Computer-Aided Detection of Stained Neurons for Histological Image Analysis in Monkey Brain Tissue

Abstract

Mapping the architecture and circuitry of the brain is essential for identifying the neural computations and functional interactions that give rise to behavior. In conventional histology, tissue slices are examined section by section under a microscope, and labeled objects of interest are marked by hand in a manually intensive and time-consuming process. An integrated hardware and software system is needed that automates image acquisition, image processing, and object detection. Such a system would allow the high throughput analysis of histologically labeled tissue needed to efficiently map the architecture and circuitry of multiple whole brains. In this paper, we propose a fully automated integrated hardware and software system for image acquisition, preprocessing and analysis of immunohistochemically stained monkey brain slices for retrogradely-labeled neurons. This system solves the problems of image acquisition process and sample through an organic combination of automated microscope device and image processing algorithm, and obtains optimal image that focuses on wide area tissue and analyzes it using image processing algorithm. The proposed image acquisition method minimizes the noise of the light source by obtaining the average image through the acquired image several times in the fixed position, and solves the focus problem by acquiring the All-in-Focus (AF) image using Focus Measure (FM). In addition, multiple images are matched to obtain a wide area image, and up-sampling is performed to solve the resolution problem. We use the Simple Linear Clustering (SLIC) superpixel-based adaptive block processing and the Maximally Stable Extremal Regions(MSERs)-based clustering algorithm to reduce the image analysis space for individual neurons recognition in acquired images and to separate individual neurons from the images using the Marker-Controlled Watershed Transform (MCWT) algorithm. The partitioned information should be classified into two classes through a classifier design as candidate group data containing neurons and non-neurons together. A Convolutional Neural Networks (CNNs)-based algorithm is designed for this classification. CNNs can automatically obtain the optimal feature information of the neurons necessary for classification through training, so that the a priori design for obtaining this information can be excluded. In addition, the hybrid CNNs classification algorithm, that classifies feature information obtained from trained CNNs and manually acquired feature information using nonlinear Support Vector Machines (SVMs) improves brain cell classification performance. The system demonstrated the efficiency and accuracy of the system through qualitative and quantitative analysis. As a result, the designed system was able to acquire the optimal wide image from the histologically stained brain slice, the analysis space was minimized and the efficiency was high, and a fully automated integrated hardware and software system was constructed to identify individual brain cells with high probability.