Radiomics analysis for distinctive identification of COVID-19 pulmonary nodules from other benign and malignant … – Nature.com

A cross-sectional observational study was performed in our department after obtaining prior approval from the Institute Ethics Committee. We initially reviewed CT thorax scans of 250 patients with RT-PCR-positive COVID-19 infection over one year as detailed in Fig. 1a. Lung nodules were present in 24 patients with COVID-19 infection of which four patients were eliminated from our study as their CT images were not suitable for radiomics analysis. Only patients in whom no other synchronous lung pathology was identified were included. The imaging features of COVID-19 lung nodules were studied in the remaining 20 patients and were assessed for size, type, margins, location, and the lobe and segment involved, Digital imaging and communications in medicine (DICOM) images of these 20 patients were thereafter subjected to segmentation analysis and radiomics post-processing.

(a) Flow diagram of the study of COVID-19-infected cases. (b) Flow diagram of the study of non-COVID-19-infected cases.

We exclusively examined patients with COVID pneumonia who exhibited solid pulmonary nodules only, totaling 20 cases, to ensure similarity with benign and malignant nodules. Our study involved individuals who underwent chest CT scans between the 3rd and 6th day after the onset of symptoms (1st week of the disease) and who tested positive for COVID-19 infection through RT-PCR. Along with these pulmonary nodules, the other common CT findings in these patients typically included ground glass opacities, crazy paving, and consolidation. However, our focus for radiomics analysis was solely on these pulmonary nodules.

We reviewed CT thorax of 1200 non-COVID-19 patients as shown in Fig.1b. Lung nodules of compatible size were present in 133 cases. The final diagnosis was available in 97 patients, of whom 44 were benign and 53 were malignant nodules. The benign lesions were diagnosed based on histopathological diagnosis or correlation with clinical features and follow-up as per the Fleishers Society guidelines18. The final diagnosis in primary malignant lesions was arrived at based on histopathological diagnosis, and in metastasis based on the histopathological diagnosis of the lung nodule or primary tumor. The DICOM images of 40 benign nodules and 50 malignant nodules were subjected to segmentation analysis. Radiomics post-processing was done in 39 benign nodules and 49 malignant nodules. Radiomics analysis was not feasible in two patients. Subsequently, the radiomics texture analysis of COVID-19 lung nodules was compared separately with each radiomics analysis of benign non-COVID-19 benign lung nodules and malignant lung nodules.

The distribution of cases included in the final radiomics analysis included, n=24 (22%) metastatic pulmonary nodules, n=25 (23%) primary malignancies, n=39 (36%) non-COVID benign nodules, and n=20 (19%) COVID-related nodules (Fig.2). The final diagnosis was available in all these nodules. n=39 of the pulmonary nodules were found to be other benign, while n=49 were malignant. The benign lesions were diagnosed based on histopathological diagnosis or correlation with clinical features and follow-up as per the Fleishers Society guidelines. The final diagnosis in primary malignant lesions was arrived at based on histopathological diagnosis, and metastasis was based on the histopathological diagnosis of the lung nodule or primary tumor. Only patients in whom no other synchronous lung pathology was identified were included to prevent overlap of pathologies. Patients with subsolid pulmonary nodules and Nodules with calcification were excluded from the study to compare purely solid pulmonary nodules. Of the other benign lesions (n=39) analyzed using radiomics, 36% were septic emboli, 28%were benign lesions monitored long-term per Fleishner society guidelines, and the remainder comprised sarcoidosis, inflammatory conditions, pulmonary tuberculosis, benign carcinoid, hematoma, hydatid disease, Sjogrens syndrome, and Wegners granulomatosis cases, as shown in Supplementary Information (Fig. S1). Among the malignant nodules (n=49), there were 25 cases of primary lung malignancies and 24 cases of metastases. Primary lung malignancies consisted of adenocarcinoma (52%), squamous cell carcinoma (40%), and other types, as depicted in Supplementary Information (Fig. S2). The predominant source of metastases was breast carcinoma, with the remainder originating from various other primary organs, as depicted in Supplementary Information (Fig. S3).

Case distribution of the lung nodules included in the study.

HRCT examinations were performed using one of the following multidetector computed tomography (MDCT) scanners: Phillips-brilliance 16 (Philips medical systems, Cleveland); GE EVO evolution 128 slices (GE healthcare, Princeton); and Siemens biograph horizon (Siemens AG, Munich). HR-CT images were obtained during breath-holding with the following parameters: 120kV, 200mA. The section thickness and reconstruction intervals were 0.650.80mm. The CT images were sent to a picture archiving and communication system (PACS) to be interpreted at workstations.

The segmentation of the DICOM images of the pulmonary nodules, a critical initial step for accurate feature extraction, was performed manually by an expert radiologist using Insight Segmentation and Registration Toolkit (ITK-SNAP) software19 and was verified by three radiologists independently. The steps described above are shown in Figs.3 and 4. By relying on the expert radiologist, we could delineate the nodules with a high degree of precision, particularly in terms of their shape and texture characteristics, which are crucial for subsequent radiomic analysis. Following the segmentation, we extracted radiomic features from the 3D representations of the nodules. The extraction process focused on a comprehensive set of features, including but not limited to, shape, size, intensity, texture, and wavelet features. The emphasis was on capturing a broad spectrum of information that reflects the underlying pathology and can be correlated with clinical outcomes. By combining expert radiological input with radiomic feature extraction techniques, we aimed to mitigate some of the challenges associated with parametric texture feature extraction.

A 57year-old male patient with RTPCR has proven COVID-19 pneumonia. (a,b) The axial section of the CT thorax in the lung window and soft tissue window shows a subpleural soft tissue nodule in the posterior segment of the left lower lobe. (c) Creation of ROI for segmentation. (d) 2D-segmented nodule. (e) 3D-volumetric rendering of the nodule. (f) Follow-up chest CT after 6months revealed partial resolution of the nodule.

A 21year-old lady with cough and hemoptysis. HPE: benign carcinoid tumor. (a,b) Axial section of CT thorax in lung window and soft tissue window showing a mass lesion in the posterior segment of the right lower lobe. (c) Creation of ROI for segmentation. (d) 2D segmented mass lesion. (e) 3D volumetric rendering of the mass lesion.

Segmented lung nodules were used to extract different types of features. These features were classified into three categories: shape features (14), first-order features (18 features), and texture-based features (69 features). Texture-based features were of four types, namely gray level co-occurrence matrix (GLCM) features (24 features)20, gray-level run-length matrix (GLRLM) features (16 features)21, gray level size zone (GLSZM) features (16 features)15,22,23 and gray level dependence matrix (GLDM) features (13 features). Each radiomics feature was given a feature rank based on a random forest classifier. Out of 101, the top 10 features were selected for classification algorithms according to Anand et al.24. Figures 5 and 6 show the top 10 selected radiomics features with rank, and Tables 1 and 2 summarize their feature importance values. Several classification algorithms, such as SUPPORT VECTOR MACHine (SVM)25, multi-layer perceptron (MLP), naive Bayes, discriminant analysis, and decision tree26, were applied to selected feature matrices to classify benign and malignant nodules. SVM with the linear kernel (L-SVM) and radial basis function kernel (RBFSVM) were used as SVM variants. Linear discriminant analysis (LDA) and quadratic discriminant analysis (QDA) were used in the category of discriminant analysis. We have also experimented with MLP classifiers for different hyperparameters which include activation, layers/number of neurons, and learning rate. To evaluate the performance of classifiers, confusion matrices were drawn on the test set. Accuracy, sensitivity, specificity, precision, and F1-measure were calculated for each classifier.

(a) Ten important features used for the classification of COVID-19 and non-COVID-19 benign lung nodules. (b) Comparison plot of the most prominent feature.

While many state-of-the-art approaches in medical image analysis today do use deep learning methods, in our experiments they showed poor performance with an accuracy of at most 55%. We evaluated models such as ResNet, DenseNet, and Vision Transformer for the same but due to the limited data available, the models showed poor performance27. The radiomic features provide a more robust basis for training on limited data as compared to the deep learning approaches.

The study was performed after obtaining prior approval from the Institutional Research Ethics CommitteeSri Ramachandra Institute of Higher Education and Research (CSPMED/19/SEP/56/122) and all methods were performed by relevant guidelines and regulations.

Informed consent was obtained from all subjects and/or their legal guardians involved in the study.

Excerpt from:

Radiomics analysis for distinctive identification of COVID-19 pulmonary nodules from other benign and malignant ... - Nature.com