Artificial Intelligence as a tipping point for the pandemic

“Artificial Intelligence is likely to be either the best thing or worst thing happen to humanity.”
– Stephen Hawking.

Today Artificial Intelligence (AI) has become a cliche in industry and business. This technology is a crucial lynchpin of much of the digital transformation taking place in our society. It helps the organisations to position themselves in capitalising the ever-growing amount of data being generated and collected. The ideal characteristic is its ability to rationalise thereby taking actions that have the best chance of achieving a special goal. Let us find out whether AI advancement is a blessing for humanity in the COVID-19 crisis.

The confirmation of SARS-CoV-2 is performed with virus-specific Reverse Transcriptase Polymerase Chain Reaction(RT-PCR) test. According to WHO, Nucleic acid detection in secretion fluid collected from a throat swab using RT-PCR is the most accurate method for detection. As it possesses high sensitivity, the false-negative rate is also high. The laboratory result will take two days for completion depending on the procedure followed. There occurs a shortage of RT-PCR test kits in outbreak regions to conduct the test. To overcome the limitations caused by the RT-PCR test, Chest Computed Tomography(CT) is a valuable component in the evaluation of suspected patients as it is quicker in action than RT-PCR. The typical findings obtained from chest CT imaging of confirmed patients are ground-glass opacities and consolidation. Also demonstrated bilateral, peripheral, lower lung distribution with a rounded morphology. In some countries like China and South Korea, Chest CT has been widely used in clinical practice due to its speed and availability during the outbreak of Covid-19. Mainly there arise two potential limitations of chest CT :

The pressure on health systems may be overburdened which may limit timely interpretation of CT by radiologists.
The morphology and accuracy of pathologic findings on CT are variable.

In the case of nosocomial infection(caused by organisms that are resistant to antibiotics), rapid detection of patients with Covid-19 is imperative and a false negative result could delay treatment thereby increasing the risk of transmittance. Also, radiologists with expertise in thoracic imaging may not be available at every institution increasing the need for AI aided detection.

AI’s Role in fighting against Covid-19

Medical AI uses computer techniques to perform clinical diagnosis and suggest treatments. The state-of-the-art AI methodology has shown great capabilities and capacities in recognition of meaningful data patterns and thus being widely experimented as a tool for clinical trials. In the study, researchers used AI algorithms for integrating chest CT findings with exposure history, clinical symptoms and laboratory testing to diagnose positive patients rapidly in the early stages. At first, they had developed a deep Convolutional Neural Network(CNN) for learning the imaging characteristics of positive patients on the initial scan. Then Support Vector Machine(SVM), Random Forest and Multilayer perceptron (MLP) are used to classify patients according to clinical information. Lastly, the neural network model which combines radiological data and clinical information to predict Covid-19 status was created.

Elucidation of modelling framework:

Three AI models are used to generate the probability of a patient being COVID-19 positive.

based on chest CT scan.
on clinical information.
on a combination of chest CT scan and clinical information.

Evaluation: Each slice was ranked by probability of containing a parenchymal abnormality as predicted by CNN model which is a pre-trained pulmonary TB model that has 99.4% accuracy to select abnormal lung slices from chest CT scans. Top ten abnormal CT images per patient were put into the second CNN to predict the likelihood of COVID-19 positivity P1. Demographic and clinical data like age and sex of patient, exposure history, symptoms and lab test were put into the machine learning model to classify COVID-19 positivity P2. Features generated by diagnosis from the CNN model and non-imaging clinical information machine-learning model were integrated by an MLP network to generate the final output of the joined model P₃.

Patient’s age, presence of exposure to SARS-CoV-2, presence of fever, cough and cough with sputum and white blood cell counts were significant features associated with SARS-CoV-2 status. The logistic regression was a good fit (P = 0.662) where P-value indicates the significance of the difference in performance metric compared with respect to the joint model. 905 patients were tested by real-time RT–PCR assay, next-generation sequencing RT–PCR and 419 (46.3%) tested positive for SARS-CoV-2. In a test set of 279 patients, the AI system achieved an area under the curve of 0.92 and had equal sensitivity as compared to a senior thoracic radiologist. The AI system also improved the detection of patients who were positive for COVID-19 via RT–PCR who presented with normal CT scans, correctly identifying 17 of 25 (68%) patients whereas radiologists classified all these patients as COVID-19 negative. When CT scans and associated clinical history are available, the proposed AI system can help to diagnose the patients rapidly. These are the observations of researchers from the above table.

Methods used are:

a)Study of patients: An initial chest CT studies and clinical from 905 patients presenting between 17 January and 3 March 2020 to 1 of 18 centres in 13 provinces in China, where patients had SARS-CoV-2 exposure, fever and RT-PCR done. The exposure is defined as:

travel history of non-resident Wuhan patients or travel to the animal market within 30 days of symptoms onset for Wuhan resident patients.
close contact with RT-PCR test positive patients.

b)Clinical information: Patient’s age, sex, exposure history, symptoms (presence or absence), white blood cell counts, absolute neutrophile number, percentage of neutrophils, absolute lymphocyte number and percentage lymphocytes were collected. Label Encoder function in the scikit-learn package was used to encode the categorical variables into numerical within a range of 0 and 1 using MinMax Scaler function for further model development.

c)Reader Studies: The prediction of AI models were compared with the reviews of two radiologists and were provided with an initial CT scan, clinical history. They evaluated initial CT scan, clinical details, combined imaging and clinical data in their separate reviews. Prediction of the status of each patient was done which was then compared to those of AI algorithm and RT-PCR results.

d) AI Models: Three different models using CT images and clinical information were evaluated. Those are Deep learning model using CNN, Conventional Machine Learning Methods(including SVM, Random forest and MLP), a Joint CNN model.

e)Convolutional Neural Network (CNN): It was trained to predict whether a CT slice from a patient was positive or negative. In the study of each patient, the average probability from the ten abnormal CT slices was used to predict the COVID-19 status of the patient.

f) Image Preprocessing: Relevant slices from hundreds of images produced by a CT scan contain pulmonary tissue and potential parenchymal abnormality. Image segmentation was used to detect parenchymal tissue. A standard lung window with width (w) = 1,500 HU and level (l) = −600 HU was used to normalize each slice to pixel intensities between 0 and 255. The body part of the CT image was segmented by finding the largest connected component, consisting of pixels with an intensity greater than 175 and was filled into a solid region. The lung region was defined as the pixels with intensity less than 175 that fall within the segmented body part. Images were discarded if the size of the lung was less than 20% of the size of the body part.

g) Slice selection CNN: To identify abnormal CT images a pretrained Inception-Residual network(ResNet-v2) model based on the Image net as the slice selection CNN was used. The TB model predicts the probabilities of classes including PTB, non-TB pneumonia and normal chest CT. Researchers applied the PTB model to a full CT scan for selection of 10 slices with a lower probability of being normal and noted that these slices don’t show any abnormal findings of a normal patient.

h)Disease diagnosis CNN: 18 layer Residual network takes images of segmented lungs as input and outputs the probability of COVID-19 positivity. The abnormal findings are localised in a subregion of CT image. By combining the predictions of the local region over the whole image, the label of an image can be predicted. An image can be labelled with Max pooling which serves as an ‘OR’ gate.

i)CNN Training: The binary cross-entropy is used as an objective function and to train the neural network Adam Optimizer with a learning rate of 0.001 was used. 20% of training samples were held out as the tuning set to monitor the progress of the training process which was iterated for 40 epochs and a batch size of 16 samples. The model with the lowest binary cross-entropy on the tuning set was selected as the final model.

j)Machine-learning classifiers: SVM, Random forest and MLP classifiers were developed on the basis of the clinical information provided. Researchers had assessed ‘C’ and kernel for SVM, number of estimators was tuned for random forest and for MLP, the number of layers, number of hidden nodes in each layer. A three-layer MLP model with 64 nodes in each layer was selected after hyper parameter optimization as it possesses the highest AUC score on tuning set.

k)Joint model: combining CT image and clinical information: Twelve clinical features of the same patient were made into series with a 512-dimensional feature vector. MLP accepts this as its input for the prediction of COVID-19 status. A three-layer MLP with 64 nodes in each layer and is composed of a batch normalization layer, a fully connected layer and a ReLU activation function was used for the study. The MLP was jointly trained with the CNN and binary cross-entropy was applied for validating predictions. The sum of these two measurements was used as the overall objective function to train the joint model.

l) Statistical Analysis: For the calculation of the two-sided P value for sensitivity and specificity between models and human readers, McNemar’s test was used. The optimal model sensitivity and specificity was determined by the Youden index. Statistical significance was defined as a P value less than 0.05. Logistic regression was used for the evaluation of each clinical variable. Hosmer-Lemeshow goodness of fit was used to evaluate the goodness of fit of the logistic regression.

With the ever-increasing population, the world has put enormous pressure on the health care sector for quality treatment. Nowadays people are demanding smart health care services, applications and gadgets that will help them in leading a better life. From the words of Tim Cook “What all of us have to do is to make sure we are using AI in a way that is for the benefit of humanity, not to the detriment of humanity.” Let’s hope that the study of researchers about the role of AI in the health sector will provide us with relaxation from these hardships.

Watch the video of Yang Yang, Assistant Professor Of Radiology to acquire the basic concepts of their study.