Multi-layer perceptron classification & quantification of neuronal survival in hypoxic-ischemic brain image slices using a novel gradient direction, grey level co-occurrence matrix image training

Abstract

Hypoxic ischemic encephalopathy (HIE) is a major global cause of neonatal death and lifelong disability. Large animal translational studies of hypoxic ischemic brain injury, such as those conducted in fetal sheep, have and continue to play a key role in furthering our understanding of the cellular and molecular mechanisms of injury and developing new treatment strategies for clinical translation. At present, the quantification of neurons in histological images consists of slow, manually intensive morphological assessment, requiring many repeats by an expert, which can prove to be time-consuming and prone to human error. Hence, there is an urgent need to automate the neuron classification and quantification process. In this article, we present a ’Gradient Direction, Grey level Co-occurrence Matrix’ (GD-GLCM) image training method which outperforms and simplifies the standard training methodology using texture analysis to cell-classification. This is achieved by determining the Grey level Co-occurrence Matrix of the gradient direction of a cell image followed by direct passing to a classifier in the form of a Multilayer Perceptron (MLP). Hence, avoiding all texture feature computation steps. The proposed MLP is trained on both healthy and dying neurons that are manually identified by an expert and validated on unseen hypoxic-ischemic brain slice images from the fetal sheep in utero model. We compared the performance of our classifier using the gradient magnitude dataset as well as the gradient direction dataset. We also compare the performance of a perceptron, a 1-layer MLP, and a 2-layer MLP to each other. We demonstrate here a way of accurately identifying both healthy and dying cortical neurons obtained from brain slice images of the fetal sheep model under global hypoxia to high precision by identifying the most minimised MLP architecture, minimised input space (GLCM size) and minimised training data (GLCM representations) to achieve the highest performance over the standard methodology.