Pyramidal Rat Neurons Segmentation in Microscopy Low-Resolution Images

Abstract

Cell analysis in image digital microscopy isa relevant tool in modern cell biology since it allows studying their behavior and morphology in differentt issues. Although there is a robust development in microscope technology, cells like live neurons are fragile due to simple factors such as illumination, which could compromise their viability. Therefore, neurons must be analyzed in a low-resolution condition. Besides, the identification and selection of neurons in images from a microscope are visually made, whichis time-consuming and increases the subjectivity of the process and human error. Computer vision techniques and Neuronal Networks help automate these tasks while guaranteeing the application of constant criteria. This work aimed to obtain automatic segmentation of neuronsin low-resolution images from an inverse microscopy used to study and test live neurons. The proposed methodology allows for separating the neuron from the background despite the high noise generated by reflectance and distortion when observing the sample through the liquid solution and the petri dish. Theresults of traditional methods and Convolutional Neural Networks (U-Net) are compared, showing that, despite the high image noise condition, it is possible to reach aDice index of 0.73±0.07 in segmentation.