models.academica.sh
01Specimen Acquisition & Input Controls
Active Foundation Architecture44.2 ms · pure ONNX
v6 · DINOv2 + SigLIP2 + RF-DETR + SAM 2.1 Procrustes
U1-JPack Procrustes Bridge + Multi-Granular Spatial Pooling + Fisher-ZCA Pullback + 4-Subspace GS
Patient-Disjoint Clinical Specimen Pickerdanjacobellis/chexpert · ood_src_patient64619
Active TB Chest X-Ray
Acquisition Modality GateIn-distribution gate threshold 0.098
Viewport Overlay ModeRF-DETR + SAM 2.1 Procrustes
Longitudinal Prior Baseline Lock256-D SupCon Re-ID delta
Single exam mode · Zero prior registration
Verified Generalization BenchmarksHeld-out hospital test_ood
Modality Gate99.96%
Lesion mIoU0.819
RF-DETR AP500.742
OOD Shrinkage0.098
02Diagnostic Viewport & 4 Clinical Answers
STAGE 1Interactive Viewport (Active TB Chest X-Ray)
boxes-mask
Active TB Chest X-Ray5 cm physical scale
Radiology · SAM 2.1 ProcrustesZero Focal Masses
STAGE 2Calibrated Clinical Output
CIC 95%: [0.88, 0.95]
S1S1 Structural Geometry (Look)[256D]
94.0%
S2S2 Diagnostic Semantics (Read)[128D]
89.2%
S3S3 Longitudinal Re-ID (Change)[256D]
96.4%
S4S4 Epistemic Abstention (Defer)[64D]
98.8%
1. LOOKSpatial / Acoustic Localization Target
Right apical fibro-cavitary opacity with blunting of right costophrenic angle consistent with pleural effusion
2. READDifferential Probability & Clinical Staging
Probable Active Pulmonary Tuberculosis with Secondary Pleural Effusion
Clinical FindingCalibrated Posterior (p)Clopper-Pearson 95% CICStatus
Pleural Effusion & Apical Consolidation0.942[0.88, 0.95]Confirmed Post-Calibration
Secondary Finding Focus0.084[0.04, 0.14]Baseline Sub-threshold
Concomitant Clinical Risk0.031[0.01, 0.08]Negated / Ruled-Out
3. CHANGELongitudinal Baseline Lock & Quantitative Biomarkers
+0.142 volume progression compared to baseline prior film (interval increase in effusion depth)
Re-ID Cosine Delta0.1420
SupCon Representation256-D Whitened
Baseline StatusLocked (Prior Exam Pair)
S9-ClinCalc Result: CHA2DS2-VASc Atrial Fibrillation Stroke Risk ScoreHigh Risk
3 (max 9)
High stroke risk; oral anticoagulation (apixaban, rivaroxaban, or edoxaban) strongly recommended
CHF(0) + HTN(1) + Age(68:1) + Diabetes(0) + Stroke(0) + Vasc(1) + Sex(male:0) = 3 / 9
Lip et al. Chest 2010 / European Society of Cardiology Guidelines
4. DEFERRisk-Controlled Epistemic Safeguards
Accepted by Mahalanobis OOD gate (DM = 1.34 < 3.50); scanner transfer artifacts suppressed via U5-LEACE
Modality Compatibility Gate99.96% Confidence (Pass)
Mahalanobis Epistemic OOD Distanced = 1.340 (threshold 0.380 · Pass)
Confounder Scrubbing (U5-LEACE)Hospital GE Discovery XR656 transfer curve, Source patient gender demographic bias
Block Gram-Schmidt Certificate||W_Sa W_Sb^T||_F = 7.10e-09 < 1e-07