病理性不对称引导的渐进学习用于急性缺血性脑卒中梗死分割| 文献速递-先进深度学习疾病诊断

Title

题目

Pathological Asymmetry-Guided, based on the principles of progressive learning, is employed to achieve accurate segmentation of the acute ischemic stroke infarct region.

病理性不对称引导的渐进学习用于急性缺血性脑卒中梗死分割

01

文献速递介绍

中风已成为继心血管后最主要的致命性疾病，在全球范围内每年造成数百万人死亡。其中约70%的中风病例为缺血性脑卒中（AIS）。计算机断层扫描（CT）与磁共振成像（MRI）是评估梗死范围的主要手段之一。尽管MRI成像虽然精确度高但操作复杂且成本高昂，在患者入院时C T扫描因其操作简便和价格亲民通常被选用作为初步筛查工具。然而手动评估虽然能在一定程度上提供参考信息但效率较低且准确性有待提高因此在临床实践中亟需一种高效可靠的评估方法来辅助C T图像分析进而精确量化AIS型梗死区域。

在非对称CT中,AIS早期阶段对缺血组织的表现较为不易察觉,这使得AIS梗死在非对称CT中的分割问题仍然较为复杂。具体而言,如图1.(a)所示,由于缺血组织引起的病理性低密度变化不易察觉,并且可能与正常生理变化或健康的脑组织特征相混淆。为了明确区分梗死区域与正常脑组织,临床实践中放射科医师常采用双侧半球的对比方法:如图1.(b)左侧所示,他们首先确定理想的中矢状线（MSL）,随后沿着MSL进行双侧半球差异比较,从而进一步定位梗死区域。

Abstract

摘要

量化梗死的估算在急性缺血性脑卒中（AIS）患者的诊断、治疗及预后预测中具有重要意义。

Method

方法

This section presents the developed PAPL model for AIS applications. Figures 2 illustrates the overall architecture.Then, three sequential stages of the PAPL outlined in Figure 2 are systematically elaborated upon in subsections III.C through III.E.

本节主要阐述了一种用于AIS梗死分割的新型PAPL方法。通过图2可以看出整体架构。具体而言，在III.C至III.E小节中详细阐述了PAPL的三个渐进阶段。

Conclusion

结论

In the early stage of AIS, the morphological alterations of ischemic brain tissue observed in NCCT are subtle and challenging to discern. As a result, achieving effective segmentation of early AIS Lesions on NCCT remains a formidable technical hurdle. Current approaches often overlook or conflate the distinct contributions of intrinsic and pathological asymmetries within bilateral hemisphere differences to the process of AIS Lesion segmentation when integrating domain-specific knowledge into deep learning frameworks. Drawing inspiration from this observation, we introduce a novel method termed Pathological Asymmetry-Guided Progressive Learning (PAPL), designed to facilitate the segmentation of AIS Lesions. PAPL mimics the incremental learning patterns observed in human cognition, comprising three sequential stages: preparation for knowledge acquisition, formal learning, and enhanced comprehension through refinement. By implementing this pre-learn-correct progressive strategy, we aim to enhance the precision of AIS Lesion segmentation. Comprehensive evaluations conducted on public datasets such as AISD and an in-house XWHD collection have demonstrated the superior performance of our proposed PAPL framework, which holds promise for improving lesion status assessment and prognosis prediction. Furthermore, our methodology is versatile enough to be integrated into other segmentation backbones, ensuring broad applicability across various computational platforms.

该种微血管病变在非对称CT成像中的特征变化不易于在急性缺血性脑卒中（AIS）早期阶段被早期识别；鉴于此，在非对称CT图像上实现AIS梗死的早期分割仍面临诸多技术挑战；现有深度学习模型往往难以有效整合这一领域的先验知识；在此背景下；我们发现现有方法在融合双侧半球内源性和病理性的不对称分布对于AIS梗死区域划分的独特影响存在不足；因此；本研究提出了一种基于路径学说的新渐进式学习框架——病理性不对称引导渐进学习法（Progressive Asymmetric Learning for Ischemic stroke, PAPL）；通过模拟人类逐步学习模式；包括知识准备阶段、正式学习阶段以及评估检验阶段；PAPL框架旨在提升AIS梗死区域划分的效果；大量实验证明该框架在公共AISD和内部XWHD数据集上均展现出显著优势；这不仅有助于提高临床评估效率及预后预测精度；也为其他分割主干集成提供了可行性方案

Results

结果

A. Segmentation Performance Against State-of-the-art Methods Table II presents a quantitative comparison of our proposed PAPL method with different SOTA segmentation techniques on the AISD and XWHD datasets. The results highlight superior performance using boldface notation. Among traditional medical image segmentation approaches such as U-Net, CENet, Attention U-Net, and U-Net++, these methods struggle to capture the subtle hypo-attenuation changes caused by AIS infarcts, resulting in poor segmentation outcomes. In contrast, TransU-Net offers long-distance dependent information that aids in distinguishing left and right image features. Consequently, our experiments demonstrate that TransU-NET surpasses these conventional methods in terms of segmentation performance. Furthermore, ablation studies conducted on the alignment process reveal that incorporating symmetric space alignment enhances the performance of these segmentation techniques. This is attributed to reduced spatial distribution discrepancies between training and test images when aligned appropriately. Additionally, specialized brain lesion segmentation methods like ISP-Net, XNet, and CLCI-NET excel at extracting hypo-attenuation features from AIS infarcts and achieve commendable results. Of particular note are brain hemisphere modeling-based approaches such as SEAN, ADN [5], SAN Net, and IS-NET; these methods further improve accuracy by integrating domain-specific knowledge about brain asymmetries into deep learning frameworks. Our PAPL method successfully mitigates adverse effects caused by normal physiological asymmetries during infarct segmentation processes. Experimental results confirm that our approach achieves a DSC score of 0.5468 on AISD and 0.4899 on XWHD datasets respectively

与现有的最先进分割方法进行对比分析 表II系统地展示了所提出的PAPL方法在AISD和XWHD数据集上的定量性能对比结果。其中最优结果以粗体形式标注并呈现更好的性能水平。值得注意的是，在一般医学图像分割任务中包括U-Net 、CENet 、Attention U-Net 和U-Net++在内的现有算法由于缺乏对远距离依赖信息的有效捕捉能力，在面对由AIS梗死引发的微小低密度变化时表现欠佳。相比之下 TransU-Net 通过其独特的架构设计能够有效地捕捉远距离依赖的空间信息 并在这一关键指标上展现出显著的优势 进一步提升分割性能的表现水平。此外 研究表明通过对非对称CT扫描进行校准处理 可以有效减少测试样本与训练样本之间的空间分布差异 进而提高模型的整体性能表现 某种程度上也验证了校准过程对于模型泛化能力提升的重要作用。基于中风病变特异性特征设计的专用分割网络如ISP-Net XNet 和CLCI-Net 在特征提取阶段展现出更强的优势 并能实现更好的分割效果 同时通过整合大脑对称建模技术的方法如SEAN ADN SANNet 和IS-Net 进一步提升了模型在这一领域的适用性。值得注意的是 所提出的PAPL方法特别针对由正常生理变化所带来的潜在不对称性问题 提出了有效的抑制机制 避免了这种潜在干扰对梗死区域分割造成的负面影响 最终在两个基准数据集AISD和XWHD上分别实现了Dice相似系数达0.5468和0.4899的最佳分割性能

Figure

图

_Fig. 1: (a) The AIS infarct segmentation on NCCT represents a highly non-trivial task due to the presence of subtle pathological changes. (b) A significant limitation of current approaches lies in their failure to properly account for the different contributions of intrinsic and pathological asymmetries in bilateral hemisphere differences during segmentation. (c) To address these challenges, we present a solution aimed at distinguishing these two types of asymmetries at the representation level, with the goal of maximizing the potential of domain-specific knowledge to enhance infarct segmentation.]

Fig. 1: (a) Acute ischemic stroke (AIS), as depicted in non-symmetric CT images, presents a highly complex task for the segmentation of infarct Lesions due to the subtle nature of pathological changes.(b) The inherent limitations of current methods lie in their tendency to either overlook or conflate the distinct contributions of intrinsic and pathological asymmetries in the two hemispheres when performing segmentation.(c) Consequently, we propose distinguishing these two types of asymmetry at the representation level to fully leverage domain knowledge and enhance infarct Lesion segmentation performance.

如图2所示：综合展示了所提出的病理不对称导向渐进式学习（Progressive Asymmetric Learning, PAPL）的整体说明，其中包括三个分阶段的过程：（1）知识准备阶段；（2）正式学习阶段；以及（3）考试提高阶段。

Fig. 2: 本研究提出的一种病理性非对称诱导的渐进式学习（PAPL）的整体示意图。该方法主要分为三个渐进步骤：(i) 知识储备环节；(ii) 核心学习环节；以及(iii) 优化提升环节。

Figure 3 illustrates the symmetric space transformation in NCCT scans. Through an affine transformation incorporating rotation (b) and translation (c), a given slice can be transformed from an unaligned state (a) to an aligned state (d).

Fig. 3展示了基于非对称CT扫描的空间对称转换示意图。该图通过仿射变换模型中包含的旋转变换（b）、平移变换（c），实现了给定切片从未配准的状态（a）到配准状态（d）的成功转换。

_Fig. 4: The generation of positive and negative samples through the designed region mask sampling process.

Fig. 4: 通过设计的区域掩码采样策略构建正负样本的示意图。

Fig. 5: Assessing the degree of differential significance φ within the stroke-related bilateral regions' intensity assessment.

Fig. 5: 评估梗死相关双侧区域强度比较中的差异显著度φ。

Fig. 6: Visualizing segmentation comparison on AISD dataset.

Fig. 6: 在AISD数据集上可视化分割比较。

Fig. 7: Visualizing segmentation comparison on XWHD dataset.

Fig. 7: 在XWHD数据集上可视化分割比较。

Fig. 8: Illustrating the acquired knowledge entities during the knowledge preparation stage. Red and blue points represent the t-SNE projections of positive and negative samples, respectively.

如图8所示：通过可视化工具展示了知识预处理阶段所学知识的表现形式。其中，在t-SNE算法作用下生成的正类样本点用红色标记表示，负类样本点用蓝色标记表示。

Fig. 9: Visualizing segmentation for the adjacent slices.

Fig. 9: 可视化相邻切片的分割结果。

Fig. 10: Visualizing feature representations and predicted outputs across various training epochs.

Fig. 10: 可视化不同训练时期的特征图和预测结果。

Fig. 11: This Bland-Altman plot shows the backbone and proposed methods in small-volume infarcts.

Fig. 11: 小体积梗死区域的主干网络和提出方法的 Bland-Altman 图。

Table

表

Table 1: Statistical distribution of AIS patients on XWHD. Age and TSS are presented as median and quartiles. NIHSS is presented as mean ± standard deviation.

数据表 I 急性缺血性脑卒中（AIS）患者的XWHD相关统计数据及其分布特征。研究者记录了纳入组患者的基线信息包括年龄以及Tremor Score（TSS），这些指标均采用中位数与百分位间距的方式进行描述；同时对尼霍夫评分系统（NIHSS）进行了标准化处理，并以均值±标准差的形式呈现结果。

TABLE II Systemically compares the performance of various approaches, including our PAPL method, on two benchmark datasets, AISD and XWHD. The † symbolizes stroke lesion-specific segmentation techniques, while the others denote brain symmetric modeling-based approaches. Among these comparisons, the top-performing models are highlighted with bold text, whereas the second-best performers are underlined.

表格II对比分析了其他方法与我们的PAPL在AISD和XWHD数据集上的性能。其中符号†代表中风病变特定分割方法，并通过粗体标记最佳性能，并用下划线标注次佳性能。

_TABLE III presents segmentation performance metrics in the AISD and XWHD datasets under altered backbone configurations.

表格III 在改变主干网络时在AISD和XWHD数据集上的分割性能。

TABLE IV Ablation analysis of different submodules on the AISD dataset.

表格 IV 在AISD数据集上不同子模块的剔除分析。

The performance evaluation of various techniques is compared for assessing small and large volumes of infarcts in the AISD dataset.

表格 V 急性缺血性脑卒中（AISD）数据集中小范围（≤70 ml）与大面积（>70 ml）梗死区域采用不同方法进行性能对比分析

_TABLE VISegmentation performance upon modifying various self-supervised contrastive learning techniques during the knowledge preparation stage.

表格 VI 在知识准备阶段使用不同自监督对比学习（CL）方法时的分割性能。

TABLE VIIAblation analysis of weight parameter λ1 and λ2 settings.

表格 VII权重参数λ₁和λ₂设置的剔除分析。

TABLE VIIIAblation analysis of weight parameter β1, β2 and β3 settings

表格 VIII 权重参数β₁、β₂和β₃设置的剔除分析