If you take a look at your desk or bedroom, you’ll probably know right away if something is out of place, even if you can’t pinpoint exactly what’s wrong without closer inspection. That’s because humans have the ability to quickly grasp the gist of a situation with just a quick glance.
A recent study published in PNAS shows that this ability goes far beyond everyday practice. Radiologists who specialize in detecting breast cancer can distinguish between normal and abnormal mammograms in just half a second. But they may not even need to look at the cancerous tissue for that.
The authors of the paper were interested in a phenomenon known as global processing, in which a quick glance at a large image provides insight into its significance. They gave radiologists a moment to look at breast tissue images and compared the results of the radiologists’ insights with carefully analyzed images.
Based on a half-second look at mammogram images, the radiologists were able to detect cancer at a rate greater than chance. This finding supports anecdotal accounts from radiologists, who often report that an image will appear “bad” to them before they even identify the abnormality in the image. In this scenario, it appears that the radiologists are responding to an overall signal of cancer in the image, and this signal is not necessarily associated with the actual location of the cancerous growth.
The real surprise, however, came when the researchers discovered that radiologists could identify a person with cancer by showing them an image of the breast unaffected by abnormal growths. (Again, this is just faster than chance.)
These more than promising results do not seem to be based on symmetry between the two breasts or on breast density. The radiologists don’t appear to be comparing a healthy breast to an unhealthy one to detect the presence of a tissue abnormality – instead, they are looking at a more fleeting quality in the image.
This idea is supported by the localization experiments. For example, when the radiologists were asked to identify the location of the abnormality with the same quick glance, their detection ability fell within the margin of chance.
Of course, this finding doesn’t mean we can hope to get good diagnoses simply by flashing images for radiologists. While the doctors can spot cancer above chance in these quick glances, they certainly can’t read the mammograms perfectly in such a short time — the authors don’t suggest that these skills could replace more robust screening methods. Instead, they’re interested in trying to identify what they call the “core” signal, the image quality that tells the radiologists something is wrong with the tissue.
The authors recommend further research into this phenomenon. In these initial studies, they have not been able to fully isolate the image features that indicate abnormalities to radiologists. The authors say a better understanding of what these people do could lead to better computerized screening methods that could complement a radiologist’s input.
PNAS2016. DOI: 10.1073/pnas.1606187113 (About DOIs).