Alskaf Ebraham, Scannell Cian M, Suinesiaputra Avan, Crawley Richard, Masci PierGiorgio, Young Alistair, Perera Divaka, Chiribiri Amedeo
School of Biomedical Engineering & Imaging Sciences, King's College London, London, UK.
Eindhoven University of Technology, Eindhoven, The Netherlands.
J Med Artif Intell. 2025 Mar;8:2. doi: 10.21037/jmai-24-94.
The prognostic value of late gadolinium enhancement (LGE) in cardiac magnetic resonance (CMR) imaging is well-established. However, the direct relationship between image pixels and outcomes remains poorly understood. We hypothesised that leveraging artificial intelligence (AI) to analyse qualitative LGE images based on American Heart Association (AHA) guidelines could elucidate this relationship.
We collected retrospective CMR cases from a stress perfusion database, selecting LGE images comprising three long-axis views and 10 short-axis views. Clinical CMR reports served for annotation. We trained a multi-label convolutional neural network (CNN) to predict each AHA segment. Additionally, we transformed LGE image pixels into features, combined them with clinical data features, and trained a hybrid neural network (HNN) to predict mortality and ventricular arrhythmia. The dataset was divided into training (70%), validation (15%), and test (15%) sets. Evaluation metrics included the area under the curve (AUC).
The total number of cases included was 2,740, with 218 patients experiencing positive mortality events (8%). The total number of cases with at least one AHA segment positive for LGE was 823 (30%), among which 111 (13%) experienced mortality events, and 84 (10%) had ventricular arrhythmia events. When assessing all segments combined, the most common cases were those classified as normal studies, with each AHA segment having a score of 0 (1,661 cases, 60.6%). The multi-label classifier demonstrated fair performance (AUC: 64%), whereas the cluster classifier did not yield any predictions (AUC: 53%, P<0.001). The mortality HNN achieved a satisfactory performance with an AUC of 77%, as did the ventricular arrhythmia HNN with an AUC of 75%.
Our study demonstrates the feasibility of generating qualitative AHA LGE maps using AI. Furthermore, the prediction of mortality and ventricular arrhythmia using HNN represents a potent new approach for risk stratification in patients with known or suspected coronary artery disease (CAD).
心脏磁共振成像（CMR）中延迟钆增强（LGE）的预后价值已得到充分证实。然而，图像像素与预后之间的直接关系仍知之甚少。我们假设利用人工智能（AI）基于美国心脏协会（AHA）指南分析定性LGE图像可以阐明这种关系。
我们从一个应力灌注数据库中收集回顾性CMR病例，选择包含三个长轴视图和10个短轴视图的LGE图像。临床CMR报告用于注释。我们训练了一个多标签卷积神经网络（CNN）来预测每个AHA节段。此外，我们将LGE图像像素转换为特征，将其与临床数据特征相结合，并训练了一个混合神经网络（HNN）来预测死亡率和室性心律失常。数据集分为训练集（70%）、验证集（15%）和测试集（15%）。评估指标包括曲线下面积（AUC）。
纳入的病例总数为2740例，其中218例患者发生阳性死亡事件（8%）。至少有一个AHA节段LGE呈阳性的病例总数为823例（30%），其中111例（13%）发生死亡事件，84例（10%）发生室性心律失常事件。在评估所有节段综合情况时，最常见的病例是分类为正常检查的病例，每个AHA节段的评分为0（1661例，60.6%）。多标签分类器表现出一般性能（AUC：64%），而聚类分类器未产生任何预测结果（AUC：53%，P<0.001）。死亡率HNN的AUC为77%，表现令人满意，室性心律失常HNN的AUC为75%，表现同样令人满意。
我们的研究证明了使用AI生成定性AHA LGE图谱的可行性。此外，使用HNN预测死亡率和室性心律失常是已知或疑似冠状动脉疾病（CAD）患者风险分层的一种有效新方法。
