Wurts Alexander, Oakley Derek H, Hyman Bradley T, Samsi Siddharth
Annu Int Conf IEEE Eng Med Biol Soc. 2020 Jul;2020:1420-1423. doi: 10.1109/EMBC44109.2020.9175832.
Alzheimers disease is characterized by complex changes in brain tissue including the accumulation of tau-containing neurofibrillary tangles (NFTs) and dystrophic neurites (DNs) within neurons. The distribution and density of tau pathology throughout the brain is evaluated at autopsy as one component of Alzheimers disease diagnosis. Deep neural networks (DNN) have been shown to be effective in the quantification of tau pathology when trained on fully annotated images. In this paper, we examine the effectiveness of three DNNs for the segmentation of tau pathology when trained on noisily labeled data. We train FCN, SegNet and U-Net on the same set of training images. Our results show that using noisily labeled data, these networks are capable of segmenting tau pathology as well as nuclei in as few as 40 training epochs with varying degrees of success. SegNet, FCN and U-Net are able to achieve a DICE loss of 0.234, 0.297 and 0.272 respectively on the task of segmenting regions of tau. We also apply these networks to the task of segmenting whole slide images of tissue sections and discuss their practical applicability for processing gigapixel sized images.
阿尔茨海默病的特征是脑组织发生复杂变化，包括神经元内含有tau蛋白的神经原纤维缠结（NFTs）和营养不良性神经突（DNs）的积累。在尸检时评估tau病理在整个大脑中的分布和密度，作为阿尔茨海默病诊断的一个组成部分。当在完全标注的图像上进行训练时，深度神经网络（DNN）已被证明在量化tau病理方面是有效的。在本文中，我们研究了三种深度神经网络在有噪声标注数据上进行训练时对tau病理分割的有效性。我们在同一组训练图像上训练全卷积网络（FCN）、SegNet和U-Net。我们的结果表明，使用有噪声标注的数据，这些网络能够在少至40个训练轮次中成功地分割tau病理以及细胞核，且成功率各不相同。在分割tau区域的任务上，SegNet、FCN和U-Net分别能够实现0.234、0.297和0.272的DICE损失。我们还将这些网络应用于组织切片全切片图像的分割任务，并讨论它们在处理千兆像素大小图像方面的实际适用性。
