An Interactive Deep Learning Toolkit for Automatic Segmentation of Images

Abstract

This thesis presents the design, analysis, and implementation of an interactive deep learning toolkit to reduce the manual labeling effort and improve the automatic segmentation of image data. While deep neural networks (DNN) models work well for image classification and segmentation, they require a large amount of training data, which is often acquired by manual annotations, a process that is time consuming and tedious. We develop our framework to train a deep neural network classifier in the background during the annotation process and provide the user with real-time feedback to guide the annotation effort. Our interactive system is developed in the context of segmentation of neuronal structures in Electron Microscopy (EM) images. This application is very important for the efficient mapping of brain structure and connectivity, enabling neuroanatomists to gain new insight into the functional structure of the brain. The architecture of the system employs a three-tiered approach that provides a web-based user interface for distributed annotations, a learning model that trains a deep neural network classifier for segmentation purposes, and a web service middle-ware that facilitates data exchange between the user interface and the learning model.