Abstract:
The arising-from-chair task assessment is a key aspect of the evaluation of movement disorders in Parkinson\'s disease (PD). However, common scale-based clinical assessment methods are highly subjective and dependent on the neurologist\'s expertise. Alternate automated methods for arising-from-chair assessment can be established based on quantitative susceptibility mapping (QSM) images with multiple-instance learning. However, performance stability for such methods can be typically undermined by the presence of irrelevant or spuriously-relevant features that mask the intrinsic causal features. Therefore, we propose a QSM-based arising-from-chair assessment method using a causal graph-neural-network framework, where counterfactual and debiasing strategies are developed and integrated into this framework for capturing causal features. Specifically, the counterfactual strategy is proposed to suppress irrelevant features caused by background noise, by producing incorrect predictions when dropping causal parts. The debiasing strategy is proposed to suppress spuriously relevant features caused by the sampling bias and it comprises a resampling guidance scheme for selecting stable instances and a causal invariance constraint for improving stability under various interferences. The results of extensive experiments demonstrated the superiority of the proposed method in detecting arising-from-chair abnormalities. Its clinical feasibility was further confirmed by the coincidence between the selected causal features and those reported in earlier medical studies. Additionally, the proposed method was extensible for another motion task of leg agility. Overall, this study provides a potential tool for automated arising-from-chair assessment in PD patients, and also introduces causal counterfactual thinking in medical image analysis. Our source code is publicly available at https://github.com/SJTUBME-QianLab/CFGNN-PDarising.
摘要:
主席起跳任务评估是帕金森病(PD)运动障碍评估的一个关键方面。然而,常见的基于量表的临床评估方法是高度主观的,并且依赖于神经科医师的专业知识。可以基于具有多实例学习的定量磁化率映射(QSM)图像来建立用于椅子产生评估的替代自动化方法。然而,这种方法的性能稳定性通常会由于存在掩盖内在因果特征的不相关或虚假相关特征而受到损害。因此,我们提出了一种基于QSM的椅子产生评估方法,使用因果图神经网络框架,其中反事实和反偏见策略被开发并整合到这个框架中,以捕获因果特征。具体来说,提出了反事实策略来抑制背景噪声引起的无关特征,通过在丢弃因果部分时产生不正确的预测。提出了去偏置策略来抑制由采样偏差引起的虚假相关特征,它包括用于选择稳定实例的重采样指导方案和用于在各种干扰下提高稳定性的因果不变性约束。大量实验的结果表明了所提出的方法在检测椅子异常方面的优越性。所选择的因果特征与早期医学研究中报道的因果特征之间的一致性进一步证实了其临床可行性。此外,所提出的方法是可扩展的另一个运动任务的腿敏捷性。总的来说,这项研究为PD患者的自动起椅评估提供了一个潜在的工具,并在医学图像分析中引入了因果反事实思维。我们的源代码可在https://github.com/SJTUBME-QianLab/CFGNN-PDarising上公开获得。