AIDCOV: An Interpretable Artificial Intelligence Model for Detection of COVID-19 from Chest Radiography Images

Abstract

As the Coronavirus Disease 2019 (COVID-19) pandemic continues to grow globally, testing to detect COVID-19 and isolating individuals who test positive remains to be the primary strategy for preventing community spread of the disease. The current gold standard method of testing for COVID-19 is the reverse transcription polymerase chain reaction (RT-PCR) test. The RT-PCR test, however, has an imperfect sensitivity (around 70%), is time-consuming and labor-intensive, and is in short supply, particularly in resource-limited countries. Therefore, automatic and accurate detection of COVID-19 using medical imaging modalities such as chest X-ray and Computed Tomography, which are more widely available and accessible, can be beneficial as an alternative diagnostic tool. We develop a novel hierarchical attention neural network model to classify chest radiography images as belonging to a person with either COVID-19, other infections, or no pneumonia (i.e., normal). We refer to this model as Artificial Intelligence for Detection of COVID-19 (AIDCOV). The hierarchical structure in AIDCOV captures the dependency of features and improves model performance while the attention mechanism makes the model interpretable and transparent. Using a publicly available dataset of 5801 chest images, we demonstrate that our model achieves a mean cross-validation accuracy of 97.8%. AIDCOV has a sensitivity of 99.3%, a specificity of 99.98%, and a positive predictive value of 99.6% in detecting COVID-19 from chest radiography images. AIDCOV can be used in conjunction with or instead of RT-PCR testing (where RT-PCR testing is unavailable) to detect and isolate individuals with COVID-19 and prevent onward transmission to the general population and healthcare workers.

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

“…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 . ResNet and DenseNet architectures showed better performance than the others, with accuracies ranging from 0.88 to 0.99.…”

“…Diagnostic models using CT scans and deep learning. Eighteen papers applied deep learning techniques to CT imaging, all of which were framed as a classification task to distinguish COVID-19 from other lung pathologies such as (viral or bacterial) pneumonia, interstitial lung disease 35,[40][41][42][43][44][45][46][47] and/or a non-COVID-19 class 40,41,44,46,[48][49][50][51][52] . The full three-dimensional (3D) volumes were only considered in seven papers 40,43,47,50,[52][53][54] with the remainder considering isolated 2D slices or even 2D patches 45 .…”

“…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 .…”

“…Table 2, we give the performance metrics quoted in each paper. Ten papers used cross-validation to evaluate model performance 21,35,36,47,49,57,65,72,74,75,77 , one used both cross-validation and an external test set 41 , one quoted correlation metrics 76 and one had an unclear validation method 17 . The other papers all had an internal holdout or external test set with sensitivity and specificity derived from the test data using an unquoted operating point (with the exception of ref.…”

Applications of artificial intelligence in battling against covid-19: A literature review

“…In order to compare different approaches in classification of X-Ray images, Sixteen versions of neural networks are compared in [337] . A hierarchical attention neural network model is proposed in [338] which captures the dependency of features and improves the model performance. The adopted mechanism is proposed to make the model interpretable and transparent.…”