PaperlayerView exact revision on arXiv

Source-linked AI summary

Self-training for cross-centre coronary vessel segmentation in X-ray angiography using pre-contrast frame subtraction

Danial Moafi, Aida Moafi, Mostafa Aryafar, Ashkan Sami, Maryam Aryafar

arXiv:2610.40302v1cs.CVcs.AI

TL;DR

Coronary segmentation can be difficult across centres, and public ARCADE labels omit vessels outside SYNTAX-defined segments. This study combines nnU-Net predictions with pre-contrast subtraction for self-training without new annotations. One round improved performance on two external datasets, with gains limited by the initial model’s sensitivity.

  • Problem

    Cross-centre coronary segmentation is difficult to generalize, while ARCADE labels omit side branches and distal vessels outside the SYNTAX score.

    Source: §1 · p. 1

  • Method

    The method combines nnU-Net predictions and same-sequence pre-contrast subtraction as pseudo-label sources for self-training without new annotations.

    Source: §1 · p. 2, §2 · p. 2, §2.2 · p. 3

  • Results

    2.7 points: the XCAD-pool model’s macro Dice gain over baseline (95% CI 2.3–3.2), while first-round models outperformed baseline and control on all three test sets.

    Source: §3 · p. 4

  • Takeaways & Limitations

    One self-training round improved performance on two external datasets without new annotations, but gains were limited by the initial model’s sensitivity.

    Source: §5 · p. 6

    The subtraction method requires a suitable pre-contrast frame in the same sequence, which may not always be available in clinical practice.

    Source: §4 · p. 5

Abstract

from arXiv · show

Deep learning models for coronary vessel segmentation in X-ray angiography often perform poorly on data from centres that were not represented in training. Apart from differences in equipment, the labels of the largest public dataset (ARCADE) are incomplete, as only the segments defined in the SYNTAX score are annotated. In this study, we evaluated whether self-training with pseudo-labels can improve performance on external data without additional manual annotation. Pseudo-labels were generated from two sources, namely the predictions of an nnU-Net model trained on ARCADE and a contrast map obtained by subtracting a pre-contrast frame from the same sequence. The nnU-Net predictions were extended along the contrast map, and pixels detected only in the subtraction image were assigned an ignore label and excluded from the loss. Six models were trained and tested on ARCADE and on two external datasets (XCAD and XACV). All first-round models initialised with the baseline predictions performed better than both the baseline and a control model fine-tuned for the same number of epochs. The model trained on the XCAD pool improved the macro-averaged Dice score by 2.7 points (95% CI 2.3-3.2), while the control decreased it by 0.6 points. A second round of self-training did not lead to further improvement. When an external model trained on 50 fully annotated images was used to generate the initial labels, the XCAD Dice score increased from 0.623 to 0.745 without any XCAD labels, although performance on XACV did not improve because large main vessels were missed. These results suggest that the benefit of self-training depends largely on the sensitivity of the model used to generate the initial pseudo-labels.

1. Introduction

Coronary segmentation models can struggle across centres, while ARCADE’s labels omit side branches and distal vessels outside the SYNTAX score. This study evaluates self-training that combines model predictions with pre-contrast frame subtraction to improve external-data segmentation without new annotations.

1. Introduction

  • External evaluation is challenging because models often perform worse across centres, and complete annotations from additional centres are costly to collect.ARCADE labels only SYNTAX-defined segments, leaving side branches and distal vessels labelled as background.Source: §1 · p. 1
  • Self-training reuses a trained model’s predictions as labels, but can reinforce its mistakes through confirmation bias.An independent second label source is one proposed way to reduce this effect.Source: §1 · p. 1
  • Pre-contrast frames from the same sequence offer another information source, and their subtraction had not been used as pseudo-labels in this self-training framework.Prior work used vesselness filters for noisy labels or learned single-frame subtraction for pretraining.Source: §1 · p. 2
  • The study evaluates pseudo-label self-training from nnU-Net predictions and pre-contrast subtraction on external data without new annotations.It also examines a second training round and the choice of initial model.Source: §1 · p. 2

2. Method

The method combines ARCADE-trained nnU-Net predictions with a registered, same-sequence pre-contrast subtraction map to create pseudo-labels, then fine-tunes on those labels and ARCADE data. Experiments compare six models across ARCADE, XCAD, and XACV test sets.

2. Method

  • Each round combines two pseudo-label sources into vessel, background, and ignore classes, then fine-tunes on pseudo-labelled frames alongside labelled ARCADE data.No manual annotations are added, and ignore pixels are excluded from the training loss.Source: §2 · p. 2, §2.3 · p. 3

2.1 Pseudo-labels from the segmentation model

  • The first source is a 2D nnU-Net trained on 1,000 ARCADE frames for 311 epochs, using its best validation checkpoint.Training stopped before the default 1,000-epoch schedule.Source: §2.1 · p. 2

2.2 Pseudo-labels from frame subtraction

  • Registered pre-contrast frames are subtracted from contrast-filled frames, and region growing extends the nnU-Net mask along the resulting contrast map.Added pixels must exceed a contrast threshold and meet vesselness or proximity criteria; width and pool-exclusion rules limit false positives.Source: §2.2 · pp. 2–3

2.3 Label fusion and training

  • Model-predicted pixels are labelled vessel, region-growing additions are ignored, and all remaining pixels are labelled background.The ignore class prevents likely missed vessels from being treated as background; ignore pixels are omitted from the loss.Source: §2.3 · p. 3

2.4 Datasets and experiments

  • Six models were compared using ARCADE labels, XCAD and XACV unlabelled training pools, and test sets from all three datasets.The comparisons included baseline and ARCADE-only control models, first-round XCAD and XACV models, a second XCAD round, and an XCAD model initialized with DeepSA labels; each model was trained once.Source: §2.4 · pp. 3–4, Table 1 · p. 3
  • Dice and clDice were averaged per test set, with macro Dice defined as the mean across the three test sets.Paired percentile-bootstrap confidence intervals against baseline used 2,000 resamples over test images.Source: §2.4 · p. 4, Table 2 · p. 4, Table 3 · p. 4

3. Results

First-round self-training improved results over both the baseline and the equally fine-tuned control on all three test sets. The XCAD-pool model had the largest gain, while a second XCAD round and the XACV-pool model did not improve on it.

3. Results

  • Macro Dice rose by 2.7 points (95% CI 2.3–3.2) with the XCAD-pool model, the largest gain; both first-round models beat the baseline and control on all three test sets.The control reduced macro Dice by 0.6 points (95% CI 0.1–1.1).Source: §3 · p. 4
  • The second XCAD round and the XACV-pool model did not improve on the first XCAD model, performing similarly or up to 1.3 points worse across test sets.Source: §3 · p. 4
  • With DeepSA initial labels, XCAD Dice reached 0.745 (+12.2 points, 95% CI 10.5–13.8) without XCAD labels, exceeding DeepSA’s own XCAD test-set Dice of 0.69.Source: §3 · p. 4
  • On XACV, the DeepSA-initialized model had the lowest self-trained Dice, 0.546 versus 0.541 for the baseline, and missed large main vessels.Figure 2 marks true positives, false negatives, and false positives in green, blue, and orange, respectively.Source: §3 · p. 5, Figure 2 · p. 5

4. Discussion

The authors report that combining pseudo-label sources improved external-data segmentation, while longer ARCADE-only training did not. Gains were constrained by initial-label sensitivity and incomplete vessel coverage; the study also has limits in training-run replication, frame availability, and clinical evaluation.

4. Discussion

  • Combining pseudo-label sources with different error types improved external segmentation, whereas longer ARCADE-only training did not.Excluding uncertain pixels from the loss let training use regions where sources agreed without treating likely errors as labels.Source: §4 · p. 5
  • On XACV, baseline predictions covered 3.9% of image pixels versus 8.9% in ground truth; incomplete ARCADE labels and full external annotations make missed branches count as errors.The authors say this coverage gap probably also explains why another round did not help: region growing added only 1.7% of XACV pixels, leaving much of the vessel tree as background.Source: §4 · p. 5
  • DeepSA-initialized self-training reached XCAD Dice 0.745, above published no-XCAD-label results of 0.47–0.55, but DeepSA’s cross-validated 0.755 is not directly comparable.The authors link weaker XACV performance probably to DeepSA top-hat preprocessing suppressing wide dark structures, an error inherited by the self-trained model.Source: §4 · p. 5
  • Each model was trained once, so confidence intervals capture test-image variability but not variability across training runs.The same-sequence pre-contrast frame needed for subtraction may not always be available clinically; the method was not evaluated on clinical data or downstream QCA measurements.Source: §4 · p. 5
  • Future work plans to fully annotate about 30 frames from a new clinical centre to obtain a better initial model for that setting.The authors also plan to explore subtraction-based or vessel-aware masked-image-modeling self-supervised pretraining.Source: §4 · p. 5

5. Conclusion

The study concludes that same-sequence pre-contrast frames can provide a second pseudo-label source for coronary vessel self-training. One round improved performance on two external datasets without new annotations, but gains were limited by initial-model sensitivity.

5. Conclusion

  • One round of self-training using a same-sequence pre-contrast frame as a second pseudo-label source improved performance on two external datasets without new annotations, with gains limited by initial-model sensitivity.Source: §5 · p. 6

Ethics statement

The study used publicly available, de-identified datasets and required no additional ethical approval.

Ethics statement

International Conference on Learning Representations (ICLR), 2023.

These passages comprise bibliography entries rather than reported findings, naming prior work on self-training, vessel segmentation, and coronary-angiography datasets.

Loading 2610.40302v1…