Heart segmentation is labeling the chambers and walls of the heart on medical images so you can measure volume, ejection fraction, wall thickness, and scar — not just watch a beating blob. The usual input is short-axis cine MRI. CT and echo show up too. The output is a mask of the left ventricle (LV), right ventricle (RV), myocardium, and — when the study supports it — atria, valves, and coronaries.
It is not the AHA 17-segment wall-motion model. That language lives at cardiac segments. It is not a generic “what is medical image segmentation” explainer — that is medical image segmentation. How you train a 3D net on your own volumes is nnU-Net for medical image segmentation.
What you actually label
The job is structure-by-structure, not “the heart” as one blob.
- Left ventricle. Cavity volume across the cycle gives ejection fraction — the number most reports actually want. Wall thickness comes from the myocardium around it.
- Right ventricle. Thinner wall, more shape variation, worse contrast than the LV. Still a chamber you measure when the question is right-sided load or congenital repair.
- Myocardium. The muscle, not the blood pool. Scar, hypertrophy, and T1 / T2 mapping need this mask, not just the cavity.
- Atria. Thin walls. Harder than ventricles. The mask is for ablation planning and atrial volume, not for EF.
- Valves. Leaflets and annuli for surgical planning. A different contrast problem than a blood-pool chamber.
- Coronaries. Vessel lumen and wall, usually on CTA, not cine MRI. Stenosis work is a vessel job sitting next to chamber work, not the same mask.
Short-axis cine MRI is the common sequence because the blood pool and myocardium have enough contrast to draw the LV and RV. Poor blood-pool / myocardium contrast, and large shape differences between patients, are why a threshold does not finish the job.
Why it is hard
The heart moves. Cardiac contraction and respiration both smear edges. A mask drawn on a blurred frame is a wrong volume. Motion correction and gating are preprocessing, not a slogan.
Pathology breaks the atlas you trained on. Scar after infarct can look like healthy muscle. Congenital shapes do not match a “normal LV” prior. Pacemakers and prosthetic valves add metal artifact. Those are the cases a model trained on clean volunteers misses.
U-Net (encoder–decoder, skip connections, sometimes attention) is the architecture papers keep using for this. Data augmentation and transfer learning exist because labeled cine stacks are expensive. None of that is a reason to publish a fake Dice table. Dice against a human mask is how a paper scores one hospital’s scanner. A 0.90 on that set is not a number you can paste onto the next vendor.
Where the mask is used
| Question | What you segment | What you do with the mask |
|---|---|---|
| Heart failure | LV / RV cavity, myocardium | Volumes and ejection fraction; wall thickness over visits |
| Cardiomyopathy | Myocardium, scar | Extent of damage; whether a device or a cut is even on the table |
| Congenital repair | Whole-heart 3D (chambers + connections) | A model the surgeon can rotate before the case |
| Rhythm (EP) | Atria, pulmonary veins | Ablation targets; not a wall-motion 17-segment bullseye |
| Coronary disease | Coronary lumen / wall on CTA | Narrowing you can point at. Clinic CTA work is a case study, not this page. |
Fusion with echo or CT is ordinary: MRI for the chamber mask, another modality for calcium or wall motion. The mask is still a measurement. It does not sign the report.
What this page is not
- Not AHA-17. Basal / mid / apical wall-motion language and coronary territories sit on cardiac segments.
- Not a methods catalogue. Threshold, region-grow, U-Net-as-idea: medical image segmentation.
- Not a train loop. Folder layout,
plan_and_preprocess,nnUNetv2_train: nnU-Net for medical image segmentation. - Not lung or mandible. Bronchopulmonary segments are lung segments on CT. Panoramic jaw work is mandible segmentation from panoramic X-ray.
PYCAD builds the imaging side of this when the mask has to live in a clinic app. Case studies.