Multi-View Stenosis Classification Leveraging Transformer-Based Multiple-Instance Learning Using Real-World Clinical Data

TL;DR

This work tackles coronary stenosis diagnosis from multi-view angiography by reframing it as a patient-level prediction problem amenable to multiple-instance learning. The authors introduce SegmentMIL, a transformer-based MIL model that uses a single, shared encoder across views, a decoder with learned queries, and a three-level supervision scheme to produce patient-, artery-, and segment-level predictions, along with zero-shot artery segmentation maps. Trained on real-world hospital data without view-level labels, SegmentMIL leverages temporal context by processing multiple frames per view and demonstrates strong AUC performance on internal and external datasets, outperforming view-level and MIL baselines. The approach offers clinically meaningful, interpretable predictions and localization without expensive annotations, highlighting practical potential and avenues for future enhancement with richer temporal modeling and clinical context.

Abstract

Coronary artery stenosis is a leading cause of cardiovascular disease, diagnosed by analyzing the coronary arteries from multiple angiography views. Although numerous deep-learning models have been proposed for stenosis detection from a single angiography view, their performance heavily relies on expensive view-level annotations, which are often not readily available in hospital systems. Moreover, these models fail to capture the temporal dynamics and dependencies among multiple views, which are crucial for clinical diagnosis. To address this, we propose SegmentMIL, a transformer-based multi-view multiple-instance learning framework for patient-level stenosis classification. Trained on a real-world clinical dataset, using patient-level supervision and without any view-level annotations, SegmentMIL jointly predicts the presence of stenosis and localizes the affected anatomical region, distinguishing between the right and left coronary arteries and their respective segments. SegmentMIL obtains high performance on internal and external evaluations and outperforms both view-level models and classical MIL baselines, underscoring its potential as a clinically viable and scalable solution for coronary stenosis diagnosis. Our code is available at https://github.com/NikolaCenic/mil-stenosis.

Figures (9)

• Figure 1: Overview of our approach: Multiple angiographic views from a single patient capture the coronary arteries from various angles, providing information on different segments of the coronary arteries, highlighted in green for the right (RCA) and blue for the left (LCA) coronary artery. Since a single view includes multiple segments and each segment appears in multiple views, patient-level stenosis diagnosis requires an integrated analysis of all views. To address this, we propose SegmentMIL, a multi-view transformer-based stenosis classification model capable of predicting patient-, artery-, and segment-level stenosis. Furthermore, by leveraging the transformer's attention maps, we derive zero-shot artery segmentation masks, providing interpretable visual explanations of the model’s decision process.
• Figure 2: Example frames from angiography views with the coronary arteries highlighted. The coronary artery system consists of two main branches: the right (RCA, marked in green) and the left coronary artery (LCA, marked in blue). Based on the Syntax Score Methodologysintax_score, RCA and LCA are divided into 16 segments. Segments 1, 2, 3, 4, and 16 correspond to the RCA, and the remaining belong to the LCA.
• Figure 3: Distribution of the horizontal and vertical angulations across views. The angulations correspond to the positioning of the C-arm of the X-ray device used to image the coronary angiography. The angulation clusters are highlighted by the K-means centroids (denoted by $\pmb{\times}$), showing the clinical practice, where acquisitions are performed from standardized angulations for visualizing specific coronary arteries and segments.
• Figure 4: Comparison of stenosis distribution for patients with different number of views. The internal test set (a) is selected to have a tight range of views-per-patient, in order to obtain a more representative evaluation. The CADICA test set distribution (b) hints that stenosis is more common among patients with more views. To evaluate such bias, we trained an XGBoost classifier to predict stenosis based only on the number of views and showed that this bias does not have a significant influence on evaluation performance.
• Figure 5: Comparison of the performance of our key frame detection algorithm ($x$-axis) against clinicians' annotations ($y$-axis). The absolute mean difference between the two is 3.77 frames, which corresponds to 0.53 seconds, given the frame rate of the angiography videos (7 frames per second).