葛威1, 刘汝璇2, 郑菲2, 刘海华2,3, 唐奇伶2(1.中南民族大学;2.中南民族大学生物医学工程学院;3.医学信息分析及肿瘤诊疗重点实验室)
目的 阿尔茨海默症(Alzheimer"s disease, AD)是一种常见于中老年人群的神经退行性疾病,并伴随着认知障碍及记忆衰退。随着全球人口老年化趋势日益加剧,阿尔茨海默症的及时诊断与病理区域的可视化及其准确定位具有重要的临床意义。目前的研究中,基于块级和区域级的检测,由于采用非线性交互很难解释影响模型决策的病理区域。为了解决这个问题,本文提出了一种AD病理区域定位及诊断的联合学习框架。方法 利用反事实推理的思想,基于前景背景注意力掩码构建注意力引导的循环生成对抗网络(Attention-guided cycle generative adversarial networks, ACGAN)可视化AD患者的病理区域,并使用生成的病理区域知识指导增强诊断模型。具体来说,通过在ACGAN模型的生成器中设计注意力掩码来引导生成方案,使模型更好地聚焦于疾病的病理区域,有效地捕捉突出的全局特征。并通过ACGAN模型中病理区域生成器实现sMRI图像在源域和目标域之间的转换清晰地划分出细微的病理区域。利用生成的病理区域知识作为指导,并结合三维坐标注意力与全局局部注意力,获取三维图像之间的依赖关系及三维空间的位置信息,优化诊断模型。结果 为了验证方法的有效性,在公开的ADNI(Alzheimer’s Disease Neuroimaging Initiative)数据集上对模型进行评估,与传统的CNN模型及几种较为先进的AD分类诊断模型相比,本文使用病理区域知识指导增强诊断模型显示出优越的诊断性能,相比与性能最好的的方法ACC、F1-score、AUC提高了3.60%、5.02%、1.94%.并对生成的病理区域图像进行定性及定量评估,本文方法得到的病理区域图像归一化互相关分数和峰值信噪比均优于对比的方法。结论 与现有方法相比,本文ACGAN模型可以学习sMRI图像在源域和目标域之间的转换,能够准确地捕获全局特征及病理区域..并将学习到的病理区域知识用于AD诊断模型的改进,使分类诊断模型取得了卓越的性能。
Enhancing diagnostic model for Alzheimer
(Departent of Biomedical Engineering,South-Central Minzu University)
Alzheimer"s disease is a neurodegenerative disease occurring commonly in middle-aged and elderly populations and is accompanied by cognitive impairment and memory loss. With the increasing trend of global population aging, timely diagnosis of Alzheimer"s disease and visualization of pathological regions and its accurate localization are of great clinical importance. In current research, one mainstream approach is to extract patch-level features based on voxel morphology and prior knowledge for detecting structural changes and identifying AD-related voxel structures; and another approach is to learn AD-related pathological regions by focusing the network on specific brain regions of interest (e.g., cortical and hippocampal regions) based on regional features, but these approaches ignore other pathological locations in the brain and cannot accurately obtain global structural information for the diagnosis of AD. In order to obtain a convincing model architecture and interpretability of the output, highlighting the information of pathological regions, we propose a joint learning model for localization and diagnosis of AD pathological regions, using the idea of counterfactual reasoning, constructing an attention-guided recurrent generative adversarial network based on foreground background attention mask Method In the vast majority of image classification methods, the task of the network model is to find which part of the input is X and which part influences the classifier"s decision to judge the final result as Y. Thinking about it another way, in a hypothetical scenario where the input X is C , would the result be Z instead of Y? This idea is defined as counterfactual reasoning. To construct the output in the hypothetical scenario, we first trained the AD classification model as a classifier in the hypothetical scenario and obtained the pathological features of AD. The hypothetical scenario was constructed using a generative adversarial network to learn the mapping of images from the source domain to the target field. However, due to the complexity of whole brain sMRI images and the huge amount of information in 3D space, it is difficult to achieve good results by directly generating image to image transformation. Inspiration from two models, CycleGAN and AttentionGAN, by changing the region in the original image that affects the category judgment, the image can be mapped from the source domain to the target domain, and using the foreground background attention to guide the model to focus on the dynamic change region, which reduces the complexity of the model and makes the model easier to fit. Therefore, this paper proposes an attention-guided recurrent generative adversarial network to construct a counterfactual mapping model for Alzheimer"s disease, so that the corresponding pathological regions are output. If a counterfactual map conditional on the target label (i.e., hypothetical scenario) is generated, this counterfactual map is added to the input image so that the transformed image is diagnosed as the target type. For example, when the counterfactual map is added to the sMRI image of a subject with AD, changing the corresponding region causes the input sMRI image to change and thus be diagnosed by the classifier as a normal subject. The pathological regions represented by the counterfactual map were used as privileged information, i.e., the location information of the counterfactual map influenced the category determination, to further optimize the diagnostic model, so that the diagnostic model focused on learning and discovering disease-related discriminative regions, and to combine the pathological region generation model with the AD diagnostic model. Result: We evaluated our model against traditional CNN models and several more advanced AD diagnostic models on a publicly available ADNI dataset with quantitative evaluation metrics including ACC, F1-score, and AUC. experimental results showed that our model improved ACC, F1-score, and AUC by 3.60%, 5.02%, and 1.94% compared to the best performing method, 1.94%. The generated pathological region images were also evaluated qualitatively and quantitatively, and the normalized correlation scores and peak signal-to-noise ratios of the pathological region images obtained by our method were better than those of the compared methods. More importantly, our AD diagnostic model can visualize the global features and fine-grained discriminative regions of the pathological regions compared with the benchmark model, and the average accuracy after three iterations is improved by +4.90%, +11.03% and +11.08%, respectively, compared with the benchmark method. Conclusion Compared with existing methods, this ACGAN model can learn the transformation of sMRI images between source and target fields, and can accurately capture global features and pathological regions. And the learned pathological region knowledge is used for the improvement of AD diagnosis model, so that the classification diagnosis model achieves excellent performance.