Study On Deep Learning Based Segmentation And Quantitative Analysis Method Of Aβ Plaques In The Whole Rat Brain

Alzheimer’s disease has created a significant burden on modern society with an increasingly aging population.The studies on developing drugs and treatment for Alzheimer’s disease have so far nearly failed.The major reason is that the transgenic mouse model that has been employed for research cannot accurately mimic the disease pathogenesis.The gene knock-in rat model can exhibit more comprehensive pathological manifestations of Alzheimer’s disease,research on this rat model will promote the clinical transformation of Alzheimer’s disease drugs.Among the pathologies of Alzheimer’s disease,the distribution of Amyloid-β(Aβ)plaques in the whole brain is a key indicator of the disease course.However,quantitative analysis of Aβ plaques in the whole rat brain is difficult due to the limitations of existing Aβ plaque labeling methods,imaging methods,segmentation and registration methods.The main purpose of this study is to obtain whole rat brain Aβ plaque datasets on mesoscale and develop a pipeline for accurate and efficient segmentation and quantitative analysis of the Aβ plaque distribution in whole rat brain.The details of this study are summarized as follows:(1)In this study,a terabyte three-dimensional microscopic brain-wide Aβ plaque dataset of the Alzheimer’s disease rat is obtained at 1 μm × 1μm × 3.5 μm resolution through immunofluorescence staining and high-speed three-dimensional microscopic imaging method.Segmentation is essential for quantitative analysis,and deep learning methods have been widely used in the image segmentation task in recent years.However,the data-driven deep learning method requires a significant amount of manual annotations,particularly for 3D microscopic dataset which requires a higher cost of manual labeling.To reduce the human-labeling cost,this study proposes a segmentation algorithm based on weakly supervised learning.This method utilizes the highresolution network for maintaining the subtle features of Aβ plaques,combined with the object detection network for accurate positioning of the Aβ plaques,and then segments Aβ plaques through post-processing.Meanwhile,the additional visual cues generated by the network are used to improve the segmentation accuracy.This study uses the Aβ plaques dataset annotated by object-level labels and pixel-level labels for the network training and evaluation separately.Extensive experiments show the effectiveness of different modules in our method,the generalization in different brain regions,and the accuracy in various plaque sizes.Comparative experiment shows that our segmentation method significantly outperforms the most popular traditional segmentation methods,fully supervised segmentation methods,and weakly supervised segmentation methods.(2)In order to quantitatively analyze the Aβ plaques distribution of different brain regions in the whole rat brain,obtaining the spatial location of Aβ plaques in the whole brain through brain atlas registration is a crucial step.The cross-modal dataset complicates the registration process since the 3D microscopic dataset collected in this study uses a different imaging technique than the rat brain atlas template dataset.This study combines an automatic registration algorithm with interactive local registration to obtain accurate contours of brain regions and generate the deformation field.Based on the above research,the segmentation method proposed in this study is further applied to the brain-wide Aβ plaque dataset and obtain the brain-wide Aβ plaque masks.The registration deformation field can be used to map the brain-wide Aβ plaque masks to the corresponding brain regions and quantitatively analyze the plaques in the whole rat brain.