Deep learning for medical computer vision: Beyond more data and more computing power

点击翻译按钮获取中文摘要

Image data is collected for a variety of applications, e.g., manufacturing, transportation, astronomy, security, and agriculture, as interpreting these images can lead to new insights and discoveries, better-informed decisions, and increased productivity. Biomedical images have revolutionized biology and medicine by giving clinicians and scientists visual access to ex/in-vivo anatomy and function of cells, tissues, organs, and whole organisms in healthy and diseased states. As the number and size of biomedical images are growing rapidly, finer details are captured in shorter times, and image dimensionality is increasing from scalar 2D to dynamic multi-valued 3D, images can no longer be interpreted via manual visual inspection. Medical computer vision (MCV), the topic of this proposed research, is the discipline tasked with developing computer systems that interpret biomedical images. In recent years, deep learning (DL), a subset of machine learning, which in turn is a subset of artificial intelligence, has become the de-facto computational methodology for tackling MCV problems across the spectrums of imaging modalities and clinical applications. The impressive superior performance results of DL methods, reported in research papers, are difficult to ignore. DL is attracting extraordinary attention and the availability of relatively easy to use DL tools is bringing an onrush of users transfixed by the perceived promise that DL is the solution to all problems. Undoubtedly DL has something valuable to offer towards addressing MCV problems. However, there are important challenges to overcome, beyond naïvely seeking more data and more computing power, before DL-based MCV technologies become trusted, reliable components that can be deployed in critical bioimaging-driven clinical workflows or biological discoveries that ultimately can lead to advancing science and improving healthcare. The proposed research focuses on creating novel automated MCV techniques capable of accurate, robust, and fast bioimage interpretation by tackling the critical issues surrounding DL and DL-based MCV, such as, fairness, generalizability, explainability, data-reliance, trust, and model design. The proposed research aims at answering the following questions: How to identify, enhance, and leverage resources to train DL MCV systems (e.g., raw image data, example interpretations, and domain-knowledge)? What is the landscape of possible systems and how to explore it in order to arrive at useful systems? What are the different criteria and tradeoffs involved in assessing such systems (e.g., accuracy and explainability)? And what are some of the computational challenges that arise when considering real-world deployment of such systems (e.g., data privacy and continual learning)?