Pipeline overview
CardiacNexus converts UK Biobank CMR inputs into modality-specific segmentations, tracking artifacts, feature CSV files, NPZ time series, QC visualizations, and aggregate phenotype tables.
- Modality
- Multimodal CMR
- Pipeline step
- Step 1 preparation, Step 2 segmentation, Step 3 modality-specific feature extraction, Step 4 combined features and aggregation
- Outputs
- Per-subject features, aggregate phenotype tables, time-series files, QC artifacts
- Maturity
- Source-audited overview page
Current workflow
| Stage | Current role | Main files |
|---|---|---|
| Preparation | Creates analysis-ready subject directories and NIfTI inputs from UKB source imaging | step1_prepare_data_cmr.py, scripts/prepare_data.py |
| Segmentation | Runs modality-specific segmentation wrappers for cine, aorta, flow, LVOT, and native T1 inputs | step2_segment.py, src/segmentation/** |
| Feature extraction | Computes modality-specific phenotype rows, time-series files, and QC plots | step3_extract_feature_separate.py, src/feature_extraction/** |
| Combined features | Computes cross-modality rows such as valve/AVPD/IPVT and corrected native T1 | step4_extract_feature_combined.py, src/feature_extraction/Combined_Features/** |
| Aggregation and validation | Concatenates or documents output contracts for downstream analysis | scripts/aggregate_csv.py, docs/data/**, website/scripts/** |
QC dependency
Many phenotypes depend on segmentation quality, cardiac frame selection, temporal smoothing, geometry metadata, BSA, ECG timing, pressure data, or DICOM acquisition parameters. The output row alone is not the full provenance.
Extraction layers
Structural phenotypes describe size and geometry: ventricular volumes, myocardial mass, wall thickness, atrial diameters and volumes, aortic diameter/length/curvature/torsion, and valve/root diameters.
Functional phenotypes describe dynamics: stroke volume, ejection fraction, filling and emptying rates, strain, torsion, recoil, distensibility, phase-contrast flow, regurgitant fraction, flow displacement, and cross-chamber coupling.
Tissue phenotypes currently describe native T1 and blood-corrected native T1. ECV remains documented as clinical context only unless a current pipeline contract emits contrast-dependent ECV rows.
What source-audited means here
A source-audited page is checked against the current implementation and data registries, but it is not a guarantee that every subject row exists. Some rows are conditional on BSA, ECG, pulse pressure, valid timing, QC success, or backend peak detection. Public pages should name these conditions instead of silently implying complete coverage.
Source audit
- Step roles were checked against the current
step*.pyorchestration files and the repository architecture rules. - Feature families were checked against the current source tree under
src/feature_extraction/**. - Output-contract and validation expectations were checked against
docs/data/**andwebsite/scripts/**. - Textbook context boundary: broad clinical textbook context is not surfaced here because this page documents pipeline architecture rather than disease interpretation.