Advances in intracoronary imaging have made it possible to distinguish different pathological mechanisms underlying acute coronary syndrome (ACS) in vivo. Accurate identification of these mechanisms is increasingly recognized as essential for enabling tailored therapeutic strategies. ACS pathogenesis is primarily classified into 2 major types: plaque rupture (PR) and plaque erosion (PE). Patients with PR are treated with intracoronary stenting, whereas those with PE may be potentially managed conservatively without stenting. The aim of this study is to develop neural networks capable of distinguishing PR from PE solely using coronary angiography (CAG). A total of 842 videos from 278 ACS patients (PR:172, PE:106) were included. To ensure the reliability of the ground truth for PR/PE classification, the ACS pathogenesis for each patient was confirmed using Optical Coherence Tomography (OCT). To enhance the learning of discriminative features across consecutive frames and improve PR/PE classification performance, we propose Sequence Contrastive Learning (SeqCon), which addresses the limitations inherent in conventional contrastive learning approaches. In the experiments, the external test set consisted of 18 PR patients (46 videos) and 11 PE patients (30 videos). SeqCon achieved an accuracy of 82.8%, sensitivity of 88.9%, specificity of 72.3%, positive predictive value of 84.2%, and negative predictive value of 80.0% at the patient-level. This is the first report to use contrastive learning for diagnosing the underlying mechanism of ACS by CAG. We demonstrated that it can be feasible to distinguish between PR and PE without intracoronary imaging modalities.
Acute coronary syndrome (ACS) is a leading cause of morbidity and mortality worldwide. Plaque rupture (PR) is the most frequent finding in autopsy studies, but plaque erosion (PE) is also an important cause of ACS. , High-resolution intracoronary imaging, particularly optical coherence tomography (OCT), has shown that PE is more common than previously recognized and differs from PR in pathobiology, prognosis, and clinical profiles. ,,,, These differences suggest the potential for mechanism-dependent tailored therapy, yet most ACS patients are uniformly treated with intracoronary stenting. Although conservative strategies may be feasible for selected PE patients, diagnosis currently requires OCT, an invasive modality that is not routinely available. Coronary angiography (CAG), the most widely performed test for ACS, provides only 2-dimensional views of the lumen and is generally considered inadequate for identifying underlying pathology. Nonetheless, certain angiographic features of PE have been reported, suggesting that pathology-related signals may exist but remain underutilized. While deep learning has been applied to CAG for stenosis segmentation and detection, , no prior study has attempted to classify ACS pathogenesis (PR vs PE) directly from CAG. To address this gap, we propose Sequence Contrastive Learning (SeqCon), a representation learning framework that leverages within-sequence information in angiographic videos for PR/PE classification. An overview of the proposed method is presented in the Central Illustration.
Central Illustration. Development of the proposed Sequence Contrastive Learning, and Evaluation at frame/video/patient level for PR and PE Classification. We proposed a sequence contrastive (SeqCon) learning method, which could capture discriminative features across consecutive frames. The learned representations are then utilized in downstream tasks, leading to improved performance in PR/PE classification.
Methods
Study population
Patients with ACS—either non–ST-segment elevation ACS (NSTE-ACS) or ST-segment elevation ACS (STE-ACS) , —who underwent coronary angiography for a clearly identifiable culprit lesion and had interpretable OCT images were enrolled in this study. All patients were admitted to Nippon Medical School Chiba Hokusoh Hospital before December 2023. The culprit lesion was defined as the site of percutaneous coronary intervention (PCI), the most severe stenotic lesion, or the lesion with angiographic evidence of recent plaque disruption, and all culprit lesions were confirmed by OCT.
The study protocol was approved by the medical ethics committee of Nippon Medical School Chiba Hokusoh Hospital (No. 898), and informed consent was obtained using an opt-out method on the hospital website. The study was conducted in accordance with the ethical principles of the Declaration of Helsinki.
OCT image acquisition and analysis
OCT images were acquired using a frequency-domain system (Abbott Vascular, Santa Clara, CA, USA) after recanalization. OCT recording was recommended; in cases of a low Thrombolysis in Myocardial Infarction (TIMI) grade, thrombus aspiration or gentle predilatation using a balloon ≤2.0 mm was permitted to obtain prompt recanalization if necessary. OCT images were analyzed using offline software at Nippon Medical School Chiba Hokusoh Hospital. An experienced interventional cardiologist assessed culprit-plaque morphology. PR was identified by the presence of fibrous cap discontinuity with communication between the lumen and the inner core of plaque or with a cavity formation within the plaque. PE was identified by the presence of an attached thrombus overlying an intact plaque or luminal surface irregularity at the culprit lesion.
Coronary angiography acquisition and analysis
The CAG data were obtained from patients with ACS at Nippon Medical School Chiba Hokusoh Hospital. The images were acquired using the Canon INFX-8000C and Canon INFX-8000V systems, both equipped with digital flat-panel detectors. The original images were acquired at either 15 or 10 frames per second (fps) as 1024 × 1024-pixel grayscale images with a 16-bit depth (ranging from 0 to 65,535 grayscale levels). These images were subsequently transferred online to a network server. On the network server, the images were processed and converted to the standard DICOM format as 512 × 512-pixel grayscale images with an 8-bit depth (ranging from 0 to 255 grayscale levels). The data were then stored in a RAID (Redundant Array of Independent Disks) system for secure long-term preservation. For analysis, the videos were anonymized and converted into the widely used MP4 (MPEG-4 Part 14) format for dynamic image processing and were submitted to the Kochi University of Technology for off-line analysis. A series of 1 to 3 right coronary and 1 to 5 left coronary angiographic acquisitions per patient was obtained at varying angles prior to guidewire crossing of the culprit lesion. In addition, CAG was performed again after thrombus aspiration or gentle predilatation with a ≤ 2.0 mm balloon to facilitate OCT imaging. Among these acquisitions, those in which the culprit lesion was clearly visualized were selected for subsequent analysis. The number of selected videos from each angiographic direction is summarized in Supplementary Figure 1 .
Statistical analysis
All statistical analyses were performed using SPSS software (version 11.0.1; SPSS Inc.). Categorical variables are presented as frequencies and compared using the Pearson chi-squared test. Continuous quantitative data are presented as mean ± standard deviation and evaluated using an unpaired Student’s t-test or 1-way analysis of variance, as appropriate. All differences were evaluated at a 95% level of significance (p < 0.05).
Development of the deep learning model
Model structure
Our proposed deep learning-based method incorporates 2 distinct learning stages: (1) sequence contrastive learning and (2) classification learning. Each learning stage is illustrated in Figure 1 and explained in the following parts.
The framework of our proposed method. It consists of 2 stages: sequence contrastive learning and classification learning. And we utilize ResNet-50 as backbone in both 2 stages. In the first stage, we extend the conventional contrastive loss function which optimizes the sequence contrastive learning, leading to a well-trained feature representation. In the second stage, load well-trained parameters to enhance the PR/PE classification learning. The neural network in classification learning will output predicted score for PR and PE using Softmax. The ACS type (PR/PE) classification for each image is determined by the maximum value between the predicted PR and PE scores. ACS =acute coronary syndromes; PR = plaque rupture; PE = plaque erosion.
Our method is a modification of the conventional contrastive learning SimCLR, using the same neural network architecture, ResNet-50. For network development, an experienced cardiologist manually annotated the lesion area using rectangular bounding boxes on the original X-ray images. As shown in Figure 2 , the red rectangle indicates the cardiologist’s annotation. Based on these annotations, we first crop the lesion area from the original 512 × 512 images. The cropped regions are then resized to 64 × 64 pixels, while maintaining the aspect ratio. These preprocessed images are used to train the neural network in both stages—sequence-based contrastive learning and classification learning—while validation and test sets are used only in the classification stage.
Sequence contrastive pairs. (A) 2 distinct sequence frames are represented, and those stenosis areas with red rectangles are cropped from the original frames. (B) Conventional contrastive pairs: Employing the way of positive and negative pairs by conventional methods. (C) Sequence contrastive pairs: We redefine the way of positive and negative pairs for ACS X-ray sequence. ACS = acute coronary syndromes.
Sequence contrastive learning (SeqCon)
Contrastive learning enhances neural network representation by mapping positive pairs to similar representations while pushing negative pairs apart. In conventional approaches, , each sample forms only 1 positive pair—2 augmentations of the same image—against a set of negatives comprising all other samples. For sequential data such as CAG videos ( Figure 2 ), this design creates an issue: consecutive frames within the same sequence, which share the same underlying pathobiology, are incorrectly treated as negatives ( Figure 2 ). Our objective is not to distinguish between individual frames from the same video, but to identify the patient’s underlying pathobiology—PR or PE. Since all frames in a video correspond to the same pathobiology, we propose SeqCon to redefine positive and negative relationships based on video-level information. Specifically, all frames within a sequence are considered positive, while frames from different sequences are negative ( Figure 2 ). During training, 2 random augmentations from the same sequence (Seq1′ and Seq1″) are encouraged to have similar representations, whereas augmentations from other sequences (e.g., Seq2) are pushed apart. To effectively capture discriminative representations from sequence-based contrastive pairs, we extend the conventional contrastive loss (see Supplementary Equation 1 , and details in the Supplementary Appendix ). This design minimizes feature distance within a sequence and maximizes separation across sequences, promoting representations that capture pathobiology rather than frame-level variation.
Classification learning
The well-trained parameters obtained through sequence contrastive learning can provide enhanced visual representations, thereby improving PR/PE classification. Therefore, we load the well-trained parameters to the same backbone in classification learning, as illustrated in Figure 1 . We also employed a weighted cross-entropy loss (see supplementary Equation 3 ) to address the class imbalance issue, as detailed in the Supplementary Appendix .
Results
Patient characteristics
For the model development, we used a total of 842 videos from 278 ACS patients (PR:172, PE:106). Detailed patient characteristics are summarized in Table 1 . To avoid any potential data leakage, patients were randomly divided into nonoverlapping subsets as shown in Table 2 , training (673 videos from 220 patients), internal validation (93 videos from 29 patients) and external test datasets (76 videos from 29 patients). It is noted that the consecutive frames from multiple videos share the same ACS type-PR/PE-for each patient. In the test, we evaluated the PR/PE classification performance at 3 levels: frame-level, video-level, and patient-level. The prevalence of male patients was lower in the external test datasets compared to the training and internal validation datasets. The level of high-density lipoprotein cholesterol was higher in the external test datasets than in the other 2 subsets. Regarding angiographic findings, the distributions of culprit vessels, Thrombolysis in Myocardial Infarction (TIMI) flow grades, and quantitative coronary angiography (QCA) parameters showed no significant differences among the 3 datasets.
Table 1
Clinical baseline characteristics
| Patients, n | Training set | Internal validation set | External test set | p |
|---|---|---|---|---|
| 220 (79.2%) | 29 (10.4%) | 29 (10.4%) | ||
| Age, years | 66.1 ± 12.2 | 64.6 ± 10.9 | 68.7 ± 12.0 | 0.397 |
| Sex, male | 185 (84%) | 24 (83%) | 17 (59%) | 0.004 |
| Hypertension, n | 151 (69%) | 19 (66%) | 20 (69%) | 0.941 |
| Hyperlipidemia, n | 124 (56%) | 18 (62%) | 15 (52%) | 0.727 |
| Diabetes mellitus | 78 (35%) | 9 (31%) | 9 (31%) | 0.820 |
| Current smoking, n | 117 (53%) | 12 (41%) | 9 (31%) | 0.052 |
| Prior MI | 7 (3%) | 0 (0%) | 1 (3%) | 0.617 |
| STE-ACS | 161 (73%) | 24 (83%) | 24 (83%) | 0.324 |
| Laboratory data | ||||
| WBC (/lL) | 9571±3141 | 9751±4135 | 9854±3218 | 0.884 |
| Hb (g/dL) | 14.3 ± 1.7 | 14.1 ± 1.6 | 13.5 ± 1.9 | 0.053 |
| T-cho (mg/dL) | 201.4 ± 38.7 | 198.7 ± 35.0 | 203.8 ± 50.7 | 0.884 |
| LDL-C (mg/dL) | 129.6 ± 34.7 | 123.0 ± 25.8 | 127.1 ± 40.3 | 0.608 |
| HDL-C (mg/dL) | 46.8 ± 10.8 | 48.9 ± 10.5 | 52.5 ± 14.1 | 0.029 |
| TG (mg/dL) | 162.2 ± 134.4 | 181.4 ± 158.8 | 124.8 ± 80.9 | 0.242 |
| CRP (mg/dL) | 0.61±2.06 | 0.88±2.53 | 2.5 ± 11.5 | 0.071 |
| HbA1c (%) | 6.5 ± 1.2 | 6.3 ± 1.4 | 6.5 ± 1.6 | 0.837 |
| eGFR (mL/min/1.73 m 2) | 70.4 ± 19.5 | 71.7 ± 22.0 | 66.3 ± 20.8 | 0.527 |
| Hs-TnT (ng/mL) | 0.659±1.659 | 0.730±1.442 | 0.390±0.806 | 0.651 |
| Peak CK (U/L) | 2322.4 ± 2182.0 | 1877.3 ± 1746.3 | 2726.5 ± 2375.5 | 0.328 |
| Peak CKMB (U/L) | 210.8 ± 194.1 | 169.7 ± 157.6 | 252.1 ± 228.3 | 0.274 |
| Angiographic findings | ||||
| Culprit vessel | 0.614 | |||
| LAD | 104 (47%) | 16 (55%) | 17 (59%) | |
| LCx | 36 (17%) | 4 (14%) | 2 (7%) | |
| RCA | 80 (36%) | 9 (31%) | 10 (34%) | |
| Culprit vessel pre-PCI TIMI flow | 0.989 | |||
| 0 | 102 (46%) | 13 (45%) | 15 (52%) | |
| 1 | 41 (19%) | 6 (21%) | 5 (17%) | |
| 2 | 47 (21%) | 5 (17%) | 5 (17%) | |
| 3 | 30 (14%) | 5 (17%) | 4 (14%) | |
| QCA profiles | ||||
| Minimum lumen diameter, mm | 0.33±0.35 | 0.28±0.32 | 0.26±0.30 | 0.452 |
| Reference diameter, mm | 2.57±0.61 | 2.53±0.60 | 2.56±0.62 | 0.953 |
| Lesion length, mm | 11.9 ± 7.1 | 9.8 ± 5.3 | 10.9 ± 5.7 | 0.451 |
| Diameter stenosis, % | 86.9 ± 14.3 | 84.5 ± 21.0 | 89.0 ± 12.4 | 0.528 |
| AHA lesion types | 0.335 | |||
| B1 | 24 (11%) | 4 (14%) | 4 (14%) | |
| B2 | 179 (81%) | 24 (83%) | 20 (69%) | |
| C | 17 (8%) | 1 (3%) | 5 (17%) | |
| Bifurcation | 45 (20%) | 7 (24%) | 6 (21%) | 0.900 |
| Calcified lesion | 36 (16%) | 6 (21%) | 6 (21%) | 0.740 |
Stay updated, free articles. Join our Telegram channel
Full access? Get Clinical Tree
