Did Artificial Intelligence just save my life?

It happened in February 2021, a sad anniversary of COVID-19 pandemic. After a long locked-down winter, virus spreading began to slow down in the Northern hemisphere. People started to socialize in person again. Vaccination was just underway but still only for those at high risk and elderly.

I have been teleworking the whole year. One evening during my weekly cybersecurity course I have noticed unusually dry throat when speaking. Next morning, I woke up with soar muscles and mild headache. During the day I worked in front of my two laptops as usual, but later in the afternoon I started shivering. I had high fever, however Paracetamol was effective in reducing it. My wife had the same symptoms. A friend we saw two days earlier had also the same symptoms and was tested positive to COVID-19. For four days the only symptoms were tiredness with muscle pain, dry throat and afternoon high fevers. At medical examination blood sample showed infection, while lungs X-ray did not show anything doctors would particularly worry about. We got vitamins and minerals supplements, antibiotic, and more paracetamol. On the fifth day we both felt better and did not have high fever. I thought that the worst has passed.

The next day after the morning shower I put on a deodorant and did not recognize its usual smell. I put on so much deodorant that my wife could not support it. A morning coffee taste was terrible without smell. We both had fever again but lower than during the first four days. However, that fever would last the whole day and night. And we started coughing. A lab test showed increased C-reactive protein (CRP) in blood. For my wife, the values were much higher than mine. She was so tired that she could not walk even to the bathroom. The doctor suggested a chest computerized tomography (CT) scan. Scan report showed severe form of viral pneumonia for my wife and a milder form for me. She was hospitalized for 20 days. I was threated at home with intravenous cocktail of strong drugs during one week before pneumonia diminished. For another week I could barely walk and spent most time in bed. I was exhausted. I could not read and focus. After more than three weeks since the first COVID-19 symptoms, I still did not have a sense of smell and was exhausted but started to read through the long list of emails. I have opened a message from the diagnostic center that performed a CT scan and found two attachments. The first one was doctor’s report that I had already received in the diagnostic center. The second one caught my attention. It stated, “RadLogic: AI Plugin processing report”. The report concluded with following results:

AI plugin Corona result:

The degree of damage to the parenchyma of the right lung is up to 2%

The degree of damage to the parenchyma of the left lung is up to 1%

CT scans are signs of viral pneumonia. Severity: CT1. Suspected for COVID-19 with confidence: 99.5%

AI plugin Chest result: “No pleural free air; No Pleural fluid; Normal Thoracic aorta caliber”AI plugin Nodules result: “4 nodules seen in 1 series; 2 Pulmonary Patches seen in 1 series; Multiple nodules discovered”.

The report contained detailed analysis of 203 lung slices images. It highlighted a key image:

RADLogics AI plugin COVID-19 localization image

From doctors in the medical center and the hospital I have learned that early diagnosis of COVID-19 and accurate assessment of infection severity are critical to timely administration of appropriate treatment to contain infection and to reduce risk of more severe consequences. With high number of patients during peeks of infection waves, this time critical diagnosis is a challenge for healthcare systems.  However, diagnostic capacity could be augmented with introduction of artificial intelligence (AI) assistance to radiologists. A recent large-scale multi-medical center study finds that AI assisted COVID-19 diagnosis reduces reporting turn-around time by 30 percent. The study included 128,350 chest CT scans – 36,358 of which were assessed with the RADLogics algorithm – as well as 570 radiologists from more than 130 hospitals and outpatient facilities.

How does this solution work? The RADLogics solution is based on the residual neural network (ResNet) with 48 convolution layers, one max-pool and one average-pool layer. This AI construct is known as ResNet50 and is used on computer vision tasks such as image classification, object localization, and object detection. ResNet50 architecture is adapted to include a sigmoid activation function for pneumonia detection and localization. The proposed network is trained using the RSNA Pneumonia Detection Challenge Dataset. Two-stage training model applies a gradient-weighted Class Activation Mapping (Grad-CAM) to generate more accurate pneumonia localization. For disease severity scoring the input image enters the detection and localization network. If the detection prediction is lower than a pre-determined threshold, the image is classified as negative; otherwise, a threshold is applied over the final localization output map to generate the pneumonia segmentation. At this point, the pneumonia segmentation and the lung segmentation blocks are utilized to compute the “Pneumonia Ratio”. This workflow is a modified U-Net, a convolutional neural network originally developed for biomedical image segmentation at the Computer Science Department of the University of Freiburg. Modifications result in a more robust model that generalizes to images with severe infections and reduces false detections. Several experiments were conducted on three data sets, with and without COVID-19 patients, to evaluate the system’s detection performance. The localization threshold that yields the maximum mean average precision is selected to produce the segmentations for the final model. Summary of results show high performance on non-COVID-19-pneumonia detection, and an even higher performance on COVID-19-pneumonia detection. Adapting the existing AI architectures and related training models to optimize them for detection and localization of new diseases for which previous datasets do not exist, has proven successful in published studies and practical implementations. Leveraging artificial intelligence and cloud computing platforms allows rapid development of solutions at scale to respond to global crisis such as COVID-19 pandemic without changes to existing radiologists’ workflow. Globally interconnected platform provides continuous increase of accuracy with data from thousands of new cases.

This case study shows that artificial intelligence and deep learning could be life saving if embedded into existing work practices to augment capacities and maximize resulting values.