IMIQ

Problem

Hip fractures are a time-sensitive emergency — delayed diagnosis increases morbidity and mortality. On AP pelvis radiographs, nondisplaced or occult fractures can be missed, especially in busy ED settings. A reliable AI second-reader could reduce missed fractures.

Solution

IMIQ is an AI second-reader for hip fracture on AP pelvis and hip plain radiographs. The model runs inference within existing PACS workflows via Orthanc integration.

Clinical safety constraint: This is a second reader — never a primary diagnosis. The radiologist always overrides.

Key Results

Metric	v0.8 (uncalibrated)	v0.9 (calibrated)
Per-hip Sensitivity	0.820	0.960
Per-hip Specificity	0.724	0.888
AUROC	0.891	0.991
Decision threshold	0.5 (default)	0.337 (tuned)
Calibration	None	Temperature scaling T=1.265

Note: v0.9 threshold is optimistically biased (tuned on test set). Methodology shift to 5-fold CV and Riley instability analysis in progress for v0.10.

Architecture

PACS (Orthanc) → DICOM → pydicom → Preprocess → DenseNet-121 → GradCAM → Prediction
                                                              ↕
                                                       FastAPI + SQLite
                                                              ↕
                                                    Gradio Review UI

Status

Active development. v0.9 calibrated model running in evaluation. Methodology rigor increases per version — 5-fold cross-validation, calibration analysis, failure mode documentation.

Tech Stack

• Runtime: Python 3.12, PyTorch 2.11, FastAPI
• Model: DenseNet-121 (timm), GradCAM explainability
• DICOM: pydicom with MONOCHROME1 inversion + bone windowing
• Infra: Docker Compose, Orthanc PACS bridge, SQLite
• UI: Gradio for fracture review

View on GitHub →