Peng Fei, Xia Jiaxiang, Zhang Fandong, Lu Shiyu, Wang Hao, Li Jiashu, Liu Xinmin, Zhong Yao, Guo Jiahuan, Duan Yonghong, Sui Binbin, Ye Chuyang, Ju Yi, Kang Shuai, Yu Yizhou, Feng Xin, Zhao Xingquan, Li Rui, Liu Aihua
Beijing Neurosurgical Institute and Beijing Tiantan Hospital, Capital Medical University, Beijing, China.
Deepwise Artificial Intelligence (AI) Lab, Deepwise Inc., Beijing, China.
Neurotherapeutics. 2025 Jan;22(1):e00505. doi: 10.1016/j.neurot.2024.e00505. Epub 2024 Nov 30.
This study aims to develop a reliable predictive model for assessing intracranial aneurysm (IA) instability by utilizing four-dimensional flow magnetic resonance imaging (4D-Flow MRI) and high-resolution MRI (HR-MRI). Initially, we curated a prospective dataset, dubbed the primary cohort, by aggregating patient data that was consecutively enrolled across two centers from November 2018 to November 2021. Unstable aneurysms were defined as those with symptoms, morphological change or ruptured during follow-up periods. We introduce a specialized ensemble learning framework, termed the Hybrid Model, which synergistically combines two heterogeneous base learning algorithms: 4D-Flow logistic regression (4D-Flow-LR) and Multi-crop Attention Branch Network (MicroAB-Net). The ability of the hybrid model to predict aneurysm instability was compared with baseline models: PHASES (population, hypertension, age, size, earlier rupture, and site) LR, ELAPSS (earlier subarachnoid hemorrhage, location, age, population, size, and shape) LR, aneurysm wall enhancement (AWE) LR, and Radiomics using the area under the curve (AUC) with Delong's test. Finally, the Hybrid Model was further validated in the validation cohort (patients enrolled between December 2021 to May 2022). In the primary cohort, 189 patients (144 women [76.2 %]; aged 58.90 years ± 10.32) with 213 IAs were included. In the validation cohort, 48 patients (35 women [72.9 %]; aged 55.0 years ± 10.77) with 53 IAs were included. The Hybrid Model achieved the highest performance both in the primary cohort (AUC = 0.854) and the validation cohort (AUC = 0.876). The Hybrid model provided a promising prediction of aneurysm instability.
本研究旨在通过利用四维流动磁共振成像（4D-Flow MRI）和高分辨率MRI（HR-MRI）开发一种可靠的预测模型，以评估颅内动脉瘤（IA）的不稳定性。最初，我们通过汇总2018年11月至2021年11月期间在两个中心连续登记的患者数据，策划了一个前瞻性数据集，称为主要队列。不稳定动脉瘤被定义为在随访期间出现症状、形态改变或破裂的动脉瘤。我们引入了一种专门的集成学习框架，称为混合模型，它协同结合了两种异构的基础学习算法：4D流动逻辑回归（4D-Flow-LR）和多裁剪注意力分支网络（MicroAB-Net）。将混合模型预测动脉瘤不稳定性的能力与基线模型进行比较：PHASES（人群、高血压、年龄、大小、早期破裂和部位）LR、ELAPSS（早期蛛网膜下腔出血、位置、年龄、人群、大小和形状）LR、动脉瘤壁强化（AWE）LR，以及使用曲线下面积（AUC）和德龙检验的放射组学模型。最后，在验证队列（2021年12月至2022年5月期间登记的患者）中对混合模型进行了进一步验证。在主要队列中，纳入了189例患者（144名女性[76.2%]；年龄58.90岁±10.32），共213个IA。在验证队列中，纳入了48例患者（35名女性[72.9%]；年龄55.0岁±10.77），共53个IA。混合模型在主要队列（AUC = 0.854）和验证队列（AUC = 0.876）中均取得了最高性能。混合模型为动脉瘤不稳定性提供了有前景的预测。
