Although COVID-19 tests are one means to determine infection, the rapid spread of the disease has created a shortage of test kits and a tight supply of reagents that are used to create test kits. Symptoms are another indication, but some of the infected show very mild or no symptoms. Another method leading to a swift diagnosis is to take an X-ray or CT scan of the patient’s lungs. CT scans provide a higher resolution image than X-rays. CT scans have also shown a higher sensitivity than real-time RT-PCR (reverse transcription-polymerase chain reaction). RT-PCR is known as the most sensitive technique for detecting mRNA, as a test in diagnosing COVID-19.
Part of the crushing burden of COVID-19 on medical personnel is also visited upon radiologists who determine whether X-ray or CT scans of a patient’s lungs indicate pneumonia, COVID-19, or something else. Other diagnostic tools include pulse oximeters, which indicate the amount of oxygen in the bloodstream. However, radiologists are vital in interpreting the severity of COVID-19 based on lung images. COVID-19 shows a “ground glass” appearance of opacity in the lungs, among other things.
Technology lends a…brain
The Novel Coronavirus Pneumonia Emergency Response Epidemiology Team (NCPERT), operating in China, built and deployed Artificial Intelligence (AI) as a preliminary analysis of COVID-19 in lung scans. In a pre-print, unreviewed paper submitted to medRxIV the authors state, “Working in an interdisciplinary team of over 30 people with medical and/or AI background, geographically distributed in Beijing and Wuhan, we were able to overcome a series of challenges in this particular situation and deploy the system in four weeks. Using 1,136 training cases (723 positives for COVID-19) from five hospitals, we were able to achieve a sensitivity of 0.974 and specificity of 0.922 on the test dataset, which included a variety of pulmonary diseases.” The AI system also highlighted all areas with legions, which allowed for a faster examination. The deep learning (DL) system was deployed in 16 hospitals and has performed over 1,300 screenings per day.
NCPERT took advantage of end-to-end deep neural network models, adapting the models they had previously developed for other diagnoses, and constructed a training-interference pipeline. After quickly evaluating several promising models, they compiled training data from positive cases, including a negative set from existing images of other lung diseases so the model could effectively learn to distinguish a COVID-19 lung image from others.
Using sample images from 5 hospitals and 11 different models of CT scanners to generalize the model, NCPERT “developed a three-stage annotation and quality control pipeline, allowing inexperienced data annotators to work with senior radiologists to create accurate annotations, with minimal time investment from the radiologists.” NCPERT continuously retrained the model to improve accuracy as more data came in. The tool was delivered in an easy-to-use, low-cost platform for hospital IT staff to set up, with a remote upgrade function. Using the tool, physicians were able to rapidly locate slides and regions for a detailed examination that assisted in improving their diagnosis, regardless of their personal experience with COVID-19.
The model missed some positives for ground-glass opacity of less than 1 cm in diameter and did not perform well with multiple types of lesions or images where subjects introduced motion or metal objects. However, the AI tool was able to help overwhelmed hospitals by providing preliminary results that sped up the triage process of filtering suspected COVID-19 patients. Also, less experienced radiologists were able to learn to better detect COVID-19 indicative features. After receiving the first set of data, NCPERT produced the first useable model in just one week.
Much more about the NCPERT team of Chinese engineers, scientists, and physicians that built this AI model for screening CT lung scans for the presence of COVID-19 can be found at medRxIV in the pre-review pre-print paper titled “AI-assisted CT imaging analysis for COVID-19 screening: Building and deploying a medical AI system in four weeks.”