In our study, we quantitatively analyzed multi-view whole lung radiomics to distinguish COVID-19 from other types of pneumonia by using machine learning methods. The radiomics model based on PSO-DELM could effectively differentiate COVID-19 from other types of pneumonia before diagnosis. Phase congruency features were used as radiomics features at the first time, it may be considered as radiomics biomarkers to predict COVID-19.
The previous studies showed that COVID-19 were significantly associated with the following CT imaging characteristics, such as bilateral lungs, peripheral or diffuse distribution, ground glass opacity, maximum lesion range, number of lesions, lobe involvement, Hilar and mediastinal lymph nodes enlargement, and so on[21–22]. All of these CT characteristics were almost expressed in the lung pulmonary window (except mediastinal lymph nodes enlargement), so the whole lung of different planes could be displayed the richer radiomics information of COVID-19 to radiologists. The current most studies were focused on the radiomics features of the lesions, these lesions were segmented by manual segmentation or by deep learning approaches[17, 21]. Manual segmentation was expensive to study and diagnosis in scarcity of clinical resources. Meanwhile, deep learning methods were need a amount of labels and the precision of the segmentation need to be improved(DSI:0.778). In our study, we tried to segment the lung unsupervised without manual operation or imaging software, the performance of segmentation was as good as classical region growing method(DSI: 0.9230 vs 0.9092) that was detailed in Appendix A1. FANG et al’s study segmented the lesion area by using a imaging software, AUC value of radiomics model for differentiating COVID-19 based on SVM was 0.826. It showed that the radiomics model based on PSO-DELM(AUC 0.9444) was better than radiomics models based on SVM(AUC 0.5556) in our study. For the transverse plane and sagittal plane, the sensitivity of radiomics model based on PSO-DELM was higher than 0.8889, which meant that the whole-lung radiomics features can accurately assess the COVID-19.
Non-invasive disease diagnosis could be realized by using imaging radiomics, 2D image features previously studied were only extracted from the transverse plane . 2D image radiomics features extracted from the transverse plane or 3D radiomics features extracted from voxel of lesions could not fully describe the whole COVID-19 radiomics features comprehensively, but sagittal plane, coronal plane and transverse plane could roundly contain radiomics information of the lesions (including the shape, size, location). In this study, we found that Hessian matrix features were all captured on the three planes. It indicated that these features can display gray gradient change information from different views. Table 4 showed that the prediction accuracy of these radiomics features achieved 0.9412 both on three different planes. Entropy feature and HOG features which described the shape and texture were captured on the transverse plane and sagittal plane, but the performances of Entropy features were better than HOG features on the transverse plane(Accuracy 0.9412 vs 0.8824).
GLCM features and phase congruency features were captured both on the transverse plane and coronal plane, but our study showed that the prediction performance of phase congruency features were better than GLCM features on different planes, that was shown in Table 4 and Fig. 3. To our knowledge, there were fewer reported researches about phase congruency features used in radiomics. The phase congruency was invariant to image contrast and could extract effective and reliable texture features under different illumination conditions . Phase congruency properly extracted features of any kind of phase angle, it was different from feature detectors based on gradient, which could only extract step features. It was usually used in palm print authentication, face representation technique, image segmentation technique, and so on [24–28]. In our study, phase congruency feature was favorable performance in differentiating COVID-19 from other types of pneumonia on the three planes (all AUC = 0.9444). It may be considered to be as diagnostic biomarker for lung in other diseases.
In this study, we constructed three radiomics models to identifying COVID-19. In the independent test cohort, the prediction accuracy, sensitivity, specificity and AUC of radiomics models based on PSO-DELM showed 0.9412, 0.8889, 1.0000 and 0.9444 on the transverse plane, respectively. It also did well on other two planes. The radiomics model based on BP(AUC = 0.9444) showed as good as the radiomics model based on PSO-DELM on the transverse plane, but the radiomics model based on SVM performed poorly on all three planes. So the radiomics model based on PSO-DELM could be as a clinical adjunct tool to help radiologists to predict COVID-19 from different views.
There were several limitations of this study. First, it was a retrospective study using a small sample size from three hospitals with no external test data, so small amount of COVID-19 may affect the problems of over-fitting and robustness of the prediction effect. Second, we did not segment the lesions in lung parenchyma that may imply that extracted radiomics features did not express the all radiomics information, because so far there were fewer accurate automatic segmentation technologies. That might affect all real radiomic information. Third, phase congruency feature was used as radiomics features at the first time, but no more proof of availability in other areas of diseases. Besides, 2D radiomics features were only used in our study. 3D radiomics features would improve the predictive ability of our models.