Annotation-Free Deep Learning of Large-Scale Nuclear Segmentation and Spatial Neighborhood Analysis on Multiplexed Fluorescence Images

Deep neural networks (DNNs) offer state-of-the-art performance for cell nucleus detection and segmentation. However, they require many manual annotations from skilled biologists for robust algorithm training, which is labor-intensive and not easily scalable. We propose an unsupervised expectation driving pipeline for Brain Cell Analysis using noisy Labels with minimal human input. 1) It uses a parametric method to generate noisy labels for cell nuclei and refines through an iterative training process. 2) We introduce a background recovery technique to enhance the detection and estimation of segmentation accuracy, especially in densely packed brain regions. 3) A novel sparse decomposition method is used to identify anomalous cell detections and automatically correct them to improve the accuracy further. We also provide extensive experiments evaluated on both supervised and unsupervised measurements to demonstrate our method's high effectiveness. The results of segmentation can be further used for phenotyping and cell localization. Besides, we proposed a spatial model to analyze the neuron-glia cells neighborhoods by cumulative influence of all neuron-glial pairs from the same circular surrounding area centered at the neuronal nuclei. A fast co-location analysis is applied to profile cell spatial neighborhoods in the healthy brain efficiently. LASSO-based feature selection methods are adopted to reveal their changes in different tissue conditions. Finally, we developed an accurate, fast-speed, and scalable method to align large-scale images from multi-round to pixel-level accuracy.