Letter to the Editor: “Radiomics for COVID-19: A time for more precise approach”
by Hamid Abdollahi, Isaac Shiri (firstname.lastname@example.org)
COVID-19 is now the leading cause of death worldwide, and is an emergency problem that should be managed quickly. To diagnose and manage COVID-19, real-time reverse transcription polymerase chain reaction (rRT-PCR) and computed tomography (CT) are used as highly accurate and precisied approaches . Although other imaging methods such as chest radiography, ultrasound and positron emission tomography are also used, CT is found as best and fastest method .
In another way, researchers applied artificial intelligence (AI) on CT images and found that this issue could be used as a more accurate and feasible method for Covid-19 detection and management . In these studies, several deep and machine learning algorithms are used for infected region detection, segmentation, quantification and screening from other pneumonia . On the other hand, AI also is recruited for virus outbreak prediction, new drug interaction and development, genomic classification, survival prediction and health workers protection [5, 6].
Radiomics is now as an acceptable image processing issue for a wide range of clinical applications including tissue characterization, disease detection, prognosis, diagnosis and therapy response prediction . In this image quantification approach, radiomics features extracted from high quality medical images, may be used for clinical decision making after some pre-processing, feature selection and advanced data analysis . Radiomics is used more for oncological purposes, but also is examined for several diseases .
On the use of radiomics for COVID-19, there are just three studies. Qi et al., developed machine learning-based CT radiomics models for predicting hospital stay in Covid-19 patients . In this study, 52 patients were studied and CT radiomics models based on logistic regression (LR) and random forest (RF) were analyzed as predictive parameters for long and short term stays at hospital. The CT radiomics models based on 6 second-order features including 1) wavelet-High Low High-Gray Level Dependence Matrix-Small Dependence Low Gray Level Emphasis, 2) wavelet- Low High High-Gray Level-Co-occurrence Matrix-Correlation, 3) wavelet-Low High Low-Gray Level Size Zone Matrix-Gray Level Variance, 4) wavelet-Low Low High-Gray Level Size Zone Matrix-Size Zone Non Uniformity Normalized, 5) wavelet- Low Low High- Gray Level Size Zone Matrix-Small Area Emphasis and 6) wavelet-Low Low High Gray Level Co-occurrence Matrix-Correlation were found as effective measures in discriminating short- and long-term hospital stay in patients. The models have obtained areas under the curves (AUC) of 0.97 and 0.92 by LR and RF, respectively.
Guiot et al., developed a fully automatic framework by applying AI on radiomic CT signature for detecting COVID-19 . In this work, 1381 patients were included and developed model has found sensitivity and specificity of 78.94% and 91.09%, respectively, with an AUC of 0.9398 with negative predictive value of 97%. Fang et al. developed a radiomic signature to screen COVID-19 from CT images . In this study, 75 pneumonia patients including 46 patients with COVID-19 and 29 other types of pneumonias were investigated and 23 radiomic features were found to be highly associated with COVID-19. They also used four radiomics features to develop a diagnostic model by using a support vector machine. Their models obtained AUCs of 0.862 and 0.826 in the training set and the test set respectively.
Based on several previous radiomics studies, imaging features could be used as highly accurate biomarkers for detection, diagnosis, prognosis and therapeutic prediction [13, 14]. In the case of COVID-19, radiomics features extracted from CT, radiography, ultrasound and PET/CT could be examined as first lines for disease diagnosis and management. Although, there are no studies on the use of radiomics for several imaging modalities, it could be examined and tested as future based strategy for such diseases including COVID-19. On the other hand, combination of radiomics and AI or machine learning algorithms, may serve as feasible approach for COVID-19. In this light, it would be critical to find most reproducible and repeatable radiomics features and highly tuned machine learning algorithms. Because studies have indicated that radiomics features are vulnerable against changes in several parameters such as image acquisition, reconstruction, segmentation and type of selected machine learning algorithm . In recent three radiomics studies for COVID-19, the issue of reproducibility and repeatability was not studied, that is necessary for future studies. Also, several dataset with enough patient data are needed for validation and more precisied modelling.
As the philosophy of radiomics is to decode tissue heterogeneity, a wide range of radiomics features could be examined to assess COVID-19 characteristics. These features are also could be correlated to clinical and genetic data for personalization of the disease and finding main biological pathways. In addition, radiomics features may be used for new therapy development and assessment by calculation delta radiomics extracted from frequently imaging. The radiomics features also may be associated with qualitative features obtaining from radiologist’s reports for more accurate COVID-19 diagnosis and management.
Finally, it may be concluded that radiomics features could act as a time saving, non-invasive and easy to use approach for detection, diagnosis and therapy assessment in COVID-19 patients. It could be applied on all medical images and is combined with clinical and other demographic data for developing personalized models. Further clinical studies are needed to examine the feasibility of the radiomics apparoch.