Projectrons - Artificial Neural Networks for Learning Medical Images

点击翻译按钮获取中文摘要

Learning to retrieve similar images has many applications. In medical imaging, locating similar images can assist the diagnosis and treatment planning. The first step in any retrieval system is describing archived images. For that, image descriptors are used, either handcrafted ones like LBP (Local Binary Patterns) or learning--based ones like deep features. Deep features are the result of training deep neural networks such as CNN and autoencoders. In spite of high accuracies reported in the literature, deep architectures face several challenges, particularly for medical image analysis: 1) they require a large, balanced, and labeled dataset to learn properly, which cannot be guaranteed in the medical field. Medical imaging archives are very large, but the nature of medical imaging and the complex human anatomy creates imbalance classes (e.g., many images for breast cancer, and comparably few images for thyroid cancer) that cannot be balanced via augmentation operations (i.e., artificially increase the number of images through rotations). Additionally, in many sub-fields of medical imaging, the clinical workflow does not include labeling (marking) regions of interest, a requirement for training of supervised deep nets that cannot be easily circumvented, 2) deep architectures are prone to adversarial attacks and may easily exhibit overfitting behaviour (memorize the data instead of actually learning it), 3) deep features are not interpretable. It is not possible to understand why a certain action was taken. In the medical filed, decisions have to be explained and described in details (i.e., in pathology reports). I will explore the design, training, and validation of a new type of neural networks called "Projectron" that are particularly useful for representing medical images. Such representations can be used for image search and identification in large archives of medical images. A Projectron is a connected graph with a relatively small number of neurons that can learn an image representation (in an unsupervised manner) and create a compact (binary) output especially learned for tagging images to facilitate efficient storage and fast search in large archives. A Projectron may not need labeled data and if does, it should not need millions. As well, the Projectron will be immune to overfitting and cannot be fooled into falsifying representations. The relationship between the image and the way it is projected into the graph should be verifiable for the physicians. As backpropgation algorithms would not work on connected graphs, new ways for learning will be explored, among others evolutionary metaheuristics, and a mixture of the random sampling of sub-graphs guided by opposition-based (anti-thetic) pairing. All designed graphs will be tested on histopathology and radiology images. In this project, I will train 2 PhD students, 1 MSc student, 10 undergraduate students, and a postdoctoral fellow. **