assessed prediction accuracy, while Decision Curve Analysis (DCA) evaluated the clinical utility.
Datasets
The study utilized a retrospective dataset of 1,245 patients diagnosed with COVID-19, collected from two hospitals in China between January and April 2020.
- Training and Validation Sets: 1,209 patients from Huoshenshan Hospital were randomly divided into a training set (80%) and a validation set (20%).
- Independent Test Set: 36 patients from the Maternal and Child Health Hospital Optical Valley Branch Hospital were used as an independent test set.
Key Insights and Ideas
- Adrenal Gland and Periadrenal Fat as Biomarkers: The study demonstrates that radiomics features from the adrenal gland and periadrenal fat on initial CT scans are significantly associated with COVID-19 disease exacerbation. This suggests these organs, often overlooked in COVID-19 imaging analysis, play a role in disease progression.
- Automated Segmentation is Feasible and Effective: The developed cascaded V-net framework successfully automated the segmentation of adrenal glands, achieving a good Dice value (average 79.02%). This overcomes the bottleneck of manual segmentation in radiomics studies.
- Radiomics Nomogram Outperforms Other Models: The integrated Radiomics Nomogram (RN), combining radiomics features with clinical indicators, showed superior predictive performance compared to models based solely on radiomics or clinical data. This highlights the synergistic value of combining different data modalities.
- Potential for Resource Allocation: The RN's strong predictive power, especially demonstrated by DCA, suggests its potential utility in identifying patients at higher risk of severe disease, thereby aiding in the fair allocation of critical healthcare resources.
Key Results
- The auto-segmentation framework achieved an average Dice value of 79.02% for adrenal glands.
- The Radiomics Nomogram (RN) achieved the highest AUC in the validation set (0.833) and test set (0.773), outperforming the clinical model (CM) and individual radiomics models (AM, PM, FM).
- The RN and FM showed significantly better predictive efficacy than the CM (P < 0.0001).
- DeLong's test indicated no statistically significant difference between the FM and RN in the validation and test sets.
- Decision curve analysis showed that the RN provided greater net benefits than FM and CM when the threshold probability for intervention was within specific ranges (0.4-0.8 in validation, 0.3-0.7 in test).
Limitations and Future Work
- Data Source: The study used chest CT images, which may limit the visualization and delineation of adrenal tissue compared to dedicated adrenal CT scans. However, the authors justify this by considering clinical practicality and radiation exposure.
- ROI Definition: The entire adrenal gland was selected as the ROI, not specific lesions. The periadrenal fat ROI might include other indistinguishable tissues.
- Future Work: The authors suggest that this approach could be extended to multi-organ fusion prediction in the future. Further research is needed to confirm whether changes in periadrenal fat are directly caused by adrenal gland inflammatory cell infiltration or indirect damage from SARS-CoV-2. The study also notes that while the RN showed good performance, there is still room for improvement in prediction calibration.
Summary updated