Abstract

Traditional computer vision and AI are often applied to tasks that humans can perform equally well or even better. In these cases, human experts typically generate the ground truth values on which algorithms are trained. The advantages of machine learning in such applications include cost savings, reduced bias, and the broad availability of expert knowledge. Typical tasks range from self-driving cars to localizing tumors in large datasets. However, machine learning can also be used for tasks that are challenging or impractical for human experts alone, or where AI is necessary to enable human experts to achieve specific goals. In medical imaging, there is a trend toward increasingly advanced imaging methods, which produce datasets that are difficult for human observers to visualize and interpret. Beyond the sheer size of these datasets, a key factor contributing to this complexity is the higher-dimensional nature of the data, which can be thought of as images with (often many) more than three channels per pixel. Examples of these advanced imaging types include hyperspectral imaging, quantitative MRI mass spectrometry imaging, and Raman imaging. Each of these fields relies on machine learning as an essential, though currently insufficient, component to integrate new imaging modalities into medical practice. In this lecture, I will outline current applications, challenges, and unmet needs of machine learning when implementing advanced imaging modalities that go beyond human vision and interpretability. https://meet.google.com/kns-mqbz-jkq