acceptodds
Withdrawn from ICLR 2027

AngioFM: a Vision Foundation Model for X-ray Coronary Angiography

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

X-ray coronary angiography (CAG) presents complex scenes in which coronary vessels, surrounding anatomy, and interventional devices overlap in two-dimensional projections. Cardiac motion, contrast filling, and changing projection geometry further complicate representation learning. We introduce AngioFM, a vision foundation model that uses these structured variations as supervision through feature prediction conditioned on acquisition time, contrast state, and imaging rays. Pretraining combines 31,559 unlabeled coronary videos with projections from 1,000 coronary CT angiography (CCTA) cases, coupling real angiographic dynamics with explicit cross-view, vessel-tangent, and branch supervision. We assess frozen representations using linear probes for controlled comparisons of accessible pathological, anatomical, and temporal information, complemented by CNN decoders for anatomical and device segmentation and a direct centerline-regression readout. With matched readouts and frozen encoders, AngioFM improves coronary-segment parsing over generic image-pretrained and CAG-pretrained baselines, with additional gains in video-based catheter and guidewire segmentation. Linear probes show higher coronary-dominance AUC, more accurate cardiac-phase matching, and competitive lesion localization. Centerline regression shows improved vessel localization in single-view and multi-view settings. Branch-supervision ablations further support improved anatomical discrimination. Together, these findings support learning reusable coronary representations from acquisition physics and vascular structure, providing a foundation for tools supporting angiographic assessment and image-guided intervention.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Vision Foundation Model, Self-supervised Learning, X-ray Coronary Angiography

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_miEKOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

AngioFM is a frozen representation model for X-ray coronary angiography. Its pretraining data include clinical cine sequences and synthetic projections generated from CCTA coronary trees. Acquisition information is supplied to the predictor. The CCTA source contributes additional structural supervision.

The downstream study includes lesion detection, cardiac-phase matching, coronary dominance, segmentation, and centerline recovery. DINO-world and LeVJEPA provide video-pretraining comparisons. DINOv3-L serves as an image-pretrained reference, and an adapted DINOv3-L checkpoint is used in the initialization control.

Performance varies substantially by task. Coronary-segment parsing and centerline recovery provide the clearest positive evidence. Several vessel, device, and lesion results remain close to DINO-world.

Critical strengths

The training design fits the CAG setting. Vessel appearance can change as the projection changes or as contrast fills the coronary tree, even though the anatomy itself has not changed. Cardiac motion creates another source of variation that ordinary image augmentation does not capture.

The CCTA source is useful because it allows controlled projections of a known coronary tree. Structural information can therefore be obtained without manually annotating a large collection of clinical cine frames.

The strongest numerical results concern coronary structure. ARCADE segment Dice is 40.52 for AngioFM, 36.42 for DINO-world, and 34.80 for DINOv3-L. CARDIAG gives the same ranking. Centerline recovery also improves over DINO-world on ARCADE-SYNTAX and CARDIAG.

The paper also reports limitations in the available data rather than hiding them. The dominance benchmark is discussed carefully. The video experiment is likewise not presented as direct proof that the representation has learned temporal reasoning.

Critical weaknesses

The branch ablation changes LbrL_{\mathrm{br}}, but the rest of the synthetic training remains in place. Cross-view training is still active. Tangent supervision remains as well. The reported difference therefore concerns one loss term inside a much larger CCTA-based setup. It does not separate the contribution of the synthetic source as a whole.

The physical-conditioning study has a larger confound. The positional-only control changes several parts of the model at once. Physical ray information is removed together with the acquisition-state input. The geometric attention term changes as well. The manuscript also notes a mismatch in pretraining history and budget.

The reported multi-view Chamfer values are 3.87 and 5.43 pixels. The numerical gap is substantial. The present experiment, however, changes too many things simultaneously to identify its source.

A more informative control would keep the training schedule fixed and alter only one part of the conditioning block at a time. Ray information could be removed first. Acquisition-state input could then be tested separately, followed by the geometric attention term.

The DINO-world comparison introduces another attribution issue. DINO-world uses the private CAG video pool and the adapted initialization. AngioFM additionally uses the CCTA-derived source. The measured difference therefore contains both a method change and a supervision change.

Two missing controls would be particularly useful. One could train with CCTA supervision while omitting physical conditioning. Another could keep the conditioning mechanism but remove the CCTA structural objectives. These controls would make the source of the gain much easier to judge.

Appendix B.2 uses unadjusted two-sided pp-values at the 0.05 level across many reported comparisons. With mm independent tests at level α\alpha, the probability of at least one false positive is 1−(1−α)m1-(1-\alpha)^m. For m=20m=20 and α=0.05\alpha=0.05, the value is about 0.64. The tests here are correlated, so the calculation is only illustrative. Even so, a large collection of nominal p<0.05p<0.05 results is difficult to interpret without a clearly defined primary family of comparisons.

The bootstrap fixes the pretrained encoder and resamples downstream examples. Variation caused by pretraining a new encoder is therefore absent from the reported intervals.

This omission matters when the observed differences are tiny. On single-frame CARDIAG vessel segmentation, the two reported Dice values are 88.11 for AngioFM and 88.05 for DINO-world. The difference is 0.06. XACV shows the reverse ordering, with 86.92 for AngioFM and 87.10 for DINO-world. Checkpoint-to-checkpoint variation could easily be comparable to these margins.

The guidewire experiment has another sampling problem. There are 640 test frames, but only 11 source sequences. Frames taken from one sequence are dependent observations. Treating all frames as separate bootstrap units can make the amount of independent evidence appear larger than it is. Resampling whole sequences would give a more appropriate uncertainty estimate.

The phase-matching result also depends heavily on study weighting. Two studies account for 58% of the valid anchors. After those studies are removed, the direction remains favorable to AngioFM, but the significance disappears. A per-study result followed by an equal-weight average would show whether the gain is distributed across the cohort.

The video experiment provides only modest improvements. CARDIAG segment Dice rises from 43.30 to 44.01. In-house calcification AP rises from 30.91 to 31.63. These changes are positive but small.

A simpler explanation is also possible. An added frame may show the vessel during a more favorable phase of contrast filling. That benefit can occur without learning a richer temporal representation. The present comparison therefore leaves the temporal claim unresolved.

The lesion-localization results are weaker than the structural results. Several comparisons with DINO-world are not significant. Under the initialization control, calcification AP is 28.95 for DINOv3-L-FT and 30.91 for AngioFM. For ARCADE stenosis the corresponding values are 75.55 and 76.38. Neither comparison is reported as significant.

The coronary-dominance benchmark has data-quality problems that affect interpretation. The released split contains acquisition overlap. Duplicate studies are also present, and some records have conflicting labels. The reported domain-shift AUC is therefore tied to an imperfect benchmark rather than to a clean external cohort.

External evaluation is limited in scale elsewhere. XCAD uses 26 test images, while XACV uses 34. These experiments are useful checks, but the sample sizes are too small to carry a broad robustness claim.

The multi-view centerline test is synthetic. Controlled ImageCAS projections are useful for studying geometry, but real multi-view angiography includes motion and calibration error. Contrast evolution and acquisition artifacts add further difficulty. A real multi-view test would provide substantially stronger evidence.

The task-specific expert models also provide useful context. nnU-Net obtains 42.84 Dice on ARCADE coronary segments, above AngioFM's 40.52. On CARDIAG vessel segmentation, the values are 89.05 and 88.11. The frozen representation is competitive, but these experiments demonstrate transfer quality rather than task-specific state of the art.

The “foundation model” terminology still appears broader than the evidence. Every downstream task remains within coronary angiography. A substantial part of the pretraining corpus is private. Some external evaluations contain only a few dozen images, and no clean multi-center validation is reported.

The present evidence supports a strong CAG representation model. Broader foundation-model generalization remains unproven.

Questions for authors

Could the conditioning block be tested with one component removed at a time while keeping the same initialization and training schedule?

Please report a control that keeps CCTA-derived supervision but removes physical conditioning.

Please also report the complementary control: retain physical conditioning and train without the CCTA structural objectives.

Were multiple independent AngioFM and DINO-world pretraining runs performed? If available, the spread across runs for segment parsing and centerline recovery would be useful.

For the guidewire experiment, please repeat the interval calculation using the 11 sequences as the resampling units.

For phase matching, please report one error value per study and then average across studies.

For coronary dominance, how do the numbers change after removing acquisition overlap? A second result after excluding the problematic duplicate records would also be useful.

Is a real multi-view CAG test available for the geometry experiment?

Is there a held-out cohort from another institution? A test acquired with a system absent from the private pretraining collection would also help assess transfer.

Reviewer_vForOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

AngioFM adapts a DINOv3 ViT-L/16 encoder for X-ray coronary angiography through conditional feature prediction. Pretraining combines 31,559 unlabeled angiography videos with synthetic projections from 1,000 CCTA coronary trees. A student predicts EMA-teacher features from masked frames, contrast transitions, and held-out synthetic views, conditioned on acquisition time, contrast state, and imaging rays; vessel-tangent and branch objectives add structural supervision. The authors evaluate frozen encoder features with task-specific readouts for lesions, coronary anatomy, cardiac-phase matching, vessel and device segmentation, and centerline regression. The most consistent gains are in coronary-segment parsing, while the benefits on other tasks are more mixed or qualified.

Critical strengths
  1. A well-motivated domain-specific formulation. Combining real cine variation with controlled projections of known coronary trees addresses two genuine difficulties of angiography: changing vessel visibility and overlapping anatomy. This makes the pretraining approach more consequential than straightforward adaptation of a generic image encoder.
  2. Convincing evidence for improved coronary-segment representation. With frozen encoders and matched readouts, AngioFM improves segment Dice over the CAG-pretrained DINO-world baseline on ARCADE and CARDIAG (Tables 2–3), and over its own CAG-adapted initializer (Table 8). Removing branch supervision under the same training schedule consistently reduces segment Dice (Appendix D.2). These results support a specific, useful anatomical transfer benefit.
Critical weaknesses
  1. The contribution of the proposed physical conditioning is not established cleanly. The positional-only comparison jointly removes ray coordinates, acquisition-state conditioning, and geometric attention bias, while also differing in pretraining history and budget (Section 6; Appendix D.4). Thus, the reported geometric gain cannot be attributed to the paper’s central conditioning choices. The controlled branch-loss ablation is valuable, but it does not resolve this question about the physical-conditioning mechanism.

  2. Evidence for broad transfer from the frozen representation is concentrated in one task family. Lesion gains over DINO-world and the CAG-adapted initializer are not significant; the phase advantage is sensitive to two studies; and the released dominance domain-shift split has known overlap and duplicate or conflicting records (Section 5.1; Appendices B.5 and D.1). For dense tasks, the trainable image stem and substantial decoders also leave the contribution of frozen features to absolute performance unclear. These issues matter because broad reusable capability is central to the foundation-model claim, whereas the strongest evidence concerns coronary-segment anatomy.

Questions for authors
  1. Physical conditioning What exactly differs in training history and budget between the full and positional-only models? Can the authors provide a matched-initialization, matched-data, matched-step comparison that separates ray/geometric conditioning from time and contrast conditioning?

  2. Frozen-feature contribution On the principal ARCADE and CARDIAG segment benchmarks, how does the same CNN-256 readout perform with (a) frozen features and the image stem, (b) frozen features without the image stem, and (c) the image stem without frozen features?

  3. Breadth of transfer Are private test patients and acquisitions disjoint from both private pretraining sources? These checks would clarify how much independent evidence supports transfer beyond coronary-segment parsing.

Reviewer_gQ4NOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

AngioFM starts from DINOv3 and learns coronary angiography features from private CAG videos and CCTA-derived projections. The training combines feature prediction conditioned on time, contrast state and imaging rays with branch and tangent supervision. The encoder is then frozen for different downstream readouts. The clearest reported gains are in coronary-segment parsing and centerline localization.

Critical strengths
  1. The anatomical motivation is well grounded. Coronary branches overlap in projection, and their visibility changes with contrast and motion. Using CCTA-derived structure together with real CAG videos is a meaningful way to address these difficulties. (2) The gains in coronary-segment parsing and centerline localization are useful. Improvements on ARCADE and CARDIAG, including the external CARDIAG centerline test, support transfer to real angiograms. The comparison with DINOv3-L-FT also shows that the anatomical gains go beyond the initial CAG adaptation. (3) The evaluation benefits from CAG-pretrained baselines, readouts with different capacities, and task-specific experts.The matched branch-loss ablation is particularly informative, since it gives direct evidence for the benefit of branch supervision. The authors also disclose several evaluation limitations, which helps the reader judge the results.
Critical weaknesses
  1. The experiments do not yet isolate the effect of physical conditioning. DINO-world lacks the same CCTA structural supervision, and Table 10 compares different pretraining histories and budgets. The current evidence therefore does not cleanly establish the added contribution of conditioning. (2) The practical route for independent verification is unclear. The manuscript has a reproducibility statement, but gives no concrete access plan for the relevant checkpoints and implementation. Core pretraining, including DINOv3-L-FT adaptation, depends on private data. ImageCAS is used for downstream training and testing and cannot reconstruct these earlier stages. (3) Some temporal and device gains have limited statistical robustness. The guidewire bootstrap resamples 640 frames from only 11 sequences. For phase matching, excluding two influential studies preserves the improvement direction but removes significance. I would be cautious about interpreting the smaller gains in light of these limitations.(4) The downstream video representation is insufficiently specified. The per-frame encoder and pretraining predictor are described, but their relationship to the features used by downstream video readouts is not fully clear. Predictor retention, frame fusion and the required acquisition inputs need an explicit account. (5) The geometric evaluation leaves an interpretation gap. Real single-view transfer is demonstrated, but the multi-view evaluation is synthetic. Chamfer distance mainly measures localization, while the calculation of Identity and Recall@10 px is insufficiently defined. The evidence for correct branch connectivity is therefore less direct. (6) The provenance of structural supervision needs clarification. The paper describes branch construction from existing CCTA trees, but not sufficiently how those trees were obtained, checked or corrected. The filtering thresholds for ambiguous projected patches are also missing. (7) Patient separation across stages is not explicit enough to audit. Appendix B.1 describes grouped downstream splits, but does not fully resolve pretraining exposure to downstream held-out patients or the patient-level grouping of ImageCAS case-sides.
Questions for authors

(1) Could the authors provide a conditioning comparison with the same initialization, CAG/CCTA data, structural losses, training budget and downstream readout protocol? Only the physical conditioning should change. For assessing what conditioning adds beyond structural supervision, this controlled comparison is the key one. (2) Which checkpoints, rendering and preprocessing code, split definitions and readout implementations will be made available for independent evaluation? Please distinguish the results that can be repeated using released weights from the pretraining stages that will still require private data. (3) Could the guidewire differences and confidence intervals be recomputed by resampling whole sequences? For phase matching, please report the difference and confidence interval with equal weight per study, alongside the existing exclusion analysis, and qualify the temporal claim according to these results. (4) Please give the complete downstream video feature path: whether the predictor is retained, where frame features are combined, which modules are frozen, and what is passed to the readout. Which acquisition conditions are required, and how are κ and ray coordinates obtained for real CAG? (5) Could the existing real centerline annotations be used to assess wrong-branch and missed-route errors alongside Chamfer distance? Please also define Identity and Recall@10 px precisely, including the matching rules, evaluation unit and denominator. These details would help assess anatomical recovery more directly. (6) How were the original CCTA trees and centerline labels produced and checked? Please describe whether manual correction was required, how incomplete branches were handled, and the thresholds used to exclude ambiguous projected patches from structural supervision. (7) Could the authors provide the patient-level split rules and overlap checks across stages, particularly between pretraining and downstream validation/test sets? Please confirm whether both coronary sides and all ImageCAS projections from one patient remain in the same partition.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.