Active Learning Methods for Computational Delineation and Cellular Profiling of Cortical Layers in Rat Whole Brain Sections using Multiplex Immunofluorescent Imaging and Classification of Histopathological Images using Convolutional Neural Networks

Most machine learning algorithms require an abundance of high-quality training data. Such a requirement creates a major obstacle when using machine learning in the medical image domain, as labeled data collection is difficult. We explore active learning solutions for cortical layer delineation and for training Convolutional Neural Networks with less amount of labeled data. For the former, we develop an objective and automated multiplex imaging-based method for delineation of cortical layers in the whole brain sections. This is an advance over current methods where layers are visually delineated by biologists. We further carryout comprehensive and quantitative profiling of the cell layers with respect to their composition (presence of neuronal and glial cell types and sub-types), cell-phenotypic status, and the spatial arrangement of cells. Our method is based on spatial cluster analysis of neuronal features using the Dirichlet Process Mixture Model and refined using active machine learning. It is versatile, modular, and readily amenable to visual inspection and proofreading. The accuracy of the computational cortical layer delineation was validated by comparing it to brain sections that were immunostained with layer-specific molecular markers (NECAB1, FOXP1) and by comparison against manual delineation by biologists. We implement our proposed method on healthy rat brains and rat brains with mild traumatic brain injury (mTBI). Our in-depth cellular profiling of the layers allows us to study the patterns of tissue perturbations in the cortex for mTBI brains. We propose whole cell morphological segmentation methods for five different types of cells which allow an in-depth analysis of the cell state activation and spatial distribution. These are also used in neuronal feature extraction for cortical layer delineation. For the second implementation of active learning, we formulate an active deep learning framework to train CNNs with less amount of labeled data. We implement two parallel active learning criteria for the same. We provide extensive experimental results and in-depth analysis to demonstrate the effectiveness of our algorithm on a breast tumor classification problem. We offer active learning solutions for addressing two different problems encountered in whole brain analysis.