Machine learning and computational methods for studying cortical neurons in the aged brain

Abstract

Ageing is one of the greatest challenges of the 21st century, with the weakening of cognitive abilities, increasing risk of certain diseases, and reduced recovery from injury. We studied injury in the aged brain to explore the fundamental mechanisms of cortical axons, to provide a baseline for injury in the healthy aged cortex, and to find potential therapeutic targets that might improve recovery following injury. To study mechanisms of axonal degeneration, regeneration, and synaptic dynamics, two-photon microscopy was used to capture images of axons before and after a laser-mediated injury. A dataset of two-photon axon images was collected to study axonal dynamics in the aged brain, and we developed tools to better automate their analysis. In particular, we focused on the use of machine learning to automatically detect synapses and segment axons, which is important for the study of axon dynamics. To detect the synapses, a combination of computer vision and machine learning was used with hand-optimised features; this was shown to achieve high performance in comparison to an existing method. To augment our image dataset, we used a deep generative model to synthesise additional training data, and showed that this can be an effective approach to improve segmentation in the presence of sparse or highly variable data. Using some of these techniques in our experimental study of injury in the aged brain, we found that while degeneration in the aged brain is comparable to the young brain, that regeneration is impaired. There was also an increase in synaptic dynamics after introducing the laser induced lesions, indicating a re-wiring mechanism. The functional consequences of synaptic re-wiring were investigated using a computational model to determine whether this response could compensate for reduced regeneration in the aged brain, and we found that partial recovery is possible according to our model.