key: cord-0890433-be1lhdg7 authors: Zhang, M.; Yin, X.; Li, W.; Zha, Y.; Zeng, X.; Zhang, X.; Cui, J.; Tian, J.; Wang, R.; Liu, C. title: Value of radiomics features from adrenal gland and periadrenal fat CT images predicting COVID-19 progression date: 2021-01-05 journal: nan DOI: 10.1101/2021.01.03.21249183 sha: d8667aa299c4dba29ca5a6ba20988ff90d0d17b0 doc_id: 890433 cord_uid: be1lhdg7 Background: Value of radiomics features from the adrenal gland and periadrenal fat CT images for predicting disease progression in patients with COVID-19 has not been studied.Methods: A total of 1,245 patients(685 moderate and 560 severe patients)were enrolled in a retrospective study.We proposed 3D V-Net to segment adrenal glands in onset CT images automatically and periadrenal fat was obtained using inflation operation around the adrenal gland. Next, we built a clinical model(CM), three radiomics models (adrenal gland model[AM], periadrenal fat model[PM] and fusion of adrenal gland and periadrenal fat model[FM])and radiomics nomogram(RN)after radiomics features extracted to predict disease progression in patients with COVID-19. Results: The auto-segmentation framework yielded a dice value of 0.79 in the training set. CM, AM, PM, FM and RN obtained AUCs of 0.712, 0.692, 0.763, 0.791 and 0.806, respectively in the training set. FM and RN had better predictive efficacy than CM (P < 0.0001) in the training set. RN showed that there was no significant difference in the validation set(mean absolute error [MAE] = 0.04) and test set (MAE = 0.075) between predictive and actual results. Decision curve analysis showed that if the threshold probability was more than 0.3 in the validation set or between 0.4 and 0.8 in the test set, it could gain more net benefits using RN than FM and CM. Conclusion: Radiomics features extracted from the adrenal gland and periadrenal fat CT images may predict progression in patients with COVID-19. hypothesized that adrenal gland changes could occur due to its adaptive mechanism and SARS-CoV-2 invasion. The periadrenal fat was bound to be affected as the closest tissue around the adrenal gland. Fundamental changes can be detected by extracting radiomics features from chest CT images using AI technology. Radiomics is an emerging technique that can convert images difficult for human eyes to distinguish into high-throughput quantitative features that may reflect potential pathological and physiological states (11) . Manual segmentation is the first and key step of successfully constructing a radiomics model, but it is time-consuming and error-prone because of repeated labor. Combining the auto-segmentation framework based on AI technology with manual revision improves both efficiency and quality (12) . Cascaded V-Nets have been used to segment brain tumors with good results (13), which could be further improved if incorporated with extra information, such as coarse localization. A total of 1,245 patients (685 moderate and 560 severe patients) were enrolled in the retrospective study from two hospitals (1,209 patients from one and 36 patients from the other). The data of 1,209 patients were formed as our training and validation sets, including 672 patients who suffered from moderate COVID-19 and 537 who suffered from severe COVID-19. The patient characteristics in training and validation sets are listed in Table 1 The auto-segmentation framework construction, radiomics feature extraction, feature selection, and predicting model building were established on the Research Portal V1.1 (Shanghai United Imaging Intelligence, Co., Ltd.). For adrenal gland segmentation, we manually delineated bilateral adrenal glands from the CT images of 315 patients, 265 of them were used for training. The remaining data from 50 patients were used to evaluate the performance and the segmentation model yielded average Dice values of 79.48% for the left adrenal gland and 78.55% for the right adrenal gland. The entire adrenal gland achieved an average All rights reserved. No reuse allowed without permission. perpetuity. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in The copyright holder for this this version posted January 5, 2021. ; https://doi.org/10.1101/2021.01.03.21249183 doi: medRxiv preprint Dice value of 79.02%. Representative auto-segmentation results are shown in Figure 1 . We also visually verified the segmentation results, and they were considered satisfactory based on radiologists' judgment. The segmentation algorithm was then used to segment all the remaining data automatically. In the training set, K-best and least absolute shrinkage and selection operator (Table 1) . Next, 17 clinical factors and serum All rights reserved. No reuse allowed without permission. perpetuity. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in biomarkers were selected using univariate logistic regression analysis, and 7 indicators, LDH, L, HB, DD, WBC, TT, and TP, were selected using multivariate logistic regression analysis. The relationship between RadScore from FM used in the construction of radiomics nomogram (RN) and 30 clinical factors plus serum biomarkers were analyzed using Pearson correlation between training, validation, and test sets (Figure 2) . The difference in RadScores with clinical factors or serum biomarkers was not significant. The radiomics information extracted from onset CT images belonged to another dimension, and this information was not affected by clinical factors and serum biomarkers. Then, 17 clinical factors and serum biomarkers were selected using univariate logistic regression analysis, and 7 indicators, LDH, L, HB, DD, WBC, TT, and TP, were selected using multivariate logistic regression analysis. perpetuity. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in The copyright holder for this this version posted January 5, 2021. Table 2 ). Box plots summarizing the RadScores and coefficients of seven clinical indicators in training, validation, and test sets directly demonstrate the difference between RadScore and coefficients of seven clinical indicators between the moderate and severe patient sets ( Figure 4 ). Multivariate analysis revealed that RadScore and seven clinical indicators were significant independent factors predicting disease progression in patients with COVID-19. Using collinearity diagnosis, variance inflation factor (VIF) for the radiomics score and seven clinical indicators were from 1.007 to 1.191, indicating no severe collinearity in these factors. Next, we used the RadScore from FM combined with seven clinical indicators to construct the RN to assess disease progression in patients with COVID-19 ( Figure 5 ). The RN showed satisfactory performance for predicting and assessing progression in patients with COVID-19 with AUC of 0.806 (95% CI, 0.780 to 0.831) in the training set, 0.833 (95% CI, 0.780 to 0.878) in the validation set, and 0.773 (95% CI, 0.603 to 0.895) in the test set ( Figure 3 , Table 2 ). DeLong's test was used to compare the AUCs of three radiomics models, CM and RN, in the training set. The result showed that the RN and FM were significantly better than CM (P < 0.0001). The difference between FM and RN was not statistically significant (P = 0.233) in the validation and test sets. All rights reserved. No reuse allowed without permission. perpetuity. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in The copyright holder for this this version posted January 5, 2021. ; https://doi.org/10.1101/2021.01.03.21249183 doi: medRxiv preprint The calibration curve showed an agreement between the predicted and actual values. The Hosmer-Lemeshow test was not significant in the validation set (mean absolute error [MAE] = 0.075) or test set (MAE = 0.04), which suggests that there was no significant departure from actual values ( Figure 6 ). Decision curve analysis (DCA) was used to evaluate the performance of RN (Figure 7 ). If the threshold probability was more than 0.3 in the validation set, the RN could get more net benefits than FM and CM. If the threshold probability was between 0.4 and 0.8 in the test set, RN can still get more net benefits than FM and CM. In this study, an adrenal gland auto-segmentation framework was constructed. We rapidly finished the auto-segmentation of adrenal glands and periadrenal fat using the framework. We found several radiomics features and seven clinical indicators related to disease progression in patients with COVID-19. Next, we built and validated a RN for predicting disease progression based on radiomics features extracted from the onset of the adrenal gland and periadrenal fat CT images combined with clinical indicators. Our study indicated that the auto-segmentation framework could robustly localize the adrenal glands and accurately refine their boundary. The RN using adrenal glands and periadrenal fat onset CT images performed well, reflecting that the microscopic changes in adrenal glands and periadrenal fat in patients with COVID-19 can be detected using radiomics features. Our results suggested that adrenal gland and periadrenal fat changes on the onset of CT images in severe patients differed from those in moderate patients. All rights reserved. No reuse allowed without permission. perpetuity. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in The copyright holder for this this version posted January 5, 2021. that may cause an acute adrenal insufficiency that may be an indicator of making the disease worse (18) . Additionally, the adrenal gland is one of the most highly vascularized organs of the human body, with almost every adrenal cell in close location to endothelial cells. Therefore, endotheliitis in COVID-19 may further increase the vulnerability of adrenal tissue. However, this needs further study and validation. We hypothesized that the same effect could occur in patients with COVID-19 due to similar virus types. Our predicting results showed that AM's AUC value was slightly better than PM's, but there was no statistical significance. That may indicate that the inner changes in adrenal glands were not different from the changes from periadrenal All rights reserved. No reuse allowed without permission. perpetuity. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in The copyright holder for this this version posted January 5, 2021. ; https://doi.org/10.1101/2021.01.03.21249183 doi: medRxiv preprint fat. Further research is needed to confirm changes in periadrenal fat were caused by adrenal gland inflammatory cells' infiltration or direct damage of SARS-CoV-2 (19) . Previous research also reported that rather than direct damage by SARS-CoV-2 itself, adrenal gland changes were believed to be caused by immune overreaction or cytokine storm-inducing endocrinological pathway impaired through a tight binding mechanism of SARS-CoV-2 and ACE2 (20) . Endocrinological pathway impairment may be related to the following aspects. First, ACE2 is highly expressed in human adrenals, which serves as the entry receptor for SARS-CoV-2 that is bound to be significantly affected. Although no studies have proved it, some work has proposed that the imbalanced ACE/ACE2 axis may mediate tissue repair and wound healing pathways in the lungs, such as fibrosis after SARS-CoV-2 (21). Second, the main substrate for ACE2 is angiotensin II. ACE2 acts as a negative regulator of the renin-angiotensin-aldosterone system (RAAS) by converting the active angiotensin and angiotensin II to the inactive angiotensin 1-7 (22) . The adrenal gland is one of the most critical end organs of RAAS that is fundamental in regulating blood pressure, maintaining homeostasis, and is mediated in the organs' inflammatory response. Finally, the SARS virus contains several permutations of amino acid sequences with homology to the antigenic relevant residues of ACTH (23) that will have potential and significant pathophysiological effects due to molecular mimicry. ACTH is a key hormone that regulates autoantibodies' release by adrenal corticosteroids and could abrogate the adrenal stress response when autoantibodies were binding ACTH. Potential cross-reactivity of antibodies may cause local infiltration with All rights reserved. No reuse allowed without permission. perpetuity. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in The copyright holder for this this version posted January 5, 2021. ; https://doi.org/10.1101/2021.01.03.21249183 doi: medRxiv preprint immunocompetent cells in adrenal glands(14). In summary, these aspects suggested that adrenal cells and surrounding tissue damage in patients with COVID-19 may be caused by a viral infection and further secondary inflammatory plus autoimmune processes located in the adrenal glands. In conclusion, we proposed and constructed an adrenal gland auto-segmentation framework based on chest CT images and AI technology. We automatically obtained the ROI from the onset CT images through the auto-segmentation framework and inflation algorithm, extracted the radiomics features, and then developed and validated a RN for predicting the disease progression of COVID-19 combining onset CT images with clinical indicators. Until now, we did not find any research using adrenal gland CT parameters as indicators to evaluate the progression in patients with COVID-19, especially in the radiomics field. Our work has some limitations. First, we used the chest CT images as data resources. However, taking clinical practicality and radiation to patients into consideration, there is no need to perform another CT scan using professional adrenal glands CT parameters to observe adrenal lesions a little better because CT examination of patients with COVID-19 is mainly to detect and observe pulmonary lesions. Second, different from other studies, we selected the entire organ as ROIs rather than the lesion itself. The periadrenal fat area may not be precise and contain some other tissue that is indistinguishable to the human eye. Although there are some deficiencies, our findings validated the potential for radiomics features extraction from adrenal glands and periadrenal fat CT images to All rights reserved. No reuse allowed without permission. perpetuity. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in Table. Patients with adrenal lesions were excluded after being evaluated by two radiologists with more than 10 years of experience. We chose chest CT images scanned within four days to the first diagnosis as the onset image. If the CT scan was done more than once, we chose the one closer to the admission date. Figure 8 demonstrates the inclusion and exclusion criteria. Figure 9 shows the workflow of our study. The objective of image segmentation is to extract ROIs, including bilateral adrenal glands and the periadrenal fat. Here, we propose a cascaded V-net network framework to segment the bilateral adrenal glands automatically, and ROI of periadrenal fat was obtained using the inflation algorithm based on the adrenal gland. An experienced radiologist (Y.F.) manually delineated the ROIs of bilateral adrenal glands which were as the ground truth labels of adrenal, and he was asked to delineate the adrenals according to the image delineating principles of BraTS 2018. For the auto-segmentation framework, we first trained a coarse localization model as Model 1, which can perform coarse segmentation to locate the adrenal gland area. The second V-Net model was used for defining segmentation as Model 2, and further divided into the left adrenal gland and the right adrenal gland. The main structure of V-net includes Down-Block and Up-Block. The high-level context information was extracted by convolutions and Down-Blocks first. Then, skip connections were used to fuse high-level context information and fine-grained local information. Through residual function and skip connection, the model can achieve high accuracy. Normalization is performed before the image is input to the network. The segmentation frame is shown in Figure 10 . The cascade model segmentation of All rights reserved. No reuse allowed without permission. perpetuity. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in The copyright holder for this this version posted January 5, 2021. ; https://doi.org/10.1101/2021.01.03.21249183 doi: medRxiv preprint the bilateral adrenal glands worked sequentially to obtain more accurate segmentation results. ROIs including bilateral adrenal glands and periadrenal fat from all CT images without annotation were segmented using the auto-segmentation framework. Next, the ROIs with serious nonconformity were manually verified by two radiologists with more than 5 years of clinical experience and validated by another radiologist with 20 years of clinical experience. Radiomic features were extracted from ROIs on the CT images using a Python package (PyRadiomics V3.0). B-spline interpolation resampling was used to normalize the voxel size, and the anisotropic voxels were resampled to form isotropic Next, the feature in the training set was preprocessed by standardization. The mean and variance in the training set were applied to the validation and test sets. To reduce dimensionality and select significant features, we performed a feature dimension reduction process to select the most relevant features. First, a univariate analysis named K-best was employed (25). It selected features according to K highest All rights reserved. No reuse allowed without permission. perpetuity. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in The copyright holder for this this version posted January 5, 2021. ; https://doi.org/10.1101/2021.01.03.21249183 doi: medRxiv preprint scores computed through ANOVA F-value and P-value between label and features. Features with a significant difference (P < 0.05) were selected. Next, the LASSO feature-selection algorithm was used to extract the most informative radiomics features to prevent the "curse of dimensionality". After feature extraction and selection, logistic regression (LR) algorithms were trained to construct three radiomics models (AM, PM, and FM) for predicting the disease progression of COVID-19 using a five-fold cross-validation strategy. In the FM model, the RadScore of the patient was calculated according to the LASSO algorithm. We used univariate analysis to assess the relationship between clinical factors plus serum biomarkers and disease outcome. Features with P < 0.05 were introduced into multivariate LR analysis with radiomics features to select the best combination using a five-fold cross-validation strategy. The best assignment of training and validation sets was chosen for the next analysis. Next, we applied the multivariate LR model to build CM using valuable clinical indicators and RN using the RadScore from FM with clinical indicators to predict the disease progression of COVID-19. We conducted collinearity diagnosis by calculating the VIF for variables in RN to detect multicollinearity among the radiomics nomogram variables. In the end, RN was verified in the validation and test sets. Calibration curves and Hosmer-Lemeshow test were used to assess the relation between the predicted risks and actual results. DCA was used to evaluate the performance of the RN. All rights reserved. No reuse allowed without permission. perpetuity. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in Dice value was used to assess the effectiveness of auto-segmentation framework. The AUC of receiver-operating characteristics (ROC) with 95% confidence interval (95% CI), sensitivity, and specificity were used to evaluate the performance of AM, PM, FM, CM and RN. Accuracy was calculated to assess the prediction performance. Differences in AUC values among different models were estimated using the DeLong test. perpetuity. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in The copyright holder for this this version posted January 5, 2021. ; https://doi.org/10.1101/2021.01.03.21249183 doi: medRxiv preprint All authors have approved the submitted version (and any substantially modified version that involves the author's contribution to the study). All authors have agreed both to be personally accountable for the author's own contributions and to ensure that questions related to the accuracy or integrity of any part of the work. This multicenter study was approved by the ethics committees of Guizhou Provincial People's Hospital (2020, NO.01). Because of its retrospective nature, the need to obtain informed consent from patients was waived. The study was performed according to the principles of the declaration of Helsinki. perpetuity. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in The copyright holder for this this version posted January 5, 2021. perpetuity. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in The copyright holder for this this version posted January 5, 2021. ; https://doi.org/10.1101/2021.01.03.21249183 doi: medRxiv preprint perpetuity. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in The copyright holder for this this version posted January 5, 2021. All rights reserved. No reuse allowed without permission. perpetuity. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in The copyright holder for this this version posted January 5, 2021. ; https://doi.org/10.1101/2021.01.03.21249183 doi: medRxiv preprint perpetuity. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in The copyright holder for this this version posted January 5, 2021. ; https://doi.org/10.1101/2021.01.03.21249183 doi: medRxiv preprint Moderate patients n = 13 Severe patients n = 23 All rights reserved. No reuse allowed without permission. perpetuity. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in The copyright holder for this this version posted January 5, 2021. ; https://doi.org/10.1101/2021.01.03.21249183 doi: medRxiv preprint Figure 9 The workflow of study All rights reserved. No reuse allowed without permission. perpetuity. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in The copyright holder for this this version posted January 5, 2021. ; https://doi.org/10.1101/2021.01.03.21249183 doi: medRxiv preprint Figure 10 The framework of the used V-net All rights reserved. No reuse allowed without permission. perpetuity. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in The copyright holder for this this version posted January 5, 2021. perpetuity. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in The copyright holder for this this version posted January 5, 2021. ; https://doi.org/10.1101/2021.01.03.21249183 doi: medRxiv preprint The treatment and follow-up of'recurrence' with discharged COVID -19 patients: data from Guizhou Clinically Applicable AI System for Accurate Diagnosis, Quantitative Measurements and Prognosis of COVID-19 Pneumonia Using Computed Tomography Digital technology and COVID-19 Endocrine changes in SARS-CoV-2 patients and lessons from SARS-CoV Endocrine complications of COVID-19: what happens to the thyroid and adrenal glands The adrenal gland microenvironment in health, disease and during regeneration COVID-19 and the endocrine system: exploring the unexplored The Renin-Angiotensin system and SARS-CoV-2 infection: A role for the ACE2 receptor? The adrenal gland in stress -Adaptation on a cellular level Radiomics: Images Are More than Pictures, They Are Data Segmentation of CT 15 Pathology and Pathogenesis of Severe Acute Respiratory Syndrome Hyperinflammation and derangement of renin-angiotensin-aldosterone system in COVID-19: A novel hypothesis for clinically suspected hypercoagulopathy and microvascular immunothrombosis Extrapulmonary lesions in influenza Two hits to the renin-angiotensin system may play a key role in severe COVID-19 Computed Tomography Highlights Increased Visceral Adiposity Associated With Critical Illness in COVID-19 SARS-CoV-2 pandemic and research gaps: Understanding SARS-CoV-2 interaction with the ACE2 receptor and implications for therapy SARS-CoV-2 Pathogenesis: Imbalance in the Renin-Angiotensin System Favors Lung Fibrosis Organ -protective effect of angiotensin -converting This study was supported Science and Technology Foundation of Guizhou Province All rights reserved. No reuse allowed without permission.