How reliable are radiographic, CT and ultra-sound appearances of coronavirus in the lungs in terms of their sensitivity/specificity? A bunch of imaging labs popped up in my town recently, all of them claiming the ability to uniquely identify or rule out the presence of covid-19 using their lung scans. I'm pretty sceptical of these claims, but I haven't been able to find any layman-friendly summaries of available research on this matter.
It has been recognized since the earliest months of the pandemic that SARS-Cov-2 infection is associated with otherwise uncommon lung CT findings including ground-glass opacities (GGO), GGO plus a reticular pattern, vacuolar sign, microvascular dilation sign, fibrotic streaks, subpleural lines, and subpleural transparent lines¹.
In a recently published paper using Deep Learning (a type of machine learning suited for image analysis), the authors reported achieving 99% positive predictive value and negative predictive value discriminating lungs infected with COVID19 from normal lungs or those with bacterial pneumonia. (Table 2 also includes sensitivity and specificity).
In my experience, machine learning rarely works as well in practice as in the manuscript, but this still demonstrates the potential feasibility of using CT images to appropriately diagnose SARS-Cov-2. Of course, the images in this study are likely biased towards critically ill patients. Thus, the discriminative ability for patients that can walk off the street into an "imaging lab" may be decreased. Likewise, the ability of a human radiologist to perform as well as the machine learning algorithm is a question which may be difficult to answer.