Cross-Architectural Feature Fusion for Coronary Artery Stenosis Detection in X-ray Angiography
Authors: Haobo Zhu, Mojtaba Lashgari, Jason Chai, Giovanni Luigi De Maria, Vicente Grau, Abhirup Banerjee
Organizations: Institute of Biomedical Engineering, Department of Engineering Science, University of Oxford, Oxford, United Kingdom · Division of Cardiovascular Medicine, Radcliffe Department of Medicine, University of Oxford, Oxford, United Kingdom · Oxford Heart Centre, Oxford University Hospitals NHS Trust, Oxford, United Kingdom
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.
Figures & tables
Dataset
ARCADE*
OxAMI
Split
Train
Validation
Test
Test
# Image/ # Stenosis
1000/998
200/376
300/412
48/85
Table 1: Dataset summary.
Figure 1: A) Overall model architecture; B) Channel attention gating (CAG) module for cross-architecture feature fusion between the UNet++ decoder (UNet++ branch) and the YOLOv11-L encoder (YOLO branch).
Hyperparameter
#Epochs
LR
BBLG
CLG
Search Space
{50, 100}
{2e-4, 1e-4, 5e-5}
{10, 7.5, 5}
{0.5, 0.3}
Table 2: Hyperparameter search space.
Optimal Hyperparameters from Optuna
Model
LR
#Epochs
BBLG
CLG
AP@50
YOLOv8-L
1e-4
100
7.5
0.5
0.579
YOLOv11-L
1e-4
100
7.5
0.3
0.574
Fusion-YOLOv11-L (Ours)
1e-4
100
5
0.3
0.585
Table 3: Comparison of AP@50 for different models on ARCADE* test set.
Confidence
Precision
Recall
F1-score
0.250
0.665
0.529
0.589
0.200
0.630
0.566
0.596
0.150
0.564
0.592
0.578
0.100
0.480
0.621
0.542
0.050
0.383
0.682
0.490
Table 4: Evaluation of the performance of Fusion-YOLO-v11 across confidence thresholds on the ARCADE* test set.
Model
Base
TTAd Only
TTAg Only
TTAd + TTAg
YOLOv11-L
0.339
0.392
0.383
0.413
Fusion-YOLOv11-L
0.434
0.432
0.480
0.470
Table 5: Generalisability of YOLOv11-L and Fusion-YOLOv11-L across inference variants on the OxAMI dataset (AP@50).
Original
TTAd Only
TTAg Only
TTAd + TTAg
Confidence
P
R
F1
P
R
F1
P
R
F1
P
R
F1
0.250
0.625
0.235
0.342
0.658
0.294
0.407
0.638
0.353
0.455
0.612
0.353
0.448
0.200
0.650
0.306
0.416
0.628
0.318
0.422
0.620
0.365
0.459
0.561
0.376
0.451
0.150
0.622
0.329
0.431
0.588
0.353
0.441
0.574
0.412
0.479
0.530
0.412
0.464
0.100
0.586
0.400
0.476
0.585
0.447
0.507
0.551
0.506
0.528
0.500
0.541
0.520
0.050
0.460
0.471
0.465
0.467
0.506
0.486
0.409
0.553
0.470
0.384
0.565
0.457
Table 6: Generalisability of detection performance across confidence thresholds and inference variants on the OxAMI dataset.
Figure 2: Comparison of ground truth (green) and predictions (red) on example XCA images from ARCADE* dataset (in-domain test dataset).
Figure 3: Comparison of ground truth (green) and predictions (red) on example XCA images from OxAMI dataset (OOD test) using Fusion-YOLOv11-L with base model without TTAd or TTAg, model with TTAd only, model with TTAg only, and model with TTAd + TTAg.
Coronary angiography uses medical imaging to visualize heart vessels and assess disease, supporting automated diagnosis, measurement, and procedural guidance.
Computer Aided Medical Procedures, Technical University of Munich, Munich, Germany · relAI – Konrad Zuse School of Excellence in Reliable AI, Munich, Germany · German Heart Center Munich, Munich, Germany +3
LTS4 Laboratory, EPFL, Lausanne, Switzerland · Chair of Mathematical Data Science, EPFL, Lausanne, Switzerland · School of AIAC, Xi’an Jiaotong-Liverpool University, Suzhou, China +3