EE & MedE Distinguished Speaker Seminar, Richard J. Cote
The microscope is perhaps the greatest scientific tool ever invented and is certainly the tool that has the longest history of use in a form that Antonie van Leeuwenhoek would recognize today. The basic parameters include the objective lens, stage, eyepiece and light from a single source focused on the sample; even an electron microscope has these features. About a decade ago Changhuei Yang and colleagues developed an entirely new way to illuminate the sample via multiple light sources at different angles, with each individual light source producing an image in the objective at a slightly different focal plane and which, after mathematical transformation of the subsequent digital images produces an "all-in-focus" image. It was through this work and work on Yang's Optifluidic microscope that my group developed a collaboration with him that has lasted over two decades, producing numerous impactful papers and multiple grant awards.
Our collaboration has recently taken an interesting and productive turn. While the microscope has proven to be a robust tool with no foreseeable end to its usefulness, it does have one major limitation: the capacity of humans to grasp the nearly endless nuances of the images it produces. Two advances have taken place that have allowed us to explore what might be in microscopic images that humans cannot quantify or even perceive, that is, extremely high resolution digital images and the development of AI platforms to interpret visual images. AI has been used to try to replicate what a human pathologist can do, for example, diagnose and grade cancers, and has shown promise in certain applications to provide more reproducible results than humans. However, we wanted to task AI with something deeper, to do what pathologists can't do, despite nearly 200 years trying; to predict the future.
In collaboration with a team from Wash U and Caltech, we have analyzed a series of diagnostic digital images from early stage lung cancer to determine if the images held information that could answer the most important question in cancer; will it metastasize and kill the patient. Our early results have been stunning, demonstrating amazing precision in this prediction. However, it also revealed a problem that has plagued the digital AI field; algorithms trained on one set of cases are not easily generalized to another set of cases. In this talk we will explore AI's ability to predict outcome, our studies on how AI learns and why the algorithms fail to generalize, and how we are exploring ways to eliminate the variability inherent in digital images and thus develop more useful AI learning systems. This is not a talk on AI platforms, but is rather an exploration of how AI can be better applied in digital microscopy.
Hosted by Changhuei Yang