A novel approach to segment cortical neurons in histological images of the near-term fetal sheep brain model.

Abstract

Oxygen deprivation (hypoxia) and reduced blood supply (ischemia) can occur before, during or shortly after birth and can result in death, brain damage and long-term disability. Assessing neuronal survival after hypoxia-ischemia in the near-term fetal sheep brain model is essential for the development of novel treatment strategies. As manual quantification of neurons in histological images varies between different assessors and is extremely time-consuming, automation of the process is needed and has not been currently achieved. To achieve automation, successfully segmenting the neurons from the background is very important. Due to presence of densely populated overlapping cells and with no prior information of shapes and sizes, the segmentation of neurons from the image is complex. Initially, we segmented the RGB images by using K-means clustering to primarily segment the neurons from the background based on their colour value, a distance transform for seed detection and watershed method for separating overlapping objects. However, this resulted in unsatisfactory sensitivity and performance due to over-segmentation if we use the RGB image directly. In this paper, we propose a semi-automated modified approach to segment neurons that tackles the over-segmentation issue that we encountered. Initially, we separated the red, green and blue colour channel information from the RGB image. We determined that by applying the same segmentation method first to the blue channel image, then by performing segmentation on the green channel for the neurons that remain unsegmented from the blue channel segmentation and finally by performing segmentation on red channel for neurons that were still unsegmented from the green channel segmentation, improved performance results could be achieved. The modified approach increased performance for the healthy and ischemic animal images from 89.7% to 98.08% and from 94.36% to 98.06% respectively as compared to using RGB image directly.