DeepCORO-CLIP: coronary stenosis checkpoint

Video model for invasive coronary angiography. From the cine runs of one study it estimates, for each of 18 coronary segments, the percent diameter stenosis and the probability of significant stenosis, calcification, chronic total occlusion (CTO) and thrombus.

This repository was previously named heartwise/deepcoro_clip_stenosis; the old name redirects here.

Files

File Use
zcb8cu0l_20260213-033432/best_model_epoch_11.pt Current checkpoint (the one config.json and the GitHub code point to). sha256 4eac7676bc24930466d84f04f813fcbc1b9e70b056232cad603f18c84d008267
config.json Model configuration: mViT video encoder, 16 frames at 224 x 224, up to 10 cine runs per study, attention pooling over runs
thresholds.json Decision thresholds for the current checkpoint (this page)
cnw09vn8_01062025-140448/, threshold_70/ Earlier checkpoints. The thresholds on this page do not apply to them

How to run

Use the GitHub repository: the README sections "Hugging Face inference weights" and "Docker Setup" describe the full pipeline (DICOM to video, run filtering with VasoVision, then this model). Head and label configuration: config/linear_probing/stenosis/docker_base_config.yaml.

python -m utils.download_pretrained_weights --models stenosis

Outputs

90 values per study: 5 outputs for each of 18 segments. Prediction columns are named <output>_pred.

Output Type Meaning
<segment>_stenosis regression Estimated percent diameter stenosis (0 to 100 scale of the angiography report; not clipped)
<segment>_stenosis_binary probability Probability of a significant stenosis
<segment>_calcif_binary probability Probability of moderate or severe calcification
<segment>_cto probability Probability of a chronic total occlusion
<segment>_thrombus probability Probability of thrombus

Segments: prox_rca, mid_rca, dist_rca, pda, posterolateral, left_main, prox_lad, mid_lad, dist_lad, D1, D2, prox_lcx, mid_lcx, dist_lcx, om1, om2, bx (ramus), lvp (posterior left ventricular branch).

Label definitions

Labels come from the operator's structured angiography report (visual assessment), one value per segment and study.

  • Significant stenosis: report stenosis of 70% or more; for the left main, 50% or more.
  • Calcification: moderate or severe calcification, or a calcified nodule. None or mild is negative.
  • CTO and thrombus: the finding is recorded for that segment in the report.

Decision thresholds

A finding is called positive when the probability is strictly greater than the threshold of that output (thresholds.json).

The probabilities are not calibrated and the findings are uncommon, so the thresholds are low. Two cut-offs that look natural do not work:

  • Do not use 0.5 on the probability: mean sensitivity 0.18 (specificity 0.99) on the internal test set.
  • Do not use 70 on the regression output: the percent estimate is compressed toward low values, and a cut-off of 70 (50 for the left main) gives a mean sensitivity of 0.08. Use the regression output as a continuous estimate.

The thresholds are those of the HeartWise deployment of this checkpoint: the point of the ROC curve with the highest geometric mean of sensitivity and specificity, on internal Montreal Heart Institute data. They are balanced operating points; sensitivity is favoured and the positive predictive value (PPV) is low.

Significant stenosis (internal held-out test set, 3,194 studies)

Segment Output column Label cut-off Threshold Positives / 3,194 AUROC Sensitivity Specificity PPV
Proximal RCA prox_rca_stenosis_binary_pred ≥ 70% 0.1079 462 0.91 0.84 0.83 0.46
Mid RCA mid_rca_stenosis_binary_pred ≥ 70% 0.1284 431 0.92 0.84 0.83 0.43
Distal RCA dist_rca_stenosis_binary_pred ≥ 70% 0.0327 215 0.90 0.92 0.71 0.18
Posterior descending artery pda_stenosis_binary_pred ≥ 70% 0.0299 146 0.82 0.84 0.62 0.10
Posterolateral branch posterolateral_stenosis_binary_pred ≥ 70% 0.0209 107 0.82 0.80 0.72 0.09
Left main left_main_stenosis_binary_pred ≥ 50% 0.0546 271 0.93 0.86 0.84 0.33
Proximal LAD prox_lad_stenosis_binary_pred ≥ 70% 0.1145 489 0.90 0.84 0.78 0.41
Mid LAD mid_lad_stenosis_binary_pred ≥ 70% 0.1754 589 0.88 0.79 0.79 0.46
Distal LAD dist_lad_stenosis_binary_pred ≥ 70% 0.0373 102 0.77 0.65 0.76 0.08
First diagonal D1_stenosis_binary_pred ≥ 70% 0.1308 326 0.77 0.69 0.69 0.20
Second diagonal D2_stenosis_binary_pred ≥ 70% 0.0568 198 0.77 0.72 0.67 0.13
Proximal LCx prox_lcx_stenosis_binary_pred ≥ 70% 0.0830 351 0.89 0.85 0.76 0.31
Mid LCx mid_lcx_stenosis_binary_pred ≥ 70% 0.0597 219 0.84 0.83 0.73 0.19
Distal LCx dist_lcx_stenosis_binary_pred ≥ 70% 0.0105 49 0.79 0.78 0.61 0.03
First obtuse marginal om1_stenosis_binary_pred ≥ 70% 0.0723 182 0.79 0.64 0.75 0.13
Second obtuse marginal om2_stenosis_binary_pred ≥ 70% 0.0767 201 0.83 0.81 0.70 0.15
Ramus bx_stenosis_binary_pred ≥ 70% 0.0299 105 0.80 0.83 0.66 0.08
Posterior left ventricular branch (LVP) lvp_stenosis_binary_pred ≥ 70% 0.0014 17 0.93 1.00 0.72 0.02
Mean of 18 0.85 0.81 0.73 0.21

Calcification, CTO and thrombus (exploratory)

Positives are few, so the estimates are imprecise. Performance is shown only for outputs with at least 10 positive studies in the test set. "Not derived" means there were too few positive cases to set a threshold; the 0.5 placeholder almost never fires and the output should be treated as not supported.

Calcification

Segment Output column Threshold Positives / 3,194 AUROC Sensitivity Specificity
Proximal RCA prox_rca_calcif_binary_pred 0.0376 166 0.85 0.88 0.72
Mid RCA mid_rca_calcif_binary_pred 0.0302 99 0.88 0.90 0.73
Distal RCA dist_rca_calcif_binary_pred 0.0171 43 0.92 0.88 0.83
Posterior descending artery pda_calcif_binary_pred 0.0066 15 0.89 0.80 0.80
Posterolateral branch posterolateral_calcif_binary_pred 0.0012 10 0.85 1.00 0.62
Left main left_main_calcif_binary_pred 0.0230 116 0.88 0.83 0.77
Proximal LAD prox_lad_calcif_binary_pred 0.0729 214 0.80 0.69 0.77
Mid LAD mid_lad_calcif_binary_pred 0.0894 168 0.83 0.72 0.78
Distal LAD dist_lad_calcif_binary_pred 0.0026 10 0.84 0.80 0.72
First diagonal D1_calcif_binary_pred 0.0206 41 0.81 0.78 0.76
Second diagonal D2_calcif_binary_pred 0.0065 30 0.87 0.97 0.64
Proximal LCx prox_lcx_calcif_binary_pred 0.0207 103 0.87 0.86 0.74
Mid LCx mid_lcx_calcif_binary_pred 0.0062 26 0.85 0.92 0.69
Distal LCx dist_lcx_calcif_binary_pred 0.0050 5 n/a n/a n/a
First obtuse marginal om1_calcif_binary_pred 0.0104 21 0.88 0.81 0.79
Second obtuse marginal om2_calcif_binary_pred 0.0091 21 0.88 0.90 0.70
Ramus bx_calcif_binary_pred 0.0029 20 0.82 0.85 0.61
Posterior left ventricular branch (LVP) lvp_calcif_binary_pred not derived (0.5 placeholder) 2 n/a n/a n/a

Chronic total occlusion

Segment Output column Threshold Positives / 3,194 AUROC Sensitivity Specificity
Proximal RCA prox_rca_cto_pred 0.0134 89 0.96 0.93 0.86
Mid RCA mid_rca_cto_pred 0.0225 57 0.95 0.91 0.85
Distal RCA dist_rca_cto_pred 0.0043 18 0.94 1.00 0.80
Posterior descending artery pda_cto_pred 0.0031 12 0.82 0.67 0.75
Posterolateral branch posterolateral_cto_pred 0.0034 13 0.89 0.77 0.81
Left main left_main_cto_pred not derived (0.5 placeholder) 3 n/a n/a n/a
Proximal LAD prox_lad_cto_pred 0.0043 22 0.94 0.95 0.78
Mid LAD mid_lad_cto_pred 0.0159 50 0.93 0.90 0.82
Distal LAD dist_lad_cto_pred 0.0017 9 n/a n/a n/a
First diagonal D1_cto_pred 0.0030 14 0.83 0.93 0.56
Second diagonal D2_cto_pred 0.0009 7 n/a n/a n/a
Proximal LCx prox_lcx_cto_pred 0.0083 34 0.92 0.91 0.78
Mid LCx mid_lcx_cto_pred 0.0112 22 0.86 0.73 0.83
Distal LCx dist_lcx_cto_pred 0.0011 4 n/a n/a n/a
First obtuse marginal om1_cto_pred 0.0019 9 n/a n/a n/a
Second obtuse marginal om2_cto_pred 0.0086 12 0.76 0.67 0.81
Ramus bx_cto_pred 0.0012 3 n/a n/a n/a
Posterior left ventricular branch (LVP) lvp_cto_pred not derived (0.5 placeholder) 4 n/a n/a n/a

Thrombus

Segment Output column Threshold Positives / 3,194 AUROC Sensitivity Specificity
Proximal RCA prox_rca_thrombus_pred 0.0109 18 0.92 0.89 0.84
Mid RCA mid_rca_thrombus_pred 0.0126 16 0.95 0.81 0.89
Distal RCA dist_rca_thrombus_pred 0.0058 10 0.96 1.00 0.81
Posterior descending artery pda_thrombus_pred not derived (0.5 placeholder) 1 n/a n/a n/a
Posterolateral branch posterolateral_thrombus_pred 0.0017 3 n/a n/a n/a
Left main left_main_thrombus_pred not derived (0.5 placeholder) 4 n/a n/a n/a
Proximal LAD prox_lad_thrombus_pred 0.0086 15 0.92 0.80 0.86
Mid LAD mid_lad_thrombus_pred 0.0038 21 0.93 0.90 0.73
Distal LAD dist_lad_thrombus_pred not derived (0.5 placeholder) 2 n/a n/a n/a
First diagonal D1_thrombus_pred not derived (0.5 placeholder) 1 n/a n/a n/a
Second diagonal D2_thrombus_pred not derived (0.5 placeholder) 1 n/a n/a n/a
Proximal LCx prox_lcx_thrombus_pred 0.0003 3 n/a n/a n/a
Mid LCx mid_lcx_thrombus_pred 0.0028 5 n/a n/a n/a
Distal LCx dist_lcx_thrombus_pred not derived (0.5 placeholder) 0 n/a n/a n/a
First obtuse marginal om1_thrombus_pred not derived (0.5 placeholder) 1 n/a n/a n/a
Second obtuse marginal om2_thrombus_pred not derived (0.5 placeholder) 2 n/a n/a n/a
Ramus bx_thrombus_pred not derived (0.5 placeholder) 1 n/a n/a n/a
Posterior left ventricular branch (LVP) lvp_thrombus_pred not derived (0.5 placeholder) 0 n/a n/a n/a

Limitations

  • Single centre (Montreal Heart Institute). The reference is the operator's visual report, not quantitative coronary angiography or a core laboratory.
  • The split used to derive the thresholds was not recorded with them. On the test set they lie close to the test-optimal points, so the sensitivity and specificity above may be slightly optimistic.
  • The test-set figures used up to 12 diagnostic runs per study without view filtering; the Docker pipeline filters runs with VasoVision and uses up to 10. Expect small differences.
  • Small and distal segments (distal LAD, diagonals, obtuse marginals) are detected less well than proximal segments.
  • Thresholds depend on prevalence, acquisition and case mix. At another site, re-derive or check them on a local labelled calibration set that is separate from the evaluation set, and report threshold-free metrics (AUROC, AUPRC, correlation with percent stenosis) next to the thresholded ones.

Citation

DeepCORO-CLIP, arXiv:2603.17675. https://arxiv.org/abs/2603.17675

Downloads last month
69
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Collection including heartwise/deepcoro_clip_generic

Paper for heartwise/deepcoro_clip_generic