Why anatomy-aware localization beats a classification score

Classification tells you what, not where

Broad multi-disease classifiers can score well across many abdominal findings [1], but a study-level probability does not tell a radiologist where to look. In emergency CT triage, where to look is half the value. The other common answer, pixel-level detection metrics like IoU and mAP, is precise but speaks a language that never appears in a radiology report.

The nine-region grid

Atlas maps every detection to the nine-region abdominal grid: epigastric, umbilical, and hypogastric down the middle, flanked by the paired hypochondriac, lumbar, and iliac regions. This is the vocabulary used in structured radiology reporting [2][3] and the same language clinicians use to describe where pain and pathology sit, which makes the output legible at a glance.

Measuring it without fooling yourself

It is easy to report a localization number that is quietly inflated by only counting cases the model already detected. Atlas reports two figures. Among detected cases, region accuracy was 99.5 percent. Counting every positive case, including the ones the model missed, it was 90.9 percent. Reporting both prevents selection bias on detected cases from flattering the result.

The region readout also has to be stable. Under random perturbation of the body envelope used to build the grid, 95.1 percent of annotations kept their region at plus or minus 5 mm, falling gracefully as the perturbation grew. The grid is coarse by design, which is exactly what makes it robust for region-spanning acute emergencies.

Precedent

This is not a fringe idea. Region-level agreement has been used as a clinically meaningful localization endpoint before, for example in detecting the transition zone in small bowel obstruction [4], where anatomic-region agreement was more useful than pixel overlap alone. Atlas applies the same logic to six acute abdominal emergencies.