Brains behind the eyes: Interpreting Medical Images

点击翻译按钮获取中文摘要

Modern medical imagers generate detailed images but they cannot quantify or `understand' these images. Computational models that can be used to interpret these medical images are needed to understand "normal" shape and function and identify deviations that may signal the onset of disease and to quantify the pace and temporal individual variability. This is the emerging field of "computational disease vision," with close parallels to computer vision, where intelligent algorithms are designed specifically for analyzing and recognizing the signals present in medical images. Developing the `brains' behind the `eyes', or the intelligent computer vision algorithms that can convert raw imaging data into measurements that can be used to detect the onset of disease, diagnose a disease with confidence, or to quantitatively monitor disease progression is the long-term goal of my research program.  The hypothesis is that machine learning models can extract and understand deeper relationships present in medical imaging data than are possible with human visual analysis. Hence, such computational vision algorithms are undoubtedly the future for clinical image interpretation. We propose two specific short-term goals towards the overarching long term quest for disease recognition from medical images. These are: (1) design of spatio-temporal multi-scale, multi-modal structured representations of normative and disease signatures, and (2) design of conventional, mixed- and deep-models for developing novel classifiers for disease recognition. Building holistic models for understanding human anatomy, shape and function, and thereby, models for recognition and quantification of disease present significant challenges. There is inherent variability across normative state in the population, and hence the signal of interest can be subtle and weak as compared to normal variability. The signals that mark the onset of disease exist and multiple scales, and are often described in relative terms of a configuration change, and hence require a semantic representation that can capture multiple levels of interaction across scales and locations within the anatomy. Often, a single modality may only capture part of the changes, for example, in the retina, the changes in retina layer geometry may be weaker but in addition to changes in the retina vasculature, provide a stronger discriminant signal to recognise the quantify disease. We propose to develop novel extensions to conventional shape models, deep-structured models that act on raw medical images directly, as well as mixed models combining the best of both conventional and deep-structured models for automated disease recognition from medical images in this proposal.