Area A—Regulated Diagnostic AI

XR Knee View Classifier & Abnormality Detector

Two cooperating models · LLM-mined ground truth

When no labels existed, an LLM turned 31,000 narrative reports into training supervision.

5-class view accuracy: 99.2%; 5-class view accuracy
reports mined into labels: ~31k; reports mined into labels

01/Problem

Musculoskeletal archives are enormous and inconsistently tagged — a laterality audit found 20 of 120 sampled studies wrongly tagged. An abnormality detector was needed with no structured labels in existence.

02/Approach

A five-class view classifier on InceptionResNetV2 and a multi-label detector across bones, joint, soft tissue and effusions. Ground truth was generated by applying an LLM to ~31,000 free-text reports, with an explicit “Unknown” state and ambiguity backlog. With heavy class imbalance, models were selected on macro G-mean, not accuracy.

03/Outcome

99.2% test accuracy on five-class view classification.
Macro AUC ≈ 0.78 / G-mean ≈ 0.72 on four-space detection, beating ResNet-50 baselines (≈ 0.70–0.73).
~985,000-DICOM corpus profiled, curated to ~31,000 studies with report linkage.

04/My role

Directed LLM-assisted labeling as an organizational method — a concrete instance of the 50% annotation-effort reduction.
Required the “Unknown” state so uncertain labels went to humans instead of being guessed.
Directed selection on macro G-mean so clinically important minority findings couldn't hide in aggregate accuracy.

Stack

TensorFlowInceptionResNetV2GPT-3.5MLflow