Reading a medical image: windows, search and a second look

Modern medical images are digital. They move between scanners, archives and reading workstations using DICOMDICOM (Digital Imaging and Communications in Medicine)The international standard for medical images and related information, first published in 1993 and recognised by ISO as ISO 12052., the international standard for medical images and related information. DICOM was first published in 1993, is recognised as ISO 12052, and is a large part of why film could be replaced by fully digital workflows.^[1]

Each pixel of a CT image stores a number on the Hounsfield scaleHounsfield unit (HU)The scale for CT values. By definition water is 0 HU and air is −1000 HU; bone ranges from several hundred to several thousand HU.. By definition water is 0 HU and air is −1000 HU, while bone ranges from several hundred to several thousand HU.^[2]

That range is far wider than the grey levels a reader can tell apart on screen, so readers choose a window using two settings: Window levelWindow level (WL, window centre)The HU value at the middle of the window, shown as mid-grey. and Window widthWindow width (WW)The span of HU values shown as shades of grey. A wide window shows many tissue types at once; a narrow one exaggerates small differences.. Values between level − width/2 and level + width/2 are spread across the grey scale; anything above is shown white and anything below black. For head CT, for example, a “brain window” and a “bone window” are two important settings.^[2]

Radiology textbooks often recommend a systematic approach: checking an image in a fixed order so that no region is skipped.^[3] The evidence is more nuanced than the advice. In two eye-tracking experiments with students, residents and radiologists reading chest radiographs, Kok and colleagues found that the data questioned the assumption that systematic viewing leads to more complete coverage and therefore better performance. Experienced readers did inspect images more systematically, but students did not benefit from being trained to view systematically.^[3]

A systematic review of eye-tracking studies describes expert search as global-focal: a fast overall impression that signals possible abnormalities, followed by a slower, detailed search of those areas. In chest CT viewed as a stack of slices, “drilling” — scrolling up and down while focusing on one region — was associated with higher expertise.^[4]

Finding one abnormality can make a second one in the same image easier to miss. Radiology has studied this for more than 50 years under the name Satisfaction of searchSatisfaction of search (subsequent search miss)A drop in the chance of noticing a further abnormality once a first one has been found in the same image.; in cognitive science it is now often called a subsequent search miss.^[5]

Several explanations have been proposed: readers may stop searching too soon once something is found, may be primed to look for more of the same kind of target, or may have fewer mental resources left for the rest of the search. In the studies reviewed by Adamo and colleagues, measures such as checklists and computer-aided detection did not reliably remove these errors.^[5]

Because any single reader can miss something, breast screening programmes in many European countries use Double readingDouble readingTwo readers interpret the same images independently, with consensus or arbitration when they disagree. It is widely used in European breast screening.: two readers interpret the same mammograms independently, and disagreements are resolved by consensus or by a third reader.^[6] The European Commission Initiative on Breast Cancer suggests double reading, with consensus or arbitration, over single reading for screening mammograms — a conditional recommendation based on moderate-certainty evidence.^[7]

In a review of the evidence, Taylor-Phillips and Stinton report that in each of five studies of cancer detection, the detection rate was somewhat higher with double reading (5.2 to 8.8 per 1,000 screens) than with single reading (4.8 to 8.0 per 1,000 screens). They also note that the effectiveness of double reading depends on whether the two readers are blinded to each other’s decisions and how disagreements are resolved, and they discuss AI as a possible second reader.^[6]