XR Chest View Position Classifier

Chest radiograph headers don't reliably carry a usable view position — empty, vendor-specific, free-text, or wrong — silently exposing every downstream model that assumes a known view.

A ResNet-50 five-class classifier (PA, AP, LL, RL, Other) behind a metadata fast path: where the header is reliable the model is bypassed entirely. An earlier two-stage cascade was consolidated into this single model. Augmentation was deliberately constrained because flips and rotations corrupt view labels. Deployed on GCP Healthcare API DICOM stores with eligibility gating, prior-result reuse and DICOM SR output.

• 95.7% accuracy (99.2% top-2) internally; near-perfect left/right lateral discrimination (F1 0.991 / 0.993).
• 91.1% accuracy (95% CI 0.887–0.935) on a facility-disjoint bench set.
• Proved the metadata systematically wrong — 148 studies tagged “LL” were actually RL — driving a corrective annotation protocol.
• Two models collapsed into one; inference cost reduced by the metadata fast path.

• Commissioned the metadata audit instead of accepting DICOM tags as ground truth — reframing the project from “build a classifier” to “our labels are wrong, prove it.”
• Approved consolidating the cascade into one model to cut long-term serving complexity.
• Required the metadata-first hybrid on cost and latency grounds.