PaperlayerView exact revision on arXiv

Source-linked AI summary

Cross-Architectural Feature Fusion for Coronary Artery Stenosis Detection in X-ray Angiography

Haobo Zhu, Mojtaba Lashgari, Jason Chai, Giovanni Luigi De Maria, Vicente Grau, Abhirup Banerjee

arXiv:2610.32730v2cs.CV

TL;DR

Manual XCA interpretation is time-consuming and operator-dependent, and explicit use of segmentation features to guide stenosis detection remains underexplored. The paper fuses pretrained UNet++ features with YOLOv11-L using channel attention. It reports improved ARCADE* detection and higher OxAMI AP@50 with test-time augmentation, while performance remains affected by domain shift.

  • Problem

    Manual XCA interpretation is time-consuming and operator-dependent, while explicit use of segmentation-derived features to guide detection networks remains underinvestigated.

    Source: §1 · p. 2

  • Method

    The framework fuses pretrained UNet++ decoder features with a YOLOv11-L backbone through a channel attention gating module.

    Source: §1 · p. 2, §2.2 · pp. 4–5

  • Results

    ARCADE* feature fusion achieves 0.585 AP@50 versus 0.574 for YOLOv11-L, and OxAMI test-time augmentation raises AP@50 from 0.434 to 0.480.

    Source: §4 · p. 8, §5 · p. 10

  • Takeaways & Limitations

    Feature fusion improves stenosis detection on ARCADE*, while test-time augmentation improves robustness on the evaluated OxAMI data.

    Source: §4 · p. 8, §5 · p. 10

    Performance degrades under domain shift, and confidence-threshold sensitivity leaves operating-point trade-offs that may require clinical-use-specific optimisation.

    Source: §4 · p. 10

Abstract

from arXiv · show

Coronary artery disease remains a leading cause of mortality, with stenosis detection from invasive X-ray Coronary Angiography (XCA) critical for diagnosis. Manual interpretation is time-consuming, subjective, and variable, especially under low contrast and vessel overlap. Existing deep learning methods often rely solely on detection or segmentation, limiting performance. We propose a hybrid framework, Fusion-YOLOv11-L, combining a YOLOv11-L detector with a pretrained UNet++ via a channel attention gating module. This enables fusion of semantic and vessel-aware structural features, improving localisation of stenotic regions. The UNet++ is pretrained on ARCADE vessel segmentation dataset to provide anatomically informed representations. Experiments on the relabeled ARCADE (ARCADE*) dataset show improved performance over YOLOv11-L, achieving AP@50 of 0.585 and F1-score of 0.596. Cross-domain evaluation on a dataset from the Oxford Acute Myocardial Infarction (OxAMI) Study reveals performance degradation due to domain shift, mitigated using test-time adaptation and augmentation. Test-time augmentation improves AP@50 from 0.434 to 0.480, indicating better generalisability. These results show that cross-architecture feature fusion and lightweight adaptation improve robustness and generalisability of automated stenosis detection in XCA.

1 Introduction

Coronary stenosis can reduce myocardial perfusion, and XCA interpretation is time-consuming and operator-dependent. The paper addresses limited use of segmentation-derived features to guide detection by proposing a hybrid fusion framework.

1 Introduction

  • Coronary artery disease is linked to plaque-driven stenosis and reduced myocardial perfusion, underscoring the clinical stakes of stenosis detection.The paper describes CAD as a leading cause of mortality worldwide, accounting for nine million deaths annually.Source: §1 · pp. 1–2
  • XCA is the gold standard for visualising coronary arteries, but manual review is time-consuming and operator-dependent, risking missed lesions or inconsistent interpretation.Source: §1 · p. 2
  • Prior work explores varied detection and segmentation strategies, while explicit use of segmentation-derived features to guide detection networks remains underinvestigated.The proposed framework fuses pretrained UNet++ features with YOLOv11-L through a Channel Attention Gating module.Source: §1 · p. 2

2 Methods

The methods use relabelled ARCADE* images for stenosis detection and OxAMI images for out-of-domain evaluation. A pretrained UNet++ decoder is fused with YOLOv11-L through channel attention, with hyperparameter search and test-time adaptation and augmentation used in the experiments.

2.1 Dataset

  • The ARCADE* dataset contains 1,500 XCA images from 1,500 patients, split into 1,000 training, 200 validation, and 300 test images; OxAMI supplies 48 images from 27 patients for out-of-domain testing.ARCADE* stenoses with over 50% diameter narrowing were reannotated; OxAMI images depict lesions with over 50% narrowing and are captured at end-diastole.Source: §2.1 · p. 3

2.2 Model Architecture

  • A pretrained UNet++ decoder is fused with YOLOv11-L backbone features through channel-attention gating at shared 1/8 and 1/16 scales, with gated features combined residually.The decoder is pretrained on the Syntax dataset of ARCADE, which provides pixel-level coronary artery branch annotations; a learnable scalar controls decoder-feature contribution at each scale.Source: §2.2 · pp. 3, 5

2.3 Experimental Setup

  • Hyperparameter search uses Optuna to test 36 combinations, selecting the checkpoint with the highest validation AP@50 for test-set evaluation.The search tunes epochs, learning rate, bounding-box loss gain, and class loss gain; standard YOLO augmentations are applied.Source: §2.3 · p. 5, Table 2 · p. 6

2.4 Test-Time Adaptation and Augmentation for Cross-Domain Evaluation

  • Cross-domain evaluation tests the ARCADE*-trained Fusion-YOLOv11-L checkpoint on OxAMI with baseline inference, AdaBN adaptation, horizontal-flip augmentation, and their combination.AdaBN recalibrates target-domain batch statistics, which are then frozen for inference; augmentation aggregates detections from original and horizontally flipped images.Source: §2.4 · p. 6

3 Results

Quantitative evaluation covers ARCADE* performance and OxAMI cross-domain generalisability, including confidence-threshold trade-offs. Qualitative comparisons show ground truth and predictions on in-domain and cross-domain images.

3.1 Quantitative Results

  • AP@50 is the primary metric, summarising performance across confidence thresholds, and Table 3 compares YOLOv8-L, YOLOv11-L, and Fusion-YOLOv11-L on ARCADE*.The table uses each model's optimum hyperparameters.Source: §3.1 · p. 6, Table 3 · p. 6
  • On ARCADE*, confidence threshold 0.2 gave the best F1-score (0.596), while lowering the threshold improved recall and worsened precision.Table 4 reports precision, recall, and F1-score across confidence thresholds.Source: §3.1 · p. 7, Table 4 · p. 7
  • TTAg alone achieved the best OxAMI AP@50, 0.480, in the cross-domain evaluation.Table 5 compares YOLOv11-L and Fusion-YOLOv11-L across inference variants.Source: §3.1 · p. 7, Table 5 · p. 7
  • On OxAMI, the best F1-score occurred at confidence threshold 0.1, unlike the ARCADE* best threshold of 0.2.Table 6 compares confidence thresholds and inference variants.Source: §3.1 · p. 7, Table 6 · p. 8

3.2 Qualitative Results

  • Figure 2 compares ARCADE* ground-truth annotations with Fusion-YOLOv11-L predictions; Figure 3 shows OxAMI predictions across baseline and adaptation or augmentation variants.The figures overlay ground truth in green and predictions in red on example images.Source: §3.2 · pp. 7–8, Figure 2 · p. 8

4 Discussion

Feature fusion improves in-domain AP@50 over YOLOv11-L, while cross-domain evaluation shows lower performance and benefits from test-time augmentation. Confidence-threshold analysis reveals a precision–recall trade-off, and the authors identify threshold sensitivity and domain shift as remaining limitations.

4 Discussion

  • 0.585 AP@50 for Fusion-YOLOv11-L exceeded the YOLOv11-L baseline’s 0.574 AP@50.YOLOv8-L reached 0.579 AP@50, but YOLOv11-L was selected for computational efficiency and improved small-object detection.Source: §4 · pp. 8–9
  • 0.596 F1-score was achieved at a confidence threshold of 0.2, while lower thresholds increased recall at the expense of precision.The authors describe this as a balanced operating point and note that clinical applications may prioritize sensitivity.Source: §4 · p. 9
  • On OxAMI, TTAd with AdaBN did not boost AP@50, while TTAg improved it; the best baseline result (0.413 AP@50) remained below Fusion-YOLOv11-L (0.480 AP@50).TTAg also improved recall and F1-score across confidence-threshold operating points under distribution shift.Source: §4 · pp. 9–10
  • Fig. 3 qualitative results show improved localisation with TTAg and feature fusion, including for small or low-contrast stenotic regions.The figure compares ground truth and predictions across baseline, TTAd, TTAg, and combined settings.Source: Figure 3 · p. 9, §4 · p. 10
  • The model remains sensitive to confidence-threshold selection, and performance degrades under domain shift, leaving a need for more robust generalisation strategies.The authors also note that precision–recall trade-offs may require further optimization for particular clinical use cases.Source: §4 · p. 10

5 Conclusion

The conclusion presents a hybrid YOLOv11-L and pretrained UNet++ framework for stenosis detection, reporting improved in-domain detection and better cross-domain robustness with TTAg. It identifies temporal information and larger, more diverse datasets as future work.

5 Conclusion

  • 0.585 AP@50 was achieved by the best feature-fusion configuration on ARCADE*.The conclusion also reports a best F1-score of 0.596 at a confidence threshold of 0.2.Source: §5 · p. 10
  • TTAg increased OxAMI AP@50 from 0.434 to 0.480 without retraining, amid performance degradation associated with domain shift.The framework fuses YOLOv11-L features with vessel-aware representations from a pretrained UNet++ decoder through a channel attention gating module.Source: §5 · p. 10
  • Future work will explore temporal information and larger, more diverse datasets.Source: §5 · p. 10

Related papers

Loading 2610.32730v2…