The role of biomarkers in cell counting with U‐Net CNN

Abstract

AbstractBackgroundImaging methods that are non‐destructive preserve tissue integrity and lead to improved topography analysis for the study of neurodegenerative diseases like Alzheimer’s Disease (AD). Immunofluorescence (IF) staining methods facilitate such investigations by addressing a gradient of intensity variations relevant for in situ cell detection [1] and morphology quantification. This paper reports on the colocalization of proteins and cell structures from human post‐mortem tissue by evaluating IF expression by means of TBR1 and NeuN biomarkers. We compare different U‐Net‐based models by exploring image representations obtained by combining image channels with different biomarkers.MethodFrom an image dataset with 21 full‐size IF micrographs, we perform data augmentation by creating a cropping mechanism resulting in 40 to 60 image crops from each micrograph, each crop with 128x128 pixels. Later, we enhance the resulting 1,200 cropped images, and construct 3‐channel (X,Y,Z) image inputs of dimensions 128x128x3 for the U‐Net model in a two‐fold process: (I) stack all 3 channels with single‐channel IF cropped images corresponding to NeuN biomarkers (NeuN, NeuN, NeuN); (II) stack the single channel images of TBR1, NeuN, and a single image representing the maximum pixel intensities from TBR1 and NeuN images (TBR1, NeuN, max (TBR1, NeuN)). With a test‐train split of 70% ‐ 30%, and parameter tuning, we compare the two models via metrics such as neuron areas, counts, accuracy and recall.ResultThe model trained with channel stacking (II) outperforms the model trained with (I) with an accuracy and recall of 0.96 and 0.5, as opposed to 0.95 and 0.4. We also found a great variation in neuron area and counts as depicted in the images below.ConclusionWhile this is the first step in accelerating the quantification of neurons for the pathophysiology studies necessary for drug development, there is a need in establishing protocols for the generation of datasets with manual segmentation to further train our models. The inclusion of prior information about the expected neuron morphology increases the performance of our networks, and can successfully deploy a foreseen automated pipeline. [1] Ehrenberg et al. A manual multiplex immunofluorescence method for investigating neurodegenerative diseases. Journal of Neuroscience Methods,June, 2020.