Rougé Pierre, Passat Nicolas, Merveille Odyssée
CReSTIC EA 3804, Université de Reims Champagne Ardenne, Reims, France.
CNRS, Inserm, CREATIS UMR 5220, U1294 INSA-Lyon, Université Claude Bernard Lyon 1, UJM-Saint Etienne, Univ Lyon, Lyon, France.
PLoS One. 2024 Dec 5;19(12):e0311439. doi: 10.1371/journal.pone.0311439. eCollection 2024.
Cerebrovascular segmentation is a crucial preliminary task for many computer-aided diagnosis tools dealing with cerebrovascular pathologies. Over the last years, deep learning based methods have been widely applied to this task. However, classic deep learning approaches struggle to capture the complex geometry and specific topology of cerebrovascular networks, which is of the utmost importance in many applications. To overcome these limitations, the clDice loss, a topological loss that focuses on the vessel centerlines, has been recently proposed. This loss requires computing the skeletons of both the manual annotation and the predicted segmentation in a differentiable way. Currently, differentiable skeletonization algorithms are either inaccurate or computationally demanding. In this article, it is proposed that a U-Net be used to compute the vascular skeleton directly from the segmentation and the magnetic resonance angiography image. This method is naturally differentiable and provides a good trade-off between accuracy and computation time. The resulting cascaded multitask U-Net is trained with the clDice loss to embed topological constraints during the segmentation. In addition to this topological guidance, this cascaded U-Net also benefits from the inductive bias generated by the skeletonization during the multitask training. This model is able to predict the cerebrovascular segmentation with a more accurate topology than current state-of-the-art methods and with a low training time. This method is evaluated on two publicly available time-of-flight magnetic resonance angiography (TOF-MRA) images datasets, also the codes of the proposed method and the reimplementation of state-of-the-art methods are made available at: https://github.com/PierreRouge/Cascaded-U-Net-for-vessel-segmentation.
脑血管分割是许多处理脑血管疾病的计算机辅助诊断工具的关键预处理任务。在过去几年中，基于深度学习的方法已被广泛应用于该任务。然而，经典的深度学习方法难以捕捉脑血管网络的复杂几何形状和特定拓扑结构，而这在许多应用中至关重要。为了克服这些限制，最近提出了clDice损失，一种专注于血管中心线的拓扑损失。这种损失需要以可微的方式计算手动标注和预测分割的骨架。目前，可微骨架化算法要么不准确，要么计算量很大。在本文中，提出使用U-Net直接从分割和磁共振血管造影图像计算血管骨架。该方法自然可微，并且在准确性和计算时间之间提供了良好的权衡。由此产生的级联多任务U-Net使用clDice损失进行训练，以便在分割过程中嵌入拓扑约束。除了这种拓扑指导之外，这种级联U-Net在多任务训练期间还受益于骨架化产生的归纳偏差。该模型能够以比当前最先进方法更准确的拓扑结构预测脑血管分割，并且训练时间较短。该方法在两个公开可用的飞行时间磁共振血管造影（TOF-MRA）图像数据集上进行了评估，同时所提出方法的代码以及最先进方法的重新实现可在以下网址获得：https://github.com/PierreRouge/Cascaded-U-Net-for-vessel-segmentation 。
