My MRI interpretation vision model pointed at the right kidney and wrote "left"
I build a tool that reads MRI and CT scans with general vision models and explains the findings in plain language. The most repeated error in that pipeline isn't a missed tumour or an invented measurement. It's left and right . We ran eight real studies through four readers (Claude, Gemini, Grok and Google's MedGemma), with the radiologist's signed report as ground truth. Here's what came back: a…
A research team tested four AI models—Claude, Gemini, Grok, and MedGemma—on MRI scan interpretations, aiming to identify common errors. The models consistently misidentified left and right anatomical structures. In one instance, MedGemma placed a cyst in the left kidney but labeled it as right. Another AI reported a medial meniscus tear as lateral, and a left paracentral disc extrusion as right-sided. This inconsistency poses a significant issue, as a left/right swap is considered a "never event" in medical reporting.
The source material reveals that the AI models are accurate in their coordinate measurements but incorrect in their descriptions. For instance, MedGemma correctly placed its crosshair 17–26 mm to the right of the midline but labeled the finding as "slightly to the left." This discrepancy arises because MRI scans are displayed as if the viewer is standing at the patient's feet, looking up, meaning the patient's right side appears on the left side of the screen.
The AI, trained predominantly on ordinary photos, interprets the image as it sees it, leading to the side mistake.
The source also explains the underlying technical cause. MRI scans are converted into NIfTI files, which contain affine matrices mapping voxel indices to world space (RAS: right, anterior, superior). By analyzing the world space coordinates, a function called `sideOfPoint` determines whether a finding is on the left or right side of the patient. The `OFF_MIDLINE_MM` constant defines a 8 mm dead zone to avoid false side assignments due to minor coordinate errors or floating-point noise.
The research also addresses language differences in MRI reports. The models' descriptions of side are written in English, so translations may inadvertently misinterpret the side information. To counter this, the researchers use prefix patterns specific to various languages to identify side words (right, left, etc.) in the reports. They account for inflection patterns in each language and avoid matching false positives, such as the word "bright" or Polish "prawdopodobnie."
Overall, the study highlights the importance of accurately interpreting left and right anatomical structures in medical imaging reports, suggesting that a model-free check can effectively identify these errors without relying on the AI itself.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
