Towards Generalizable Reasoned Deep Learning for Efficient Interpretable Medical Image Computing

图像计算是一个关键的技术领域，从工程到医学的许多领域都有应用。人工智能（AI）的一个子集，涉及创建能够模拟人类视觉的计算机系统或机器，该领域仍在努力应对成像数据的巨大可变性和不同应用相关分析需求所带来的技术挑战。最近，一股连接主义机器学习浪潮席卷了该领域，其中系统通过训练来自庞大数据库的标记图像示例来动态学习执行任务，而不是明确编程。特别是深度学习（DL），其中使用多层人工神经网络松散地模拟人类大脑，在一些应用中以前所未有的性能水平占据主导地位。医学图像计算对自动化分析提出了独特的技术挑战，限制了该领域的成功。与纯粹的认知视觉应用（如自动驾驶）截然不同，医学图像数据的分析和解释涉及与复杂知识库相关的高度专业化过程。因此，尽管数据驱动的深度学习模型具有明显的能力和普遍的适用性，但目前的数据驱动的深度学习模型通常会导致表面学习，并伴有模型脆弱性和意外故障的惊人报告，对噪声的过度拟合和对有意义数据的欠拟合，新数据的泛化性差和低性能，以及所需训练数据库的不实用性和过高的标签成本。

Image computing is a key technology area with applications in numerous fields from engineering to medicine. A subset of artificial intelligence (AI) concerned with creating computer systems or machines capable of simulating human vision, the field still grapples with technical challenges arising from the vast variability in imaging data and disparate application-dependent analysis needs. A wave of connectionist machine learning has recently swept the field, in which systems dynamically learn' to perform a task through training on labelled image examples from huge databases rather than being explicitly programmed. In particular, deep learning (DL), in which multi-layer artificial neural nets that loosely model the human brain are used, has dominated with unprecedented performance levels reported in some applications. Medical image computing poses unique technical challenges to automated analysis that limited success in this area. Quite distinct from pure cognitive vision applications, such as autonomous driving, analysis and interpretation of medical image data involves highly specialized processes connected to complex knowledge bases. Hence, despite their apparent power and universal applicability, current data-driven DL models often result in superficial learning with alarming reports of model fragility and unexpected failures, overfitting to noise and underfitting to meaningful data, poor generalizability and low performance on new data, as well as impracticality of required training databases and prohibitive cost of labelling. This grant will advance DL methods towards deep reasoning' to enable reliable and safe AI for critical data analysis such as medical image computing. Specifically, the proposed technical methods will evolve data-driven DL models beyond visual perception (seeing) towards integrating logic and context based reasoning (understanding, extrapolating) in order to achieve common sense' solutions that observe physical realities and incorporate domain knowledge. My work will result in more generalizable, interpretable, efficient and reasoned learning models that can self-test and self-improve, ensuring real life utility and significantly improving the prospects of successful adoption in areas of strategic importance to Canadians. Automation of heterogenous complex medical image data analysis enabled by the proposed technical methods will have significant socio-economic impact including overcoming poor physician to patient ratio, reducing healthcare costs and delays, increasing accuracy and efficiency in diagnostics and care delivery, and enabling new findings and opportunities in health research. My research will train a diverse group of future experts and leaders in AI, a field where severe shortages in much needed talent exist. It will also help ensure Canada remains a world leader in AI research and a major contributor to the DL revolution'.