Cardiac MRI Annotation: Chamber, Vessel and Function Labelling

Why Cardiac MRI Is the Most Complex Medical Imaging Annotation Task

A cardiovascular magnetic resonance (CMR) examination differs from a static radiology study in a fundamental way: the heart moves. A typical CMR protocol for cardiac function assessment produces 20–50 temporal frames per slice, 8–15 short-axis slices, plus long-axis views and additional sequences for tissue characterisation. The total image count for a single study ranges from 400 to over 1,000 frames — all of which must be considered in aggregate when annotating cardiac structure boundaries.

The annotation challenge compounds because each MRI sequence has different tissue contrast. The myocardium appears bright on T2-STIR in acute myocarditis, dark on T1 mapping in iron overload, and bright with focal enhancement on late gadolinium enhancement (LGE) in myocardial infarction. An annotator who can reliably identify ventricular endocardium on cine sequences cannot necessarily distinguish artefact from true scar on LGE without subspecialty CMR training.

According to a 2024 meta-analysis in European Heart Journal — Cardiovascular Imaging, inter-observer variability in left ventricular ejection fraction (LVEF) measurement from CMR is approximately 3–5 percentage points among expert cardiologists — small enough to be clinically acceptable but large enough to require multi-reader consensus protocols when producing AI training labels. AI models trained on single-reader CMR annotations have demonstrated systematic LVEF biases matching the individual annotator's measurement tendencies, limiting generalisability across clinical sites.

What Gets Annotated: Structures and Measurements

Cardiac MRI annotation tasks fall into three categories based on the AI model's clinical objective.

Functional segmentation (volumetry and EF)

The most common cardiac AI application is automated LVEF measurement for heart failure and chemotherapy cardiotoxicity monitoring. Annotation involves contouring the left ventricular endocardium and epicardium on cine SSFP short-axis slices at end-diastole and end-systole. The right ventricular endocardium is also contoured for biventricular function assessment — a more challenging task because the RV has complex trabecular anatomy without a sharp inner wall.

Key annotation decisions that must be standardised across annotators:

• Papillary muscle inclusion: Whether papillary muscles are included in or excluded from the LV blood pool (different conventions affect LVEF by 2–4 percentage points)
• Basal slice selection: Identifying the most basal slice where ≥50% of the circumference is myocardium rather than atrioventricular valve plane
• Trabeculation classification: LV non-compaction and some dilated cardiomyopathies have heavy trabeculation — the annotation protocol must specify the compacted/non-compacted boundary rule
• End-diastole frame selection: Whether to use the R-wave-triggered frame or the visually largest LV blood pool frame, which may differ in arrhythmia patients

Tissue characterisation (LGE, T1, T2)

Late gadolinium enhancement (LGE) imaging is the reference standard for myocardial scar and fibrosis detection. Annotation requires a cardiologist to delineate enhancing regions within the myocardium, classify them by pattern (subendocardial vs. transmural infarct pattern, mid-wall fibrosis in cardiomyopathy, epicardial enhancement in myocarditis), and measure the percentage of scar burden per segment (AHA 17-segment model).

T1 mapping annotation involves delineating the myocardium on T1-map reconstructions and computing native T1 values per segment — these are absolute quantitative measurements where the annotation boundary has a direct effect on the measured value, making precision more critical than on cine sequences. T2-STIR and T2* annotation is used for myocardial oedema (acute myocarditis, myocardial infarction) and iron overload (haemochromatosis, transfusion-dependent anaemia) respectively.

Structural and vascular annotation

Congenital heart disease AI models require annotation of complex three-dimensional anatomy: atrial septal defects, ventricular septal defects, pulmonary artery anatomy, and aortic valve morphology. These annotations are performed on 3D volumetric sequences and require paediatric cardiologists or congenital heart disease subspecialists — not general adult cardiologists. Phase-contrast velocity mapping (4D flow) annotation involves segmenting vessels and labelling flow direction planes for haemodynamic analysis.

The Multi-Sequence Annotation Protocol

A comprehensive cardiac AI training dataset typically requires a layered annotation protocol spanning three to five MRI sequences per study. Each sequence has its own annotation guidelines, acceptable tools, and quality thresholds.

Not every study needs all sequences annotated. A cardiac function AI model targeting LVEF measurement needs cine annotation only. A cardiomyopathy AI that classifies hypertrophic vs. dilated vs. ischaemic patterns needs cine plus LGE plus T1 mapping. Defining the sequence scope upfront — before annotation begins — is the most important cost control lever for cardiac AI development.

Case Study: Building a Cardiac Function AI From Scratch

A cardiology AI company developing an automated CMR reporting tool for community hospitals needed a training dataset of 3,200 CMR studies covering heart failure, ischaemic cardiomyopathy, and structurally normal hearts. Their initial dataset of 800 studies had been annotated by three general radiologists without CMR subspecialty training. On internal validation, the LVEF model performed well (mean absolute error 4.1 EF%), but external deployment at a tertiary cardiac centre revealed a systematic bias: the model overestimated LVEF by 6–8 percentage points in patients with heavy trabeculation (a subgroup comprising approximately 15% of clinical referrals).

Audit of the training annotations revealed that the three general radiologists had inconsistently applied the papillary muscle exclusion rule and had systematically over-segmented the LV blood pool in trabeculated patients — including LV trabeculations in the blood pool rather than the myocardium. This produced artificially high LVEF values in the training labels for this subgroup.

A reannotation project with four board-certified cardiologists — each with EACVI Level 2 CMR certification — was commissioned using a standardised protocol:

• Papillary muscles included in myocardium (ESC/EACVI guideline default)
• Trabeculation classified per LVNC criteria: compacted layer only traced for epicardium
• Dual annotation for all 3,200 studies; adjudication for cases with LVEF disagreement >5 points absolute
• Post-annotation LVEF correlation with clinical CMR reports used as an external quality check

Results after retraining on adjudicated cardiologist labels:

The reannotation took 14 weeks with a panel of four cardiologists. The retrained model entered clinical validation at three cardiac centres and subsequently received CE mark under the EU MDR pathway. The original general-radiologist-annotated model would not have met the accuracy specifications in the CE clinical evaluation.

Temporal Annotation: Handling the Cardiac Cycle

Cine MRI captures the heart at 20–50 temporal phases across the cardiac cycle. Most cardiac AI models require annotation at a minimum of two time points — end-diastole (ED, maximum LV blood pool volume) and end-systole (ES, minimum LV blood pool volume). Some models, particularly those training for myocardial strain analysis or wall motion abnormality detection, require annotations at every temporal frame.

The technical challenge is annotation propagation: once the annotator draws contours at ED and ES, intermediate frames are typically interpolated from these anchor points. The quality of this interpolation determines whether the model learns realistic cardiac motion trajectories. Linear interpolation (the simplest approach) fails at peak systole and isovolumetric contraction/relaxation phases because the LV volume change is not linear across the cardiac cycle.

Production cardiac AI annotation pipelines use one of three approaches:

HIPAA, De-identification, and FDA Compliance for Cardiac MRI Data

CMR studies sourced from clinical archives are DICOM files containing patient-identifiable information: name, date of birth, MRI accession number, and often the performing institution and requesting physician. Before these studies are used for AI annotation, HIPAA Safe Harbour or Expert Determination de-identification must be applied — removing or transforming all 18 HIPAA-defined identifiers from DICOM metadata.

DICOM de-identification for cardiac MRI has specific complications:

• DICOM tags embedded in overlay data and private tags (vendor-specific) must be inspected separately — standard de-identification tools frequently miss private tags
• Study date must be shifted consistently across all sequences for the same patient so that temporal relationships (pre/post gadolinium, stress/rest) are preserved after de-identification
• Pixel-burned patient information (some older scanners burned patient name into the image pixel data) requires pixel-level scrubbing — metadata-only de-identification is insufficient
• Gadolinium injection timing metadata (trigger delay, gadolinium dose) should be retained for LGE analysis but may need to be generalised if it enables patient re-identification in small clinical datasets

For FDA 510(k) or De Novo submissions, the annotation records must comply with FDA 21 CFR Part 11: a tamper-evident audit trail per annotation region, documenting the cardiologist annotator's identity and credential, the annotation software version, and the timestamp of each contour and modification. The same cardiologist identity and credential records must appear in the training data appendix submitted to FDA.

A 2024 analysis of FDA-cleared cardiac AI submissions (ECRI Institute) found that 34% of deficiency letters cited inadequate documentation of training data annotation methodology — specifically the absence of annotator credential records and inter-observer agreement statistics. These are administrative failures, not technical ones, and they are entirely preventable with a well-designed annotation data management process.

The Cardiac AI Market and Why Annotation Quality Determines Product Outcomes

The global cardiac AI market was valued at USD $1.8 billion in 2024 and is projected to grow at 26% CAGR to USD $7.4 billion by 2030 (MarketsandMarkets, 2025). The primary drivers are LVEF automation for heart failure monitoring, cardiomyopathy classification, and CMR reporting efficiency in tertiary centres where CMR volumes exceed reporting capacity.

The current leading cardiac AI products — including those with CE mark and FDA clearance — share one characteristic: their training datasets were annotated by credentialed cardiologists under validated multi-reader protocols, not by general radiologists or non-clinician annotators. This is not a regulatory requirement per se; it is a clinical performance requirement. General-annotator cardiac MRI labels produce models with systematic biases that fail clinical evaluation, regardless of their performance on internal validation sets.

For AI teams building in cardiac MRI, the annotation quality decisions made at the start of a project determine the clinical performance ceiling. Investing in cardiologist annotators, multi-reader protocols, and Part 11-compliant provenance from the outset is substantially cheaper than the alternative: a failed clinical validation, regulatory deficiency letters, and an expensive reannotation cycle. The teams that have succeeded in clearing FDA and CE mark have learnt this the hard way; the teams building now can skip that lesson.

Frequently Asked Questions