An Integrated Framework with Machine Learning and Radiomics for Accurate and Rapid Early Diagnosis of COVID-19 from Chest X-ray

Abstract

Early diagnosis of COVID-19 is considered the first key action to prevent spread of the virus. Currently, reverse transcription-polymerase chain reaction (RT-PCR) is considered as a gold standard point-of-care diagnostic tool. However, several limitations of RT-PCR have been identified, e.g., low sensitivity, cost, long delay in getting results and the need of a professional technician to collect samples. On the other hand, chest X-ray (CXR) is routinely used as a cost-effective diagnostic test for diagnosis and monitoring different respiratory abnormalities and is currently being used as a discriminating tool for COVID-19. However, visual assessment of CXR is not able to distinguish COVID-19 from other lung conditions. Several machine learning algorithms have been proposed to detect COVID-19 directly from CXR images with reasonably good accuracy on a data set that was randomly split into two subsets for training and test. Since these methods require a huge number of images for training, data augmentation with geometric transformation was applied to increase the number of images. It is highly likely that the images of the same patients are present in both the training and test sets resulting in higher accuracies in detection of COVID-19. It is, therefore, vital to assess the performance of COVID-19 detection algorithm on an independent data set with different degrees of the disease before being employed for clinical settings. On the other hand, machine learning techniques that depend on handcrafted features extraction and selection approaches can be trained with smaller data set. The features can also be analyzed separately for various lung conditions. Radiomics features are such kind of handcrafted features that represent heterogeneous appearance of the lung on CXR quantitatively and can be used to distinguish COVID-19 from other lung conditions. Based on this hypothesis, a machine learning based technique is proposed here that is trained on a set of suitable radiomics features (71 features) to detect COVID-19. It is found that Support Vector Machine (SVM) and Ensemble Bagging Model Trees (EBM) trained on these 71 radiomics features can distinguish between COVID-19 and other diseases with an overall sensitivity of 99.6% and 87.8% and specificity of 85% and 97% respectively. Though the performance is comparable for both methods, EBM is more robust across severity levels. Severity, in this case, was scored between 0 to 4 by two experienced radiologists for each lung segment of each CXR image represents the degree of severity of the disease. For the case of 0 severity, sensitivity and specificity of the EBM method are 91.7% and 100% respectively indicating that there are certain radiomics pattern that are not visibly distinguishable. Since the proposed method does not require any manual intervention (e.g., sample collection etc.), it can be integrated with any standard X-ray reporting system to be used as an efficient, cost-effective and rapid early diagnosis device. It can also be deployed in places where quick results of the COVID-19 test are required, e.g., airports, seaports, hospitals, health clinics, etc.

Common pitfalls and recommendations for using machine learning to detect and prognosticate for COVID-19 using chest radiographs and CT scans

“…Figure 3 shows the number of images of each class used in the holdout/test cohorts. We found that 6/32 papers had an imbalanced testing dataset 17,24,33,36,37,61 . Only 6/32 papers tested on more than 1,000 images 17,27,36,41,54,81 .…”

“…Twenty-two papers considered diagnosis of COVID-19 from CXR images . Most of these papers used off-the-shelf networks, including ResNet-18 or ResNet-50 16,17,20,26,29,32,37 , DenseNet-121 27,28,31,32,34 , VGG-16 or VGG-19 19,33,35 , Inception 21,38 and EfficientNet 30,39 , with three considering custom architectures 18,25,36 and three using hand-engineered features [22][23][24] . Most papers classified images into the three classes, that is, COVID-19, non-COVID-19 pneumonia and normal 16,19,21,23,25,26,28,30,[32][33][34][35][36][37] , while two considered an extra class by dividing non-COVID-19 pneumonia into viral and bacterial pneumonia 17,29 .…”

“…Almost all papers had a high (45/62) or unclear (11/62) risk of bias for their participants, with only six assessed as having a low risk of bias. This was primarily due to the following issues: (1) for public datasets it is not possible to know whether patients are truly COVID-19 positive, or if they have underlying selection biases, as anybody can contribute images 16,24,26,28-32,34,35,37,41,44,48,49,76 ; (2) the paper uses only a subset of original datasets, applying some exclusion criteria, without enough details to be reproducible 16,43,44,48,49,51,61,70,71,75,76 ; and/or (3) there are large differences in demographics between the COVID-19 cohort and the control groups, with, for example, paediatric patients as controls 17,24,28,29,31,32,35,37,45,46,59,81 .…”

“…We see that 20/32 papers have a reasonable balance between classes (with exceptions being refs. 17,24,26,30,31,33,36,37,40,51,61,62 . However, the majority of datasets were quite small, with 19/32 papers using fewer than 2,000 datapoints for development (with exceptions being refs.…”

Prototype Development of an Expert System of Computerized Clinical Guidelines for COVID-19 Diagnosis and Management in Saudi Arabia

“…The radiomic features obtained from images taken by CXR can be used to detect and classify the COVID-19 stages. The heatmap of z-scores and the one-way ANOVA test [29] are used to find significant features that can be used to classify COVID-19 stages. A number of machine learning classifiers can be used for the best performance.…”