Baaklini Makram, Valdés Hernández Maria Del C
Edinburgh Imaging Academy, College of Medicine and Veterinary Medicine, University of Edinburgh, Edinburgh, United Kingdom.
Department of Neuroimaging Sciences, Centre for Clinical Brain Sciences, University of Edinburgh, Edinburgh, United Kingdom.
Front Med Technol. 2025 Jun 10;7:1491197. doi: 10.3389/fmedt.2025.1491197. eCollection 2025.
Segmentation of ischaemic stroke lesions from magnetic resonance images (MRI) remains a challenging task mainly due to the confounding appearance of these lesions with other pathologies, and variations in their presentation depending on the lesion stage (i.e., hyper-acute, acute, subacute and chronic). Works on the theme have been reviewed, but none of the reviews have addressed the seminal question on what would be the optimal architecture to address this challenge. We systematically reviewed the literature (2015-2023) for deep learning algorithms that segment acute and/or subacute stroke lesions on brain MRI seeking to address this question, meta-analysed the data extracted, and evaluated the results.
Our review, registered in PROSPERO (ID: CRD42023481551), involved a systematic search from January 2015 to December 2023 in the following databases: IEE Explore, MEDLINE, ScienceDirect, Web of Science, PubMed, Springer, and OpenReview.net. We extracted sample characteristics, stroke stage, imaging protocols, and algorithms, and meta-analysed the data extracted. We assessed the risk of bias using NIH's study quality assessment tool, and finally, evaluated our results using data from the ISLES-2015-SISS dataset.
From 1485 papers, 41 were ultimately retained. 13/41 studies incorporated attention mechanisms in their architecture, and 39/41 studies used the Dice Similarity Coefficient to assess algorithm performance. The generalisability of the algorithms reviewed was generally below par. In our pilot analysis, the UResNet50 configuration, which was developed based on the most comprehensive architectural components identified from the reviewed studies, demonstrated a better segmentation performance than the attention-based AG-UResNet50.
We found no evidence that favours using attention mechanisms in deep learning architectures for acute stroke lesion segmentation on MRI data, and the use of a U-Net configuration with residual connections seems to be the most appropriate configuration for this task.
https://www.crd.york.ac.uk/PROSPERO/view/CRD42023481551, PROSPERO CRD42023481551.
从磁共振图像（MRI）中分割出缺血性中风病灶仍然是一项具有挑战性的任务，主要原因是这些病灶与其他病变的外观混淆，以及其表现会因病灶阶段（即超急性期、急性期、亚急性期和慢性期）而有所不同。关于该主题的研究已有综述，但没有一篇综述探讨过应对这一挑战的最佳架构这一关键问题。我们系统地回顾了2015年至2023年期间关于在脑MRI上分割急性和/或亚急性中风病灶的深度学习算法的文献，对提取的数据进行了荟萃分析，并评估了结果。
我们的综述已在PROSPERO（编号：CRD42023481551）注册，涉及在2015年1月至2023年12月期间在以下数据库中进行系统检索：IEEE Explore、MEDLINE、ScienceDirect、Web of Science、PubMed、Springer和OpenReview.net。我们提取了样本特征、中风阶段、成像方案和算法，并对提取的数据进行了荟萃分析。我们使用美国国立卫生研究院的研究质量评估工具评估偏倚风险，最后，使用ISLES - 2015 - SISS数据集的数据评估我们的结果。
从1485篇论文中，最终保留了41篇。41项研究中有13项在其架构中纳入了注意力机制，41项研究中有39项使用骰子相似系数来评估算法性能。所综述算法的通用性普遍低于标准水平。在我们的初步分析中，基于从综述研究中确定的最全面的架构组件开发的UResNet50配置，表现出比基于注意力的AG - UResNet50更好的分割性能。
我们没有发现证据支持在基于MRI数据的急性中风病灶分割的深度学习架构中使用注意力机制，并且使用具有残差连接的U - Net配置似乎是这项任务最合适的配置。
https://www.crd.york.ac.uk/PROSPERO/view/CRD42023481551，PROSPERO CRD42023481551。
