MICCAI 2018 Daily - Tuesday

4 Tuesday Oral Presentation Mining Clinical Reports to Gain a Broad Understanding of Chest X-rays . Jonathan tells us that Zebra Medical has access to millions of medical images and reports. Most of these reports are in Hebrew, because their partner is an Israeli health provider. By analysing all those reports and figuring out what findings exist in them, they have created a one million study dataset. They focused specifically on chest X-ray, which is the most widespread medical image and harder to interpret, and now have the largest chest X-ray dataset that has been constructed, with around one million medical images. This is 2D work, although they have both frontal and lateral views. A lot of people use only the frontal X-ray image, but they used both, and he says it has been interesting to see in which findings the lateral view helps. They have the analysis to show it. The reports themselves turned out to be simple enough to analyse, but they had to do a rigorous process of figuring out exactly what the findings are that people look for in X-rays. For that, they used an expert radiologist who reviewed the reports and every time she saw a new finding she would add it to a list. Eventually they came up with 40 different findings that are most prevalent in chest X-ray reports. Jonathan tells us they used noisy labels. They read the reports and used an NLP algorithm to figure out what the findings were based on those reports, but they didn’t have any further analysis to confirm if this was true. Inherently, in the process, there was noise. While looking at the graphs of their performance, it was easy to see that for a lot of the findings that indicate the presence of artificial objects – such as pacers and implants and stitches that are used in the With Jonathan Laserson Jonathan Laserson is the Lead AI Researcher at Zebra Medical. They look at medical images – such as X-ray, CT, MRI, mammography – and try to say what an expert radiologist would say when looking at those medical images. He speaks to us ahead of his oral and poster today. . “ The value that we get from it by having all of our million studies already labelled far outweighs the noise… ”