GRC-ProbNet 不确定感知特征提取用于心血管疾病分类

Jul 16, 2026

Upload by Asher

Automatic detection and classification of cardiovascular disease (CVD) from cardiac CT images support clinical decision-making, yet existing hybrid pipelines like GRC-Net depend on deterministic segmentation masks and ignore anatomical ambiguity. GRC-ProbNet uses deep ensembling to generate multiple segmentation masks, extracts uncertainty features from them, and evaluates how these features correlate with segmentation error and propagate to downstream CVD classification. Experiments on MM-WHS and ASOCA show improved AUROC versus GRC-Net.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages10

Reading Time25 min

Abstract

Introduction

Method

arXiv :2607 . 10357v1 [ cs .CV] 11 Jul 2026 GRC-ProbNet: Uncertainty-aware Feature Extraction for Cardiovascular Disease Classification Yash Shah 1 , Omar Todd 1 , Philipp Seeböck2 , Georg Langs2 , Ben Glocker 1 , and Raghav Mehta\#, 1 1 Imperial College London, UK 2 Medical University of Vienna, Austria \# Email: [raghav.mehta@imperial.ac.uk](raghav.mehta@imperial.ac.uk) Abstract The automatic detection and classification of cardiovascular disease (CVD) from computed tomography (CT) images plays an important role in clinical practice. Recently, a hybrid pipeline (GRC-Net) for CVD classification was proposed, which leverages a deep-learning-based segmentation and registration method to extract radiomic and geometric features. However, GRC-Net relies on a deterministic segmentation mask, without considering the inherent ambiguity associated with cardiac anatomy. In this paper, we propose GRC-ProbNet, which takes advantage of a deep ensemble to produce multiple segmentation masks for a given input. From these masks, we extract multiple uncertainty features. We analyze these uncertainty features for both their correlation with segmentation error and their propagation effects on downstream CVD classification performance. Our experiments on the publicly available MM-WHS and ASOCA datasets show that the uncertainty measure that best reflects segmentation quality is not necessarily the one that provides the strongest signal for downstream CVD classification. Overall, our results demonstrate that GRC-ProbNet utilizing uncertainty features substantially improves CVD classification AUROC (92.92%) compared to the baseline GRC-Net model (91.25%). Our code is publicly available: [https://github.com/biomedia-mira/](https://github.com/biomedia-mira/)[ ](https://github.com/biomedia-mira/)GRC-ProbNet. Keywords: Cardiovascular Disease · Uncertainty Quantification · Cardiac CT · Foundation Models. 1 Introduction Cardiovascular disease (CVD) remains one of the leading causes of mortality worldwide, accounting for an estimated 20.5 million deaths in 2025, with cardiovascular mortality projected to rise by 73.4% between 2025 and 2050 as populations age [1] . Cardiac computed tomography (CT) has become a central non-invasive imaging modality for the diagnosis of CVD. However, analysing these scans manually is time-consuming, requiring significant clinical expertise [2] . These challenges motivate the development of automated, segmentation-based pipelines that can deliver fast, consistent, and reproducible analysis at scale. Deep learning is the leading approach for analysing cardiac images. It has been extensively applied to the automated detection and quantification of coronary artery disease in CT angiography, demonstrating strong agreement with expert radiologist annotations [6 , 3 , 7] . In addition to improving accuracy, these methods often reduce reporting times and support decision-making in time-critical settings such as acute cardiac care [4] . A key limitation of existing cardiac image analysis pipelines is their inability to accurately represent the inherent ambiguity observed in cardiac anatomy. Their deterministic outputs do not align with the inter-rater variation that would be present across multiple expert annotations. This limitation reduces their efficacy for meaningful uncertainty propagation and analysis in downstream stages. Recently, many methods have been proposed that allow approximating the uncertainty associated with model outputs; these include approximate Bayesian methods, such as Monte Carlo Dropout [11], Deep Ensembles [12], and Stochastic Segmentation networks [13] . Each of these produces multiple plausible predictions, and their variability quantifies model uncertainties. Recent works report that propagating these uncertainties in cascaded inference tasks leads to improvement in downstream tasks [5 , 14 , 9 , 10] . However, most of these existing approaches propagate uncertainty directly b

FAQs

GRC-ProbNet 相比 GRC-Net 解决了什么核心问题？

GRC-Net 使用确定性的分割掩膜，无法反映心脏解剖中的固有不确定性；GRC-ProbNet 通过深度集成生成多种分割结果，并据此提取不确定性特征。

GRC-ProbNet 提取的不确定性特征包括哪些类型？

从多张分割掩膜中提取多种不确定性表征，文中提到包含 predictive entropy、KL divergence 以及 aleatoric maps。

不确定性特征与分割误差的相关性是否等同于分类性能提升？

不一定。结果表明与分割误差最相关的不确定性度量不一定带来最大的下游分类提升；predictive entropy 更贴近分割误差，而 KL divergence 更一致地提升分类表现。

Jul 16, 2026

Upload by Asher