Automatic Cell Segmentation by Adaptive Thresholding (ACSAT) for large scale calcium imaging datasets

Abstract

Advances in calcium imaging have made it possible to record from an increasingly larger number of neurons simultaneously. Neuroscientists can now routinely image hundreds to thousands of individual neurons. With the continued neurotechnology development effort, it is expected that millions of neurons could soon be simultaneously measured. An emerging technical challenge that parallels the advancement in imaging such a large number of individual neurons is the processing of correspondingly large datasets, an important step of which is the identification of individual neurons. Traditional methods rely mainly on manual or semi-manual inspection, which cannot be scaled to processing large datasets. To address this challenge, we have developed an automated cell segmentation method, which is referred to as Automated Cell Segmentation by Adaptive Thresholding (ACSAT). ACSAT includes an iterative procedure that automatically calculates global and local threshold values during each iteration based on image pixel intensities. As such, the algorithm is capable of handling morphological variations and dynamic changes in fluorescence intensities in different calcium imaging datasets. In addition, ACSAT computes adaptive threshold values based on a time-collapsed image that is representative of the image sequence, and thus ACSAT provides segmentation results at a fast speed. We tested the algorithm on wide-field calcium imaging datasets in the brain regions of hippocampus and striatum in mice. ACSAT achieved precision and recall rates of approximately 80% when compared to individual neurons that are verified by human inspection. Additionally, ACSAT successfully detected low-intensity neurons that were initially undetected by humans.SignificanceACSAT automatically segments cells in large scale wide-field calcium imaging datasets. It is based on adaptive thresholding at both global and local levels, implemented in an iterative process to identify individual neurons in a time-collapsed image from an image sequence. It is therefore capable of handling variation in cell morphology and dynamic changes between different calcium imaging datasets at a fast speed. Based on tests performed on two datasets from mouse hippocampus and striatum, ACSAT performed comparable to human referees and was even able to detect low-intensity neurons that were initially undetected by human referees.

Evaluation of cell segmentation methods without reference segmentations

“…Existing segmentation methods can be divided into two categories, the geometry-based segmentation techniques and deep learning-based approaches. The former includes but is not limited to threshold-based segmentation (Shen et al, 2018), region-based segmentation (Panagiotakis & Argyros, 2018), watershed algorithm and its variants (Ji et al, 2015), active contour (Wu et al, 2015), Chan-Vese segmentation (Braiki et al, 2020;Fan et al, 2013), and Graph-cut based segmentation (Oyebode & Tapamo, 2016). In the deep learning category, the conventional deep convolutional neural network (CNN) was initially applied (Jung et al, 2019;Sadanandan et al, 2017).…”