AIMA 2026 projects ✦ Project IV ✦ Cardiovascular AI
Reading the flow: AI on dynamic coronary angiograms
An angiogram is a film, not a photograph. Contrast flows in, fills the arteries and washes out, and where it lingers may show which segments are under hemodynamic stress. Our cardiologists and physicists propose that a backward pressure wave leaves this footprint. We will build the AI that can measure it: segmentation that holds on every frame, and contrast-transport maps that can be tested against how lesions grow.
In partnership with
- Methodist HospitalsIndiana
From one frame to the whole run
Kymograph · contrast along the vessel over time
Coronary angiogram: Bleiglass, Wikimedia Commons, CC BY-SA 3.0. This is one real frame. The inflow, washout, retention zones and AI layers are simulated on it for illustration; they are not measurements.
What is known
- A backward-travelling wave in the coronary arteries is established: in humans, a dominant backward wave carried about 30% of all wave energy1.
- Low wall shear stress and large plaque burden independently predicted which lesions narrowed in 374 re-imaged patients2. This is the benchmark a new angiographic marker must beat.
- Angiography already measures flow: the TIMI frame count3, and QFR, which agreed with invasive FFR in 92.7% of 328 vessels4.
The gap we target
Public vessel masks cover only the best-filled frame or two of each view5; the full-video set CADICA marks stenoses with boxes, not vessels6, and video methods are only now appearing7. Stenosis readers are already strong, with AUC 0.869 at an outside hospital8, but they score how narrow a vessel is, not how contrast moves through it. No one yet measures washout segment by segment, frame by frame. That is the instrument the hypothesis needs, and it is useful even if the hypothesis fails.
Segment the whole run
Reproduce nnU-Net9 on ARCADE, then make segmentation stable across full sequences, including the faint washout frames. Label a small washout-frame test set.
Measure transport
Centerlines, time-density curves and kymographs per segment: arrival and clearance times, checked against manual frame counts and their known variability3.
Test it honestly
With the clinical team, a pre-registered pilot that compares retention maps with plaque and shear-stress measures. Release code and the test set.
Data
You will learn
- Video and temporal segmentation
- Topology-aware metrics beyond Dice
- Quantitative angiography and coronary physiology
- Pre-registered, honest validation
Good fit if you
- Like computer vision on real, messy video
- Are curious about the physics of blood flow
- Are a medical student who can read angiograms
Data & labels
Sequences, splits by patient, the washout test set.
Segmentation
nnU-Net baseline and the temporal model.
Transport
Centerlines, time-density curves, kymographs.
Validation
Frame counts, physiology pilot and statistics.
The plan · 6 months
From first paper to submission.
Six months, from first paper to submission, with weekly group meetings and a monthly review with mentors.
Onboard
Read 10–15 key papers, set up data access and compute, write a one-page proposal.
Reproduce
Match a published baseline. Share a reproducibility report.
Propose
First version of the method, defended in a mentor review.
Iterate
Ablations, failure analysis, fix what the review found.
Validate
External data or pilot study; freeze results.
Submit
Paper, code release and a talk for the AIMA community.
Project IV · Cardiovascular AI
Want to work on
this question?
Apply through the form by November 7, 2026, choosing Project IV, with an academic CV and a statement of purpose of up to 300 words. Interviews run November 9 to 13, and groups start around November 18. Mentors from Yale University · Harvard University · University of Chicago · Methodist Hospitals, Indiana.
References
- Davies et al. Evidence of a Dominant Backward-Propagating “Suction” Wave Responsible for Diastolic Coronary Filling in Humans, Attenuated in Left Ventricular Hypertrophy. Circulation 2006. DOI
- Stone et al. Prediction of Progression of Coronary Artery Disease and Clinical Outcomes Using Vascular Profiling of Endothelial Shear Stress and Arterial Plaque Characteristics. Circulation 2012. DOI
- Gibson et al. TIMI frame count: a quantitative method of assessing coronary artery flow. Circulation 1996. DOI
- Xu et al. Diagnostic Accuracy of Angiography-Based Quantitative Flow Ratio Measurements for Online Assessment of Coronary Stenosis. JACC 2017. DOI
- Popov et al. Dataset for Automatic Region-based Coronary Artery Disease Diagnostics Using X-Ray Angiography Images. Scientific Data 2024. DOI
- Jiménez-Partinen et al. CADICA: A new dataset for coronary artery disease detection by using invasive coronary angiography. Expert Systems 2024. DOI
- Wu et al. DeNVeR: Deformable Neural Vessel Representations for Unsupervised Video Vessel Segmentation. CVPR 2025. DOI
- Avram et al. CathAI: fully automated coronary angiography interpretation and stenosis estimation. npj Digital Medicine 2023. DOI
- Isensee et al. nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation. Nature Methods 2021. DOI
- Ma et al. Self-Supervised Vessel Segmentation via Adversarial Learning. ICCV 2021. DOI
- Cervantes-Sanchez et al. Automatic Segmentation of Coronary Arteries in X-ray Angiograms using Multiscale Analysis and Artificial Neural Networks. Applied Sciences 2019. DOI