Alzheimer's disease and Frontotemporal dementia are common forms of neurodegenerative dementia. Behavioral alterations and cognitive impairments are found in the clinical courses of both diseases and their differential diagnosis is sometimes difficult for physicians. Therefore, an accurate tool dedicated to this diagnostic challenge can be valuable in clinical practice. However, current structural imaging methods mainly focus on the detection of each disease but rarely on their differential diagnosis. In this paper, we propose a deep learning based approach for both problems of disease detection and differential diagnosis. We suggest utilizing two types of biomarkers for this application: structure grading and structure atrophy. First, we propose to train a large ensemble of 3D U-Nets to locally determine the anatomical patterns of healthy people, patients with Alzheimer's disease and patients with Frontotemporal dementia using structural MRI as input. The output of the ensemble is a 2-channel disease's coordinate map able to be transformed into a 3D grading map which is easy to interpret for clinicians. This 2-channel map is coupled with a multi-layer perceptron classifier for different classification tasks. Second, we propose to combine our deep learning framework with a traditional machine learning strategy based on volume to improve the model discriminative capacity and robustness. After both cross-validation and external validation, our experiments based on 3319 MRI demonstrated competitive results of our method compared to the state-of-the-art methods for both disease detection and differential diagnosis.

翻译：阿尔茨海默病与额颞叶痴呆是常见的神经退行性痴呆类型。两种疾病的临床病程均涉及行为改变和认知障碍，临床医师有时难以进行鉴别诊断。因此，针对这一诊断难题的精准工具在临床实践中具有重要价值。然而，当前结构影像学方法主要侧重于单种疾病的检测，鲜少关注鉴别诊断。本文提出一种基于深度学习的方法，同时解决疾病检测与鉴别诊断两个问题。我们建议在应用中采用两类生物标志物：结构分级与结构萎缩。首先，我们提出训练由3D U-Net组成的大型集成模型，以结构MRI为输入，在局部确定健康人群、阿尔茨海默病患者及额颞叶痴呆患者的解剖模式。集成模型输出为双通道疾病坐标图，可转化为临床医师易于解读的三维分级图。该双通道图与多层感知机分类器结合，用于不同分类任务。其次，我们提出将深度学习框架与基于容积的传统机器学习策略相结合，以提升模型的区分能力与鲁棒性。经过交叉验证与外部验证，基于3319例MRI数据的实验表明，我们的方法在疾病检测与鉴别诊断方面均达到了与当前最优方法竞争的性能。