A leading researcher in the field of medical image analysis has cautioned against the rush to provide AI solutions to the COVID pandemic, arguing that the need to help out must not compromise scientific principles.
Prof Hamid Tizhoosh, head of KIMIA Lab, Faculty of Engineering at the University of Waterloo wrote a piece on Medical-News.Net where he laid out the problems involved in such “quick fix” solutions.
He explains that AI researchers often make “toy” datasets which they use to experiment with in their own labs. In the middle of the pandemic, it is difficult to collaborate with radiographers who have their hands full dealing with COVID patients’ images.
AI research requires the acquisition and curation of large amounts of high-quality data, and currently there is an absence of this. While there are still few publicly available X-ray images or CT images of COVID patients’ lungs, they are beginning to crop up on the internet. AI researchers and enthusiasts are scraping together these images for their data sets and supplementing them with those of pneumonia patients, which are much more readily available. The results of their AI work are being released in papers that are not peer reviewed, yet some claim to be authoritative solutions.
Tizhoosh draws attention to the validity of this data. In one instance, he saw that the data included a pneumonia case from a paediatric patient. He cautions that, “AI is neither a ventilator nor a vaccine nor a pill; it is extremely unlikely that the exhausted radiologists in Wuhan, Qom or Bergamo download the Python code of our poorly trained network (using insufficient and improper data and described in quickly written papers and blogs) to just obtain a flawed second opinion.”
He concludes that the AI developments must come after appropriate images are made available by hospitals, that ethics approval is received and the data is properly de-identified.
