Tsuchida Ami, Goubet Martin, Boutinaud Philippe, Astafeva Iana, Nozais Victor, Hervé Pierre-Yves, Tourdias Thomas, Debette Stéphanie, Joliot Marc
GIN, IMN-UMR5293, CEA, CNRS, Université de Bordeaux, Bordeaux, France.
BPH-U1219, INSERM, Université de Bordeaux, Bordeaux, France.
Sci Rep. 2024 Dec 28;14(1):30901. doi: 10.1038/s41598-024-81870-5.
Cerebral microbleeds (CMB) represent a feature of cerebral small vessel disease (cSVD), a prominent vascular contributor to age-related cognitive decline, dementia, and stroke. They are visible as spherical hypointense signals on T2*- or susceptibility-weighted magnetic resonance imaging (MRI) sequences. An increasing number of automated CMB detection methods being proposed are based on supervised deep learning (DL). Yet, the lack of open sharing of pre-trained models hampers the practical application and evaluation of these methods beyond specific data sources used in each study. Here, we present the SHIVA-CMB detector, a 3D Unet-based tool trained on 450 scans taken from seven acquisitions in six different cohort studies that included both T2*- and susceptibility-weighted MRI. In a held-out test set of 96 scans, it achieved the sensitivity, precision, and F1 (or Dice similarity coefficient) score of 0.67, 0.82, and 0.74, with less than one false positive detection per image (FPavg = 0.6) and per CMB (FPcmb = 0.15). It achieved a similar level of performance in a separate, evaluation-only dataset with acquisitions never seen during the training (0.67, 0.91, 0.77, 0.5, 0.07 for the sensitivity, precision, F1 score, FPavg, and FPcmb). Further demonstrating its generalizability, it showed a high correlation (Pearson's R = 0.89, p < 0.0001) with a visual count by expert raters in another independent set of 1992 T2*-weighted scans from a large, multi-center cohort study. Importantly, we publicly share both the pipeline ( https://github.com/pboutinaud/SHiVAi/ ) and pre-trained models ( https://github.com/pboutinaud/SHIVA-CMB/ ) to the research community to promote the active application and evaluation of our tool. We believe this effort will help accelerate research on the pathophysiology and functional consequences of CMB by enabling rapid characterization of CMB in large-scale studies.
脑微出血（CMB）是脑小血管疾病（cSVD）的一个特征，cSVD是导致与年龄相关的认知衰退、痴呆和中风的一个主要血管因素。在T2加权或敏感性加权磁共振成像（MRI）序列上，它们表现为球形低信号。越来越多基于监督深度学习（DL）的自动CMB检测方法被提出。然而，缺乏预训练模型的公开共享阻碍了这些方法在每项研究中使用的特定数据源之外的实际应用和评估。在此，我们展示了SHIVA-CMB检测器，这是一种基于3D U-Net的工具，它在来自六项不同队列研究的七次采集的450次扫描上进行训练，这些扫描包括T2加权和敏感性加权MRI。在一个包含96次扫描的保留测试集中，它的灵敏度、精度和F1（或骰子相似系数）得分分别为0.67、0.82和0.74，每张图像的假阳性检测少于1次（平均假阳性率FPavg = 0.6），每个CMB的假阳性检测为0.15（FPcmb）。在一个单独的、仅用于评估的数据集上，该数据集的采集在训练期间从未见过，它实现了相似的性能水平（灵敏度、精度、F1得分、FPavg和FPcmb分别为0.67、0.91、0.77、0.5、0.07）。在来自一项大型多中心队列研究的另一组独立的1992次T2*加权扫描中，它与专家评分者的视觉计数显示出高度相关性（皮尔逊相关系数R = 0.89，p < 0.0001），进一步证明了其通用性。重要的是，我们向研究社区公开共享了管道（https://github.com/pboutinaud/SHiVAi/ ）和预训练模型（https://github.com/pboutinaud/SHIVA-CMB/ ），以促进对我们工具的积极应用和评估。我们相信，这项工作将有助于通过在大规模研究中快速表征CMB来加速对CMB病理生理学和功能后果的研究。
