AI Medical Imaging Tools Consistently Swap Left and Right in Scan Reports

A developer building an AI tool to interpret MRI and CT scans discovered that the most frequent error across four vision models — Claude, Gemini, Grok, and MedGemma — was incorrectly labeling left and right anatomical sides. Testing eight real studies against verified radiologist reports revealed mistakes such as a right-kidney cyst placed on the left and a medial meniscus tear attributed to the lateral meniscus. The root cause is that models trained on everyday photos describe what appears on screen, while radiology convention displays axial scans mirrored — the patient's right appears on the image's left. The developer addressed this by cross-checking each model's coordinate output against the NIfTI affine matrix, which maps voxel positions to real-world space where positive x always corresponds to the patient's right side. A language-agnostic regex covering multiple languages and an 8 mm midline dead zone were also added to reliably flag any mismatch between the model's written description and its coordinate data.