Machine learning approaches to image analysis

点击翻译按钮获取中文摘要

My research program focusses on developing new computational approaches and analytical tools for healthcare policy and delivery, aligned with opportunities afforded by the intersection of machine learning (ML) with advances in computing power. This has included development and optimization of agent-based modeling frameworks to simulate large-scale infection spread and advances in image processing in mHealth contexts. The overall research focus is on better technology-based decision-making tools for policy and clinical practice. Image recognition and image analysis is a very rapidly growing form of artificial intelligence used in a myriad of social and business contexts, from auto-organizing images in social media to driverless vehicles recognizing obstacles in their path. Medical image processing is a specialization within the field, gaining attention in recent years for its anticipated game-changing applications in clinical decision support, diagnostic tools, precision medicine, and population health. Medical image processing is complex. It is widely recognized that the stakes are generally higher than in other image recognition application areas, with required levels of reliability lagging. Meaningful medical image analysis requires massive amounts of data, which are currently not available for many classes of medical images. The trend toward higher resolution and 3-dimensional medical images, combined with expectation of efficiency create new computational challenges. There is no single best algorithm applied to medical image processing, leaving the best methods as an open question. This research program will construct and optimize algorithmic pipelines for medical image analysis, leveraging the enormous progress in ML combined with analytical tools in image processing. The lack of labelled wound data creates challenge and complexity. The research will unfold within the constraints of off-the-shelf mobile devices with only built-in sensors and cameras. This introduces significantly higher variability in images than MRI or CT imaging, yet reflects patient-driven healthcare in which we collect, research, and share information on our mobile devices. The goal is to develop new approaches to image analysis that can offer decision, diagnostic, and telehealth tools to Canadian healthcare. To do so, the work will develop tools to crowdsource medical image data for wound images, addressing a current gap required for the research and providing a database for other researchers' use. The work will construct and optimize algorithms to accurately and efficiently identify features and classify them in medical images. Of specific interest are generative adversarial networks, a promising ML approach when training data are limited, and convolutional neural networks, one of several ML approaches modeled loosely on the human brain and well-adapted to image classification.