[Paper Review] CNN-CASS: CNN for Classification of Coronary Artery Stenosis Score in MPR Images

Mariia Dobko, Bohdan Petryshak+1 more|arXiv (Cornell University)|Jan 23, 2020

Cardiac Imaging and Diagnostics4 citations

TL;DR

This paper proposes CNN-CASS, a fully automated deep learning framework based on ShuffleNet V2 for classifying coronary artery stenosis score in curved Multiplanar Reconstruction (MPR) images from CCTA scans. It achieves 80% patient-level accuracy without centerline extraction or handcrafted features, using a newly collected dataset of 708 patients for training and 105 for testing, with model interpretability demonstrated via gradient-based visualization.

ABSTRACT

To decrease patient waiting time for diagnosis of the Coronary Artery Disease, automatic methods are applied to identify its severity using Coronary Computed Tomography Angiography scans or extracted Multiplanar Reconstruction (MPR) images, giving doctors a second-opinion on the priority of each case. The main disadvantage of previous studies is the lack of large set of data that could guarantee their reliability. Another limitation is the usage of handcrafted features requiring manual preprocessing, such as centerline extraction. We overcome both limitations by applying a different automated approach based on ShuffleNet V2 network architecture and testing it on the proposed collected dataset of MPR images, which is bigger than any other used in this field before. We also omit centerline extraction step and train and test our model using whole curved MPR images of 708 and 105 patients, respectively. The model predicts one of three classes: 'no stenosis' for normal, 'non-significant' - 1-50% of stenosis detected, 'significant' - more than 50% of stenosis. We demonstrate model's interpretability through visualization of the most important features selected by the network. For stenosis score classification, the method shows improved performance comparing to previous works, achieving 80% accuracy on the patient level. Our code is publicly available.

Motivation & Objective

• To reduce patient waiting time for coronary artery disease diagnosis by automating stenosis score classification in CCTA scans.
• To overcome limitations of prior work by using a larger, publicly available dataset of 800+ patients, surpassing previous data scales in this domain.
• To eliminate reliance on manual preprocessing steps such as centerline extraction and handcrafted features, enabling full automation.
• To improve model interpretability through visualization of salient features using Integrated Gradients, supporting clinical trust and decision validation.
• To demonstrate state-of-the-art performance in stenosis classification using whole curved MPR images without anatomical alignment or segmentation.

Proposed method

• The method employs a pre-trained ShuffleNet V2 architecture fine-tuned on a custom dataset of 708 patient MPR images for training and 105 for testing.
• Input data consists of whole curved MPR images extracted from CCTA scans, without any preprocessing such as centerline extraction or vessel alignment.
• The model classifies stenosis into three classes: 'no stenosis', 'non-significant' (1–50% stenosis), and 'significant' (>50% stenosis).
• Model interpretability is achieved using Integrated Gradients to visualize the most influential regions in the input image for each prediction.
• For final classification per coronary artery, 50 predictions per MPR image are aggregated, and the majority class is assigned.
• The framework avoids 3D CNNs and segmentation by relying on end-to-end learning from 2D MPR images, reducing complexity and computational cost.

Experimental results

Research questions

• RQ1 Can a deep learning model achieve high accuracy in classifying coronary artery stenosis severity without centerline extraction or handcrafted features?
• RQ2 Does a larger, publicly available dataset of MPR images improve the reliability and performance of stenosis classification models compared to prior small-scale studies?
• RQ3 Can model interpretability through gradient-based saliency maps enhance clinical trust and assist radiologists in evaluating borderline cases?
• RQ4 How does the performance of a 2D CNN on curved MPR images compare to previous methods relying on preprocessed or aligned vessel data?
• RQ5 What are the main sources of model confusion, and can they be mitigated through architectural or preprocessing improvements?

Key findings

• The proposed CNN-CASS model achieves 80% accuracy on the patient-level classification of stenosis score, outperforming previous methods in this task.
• The method eliminates the need for centerline extraction and handcrafted features, enabling full automation of the classification pipeline.
• Visualization using Integrated Gradients reveals that the model focuses on anatomically relevant regions, such as the vessel lumen and plaque areas, enhancing interpretability.
• Despite high accuracy, the model occasionally confuses calcified plaques with pathological stenosis, particularly in regions like the sternum and ribs, due to similar Hounsfield unit intensities.
• The model's performance is robust across different MPR views, with consistent attention to the artery lumen and stenotic regions, even in complex anatomical configurations.
• The authors release a new dataset of over 800 patient MPR images, which is larger than any previously used in this domain, supporting future research.