Computational and machine learning methods have been essential in analyzing biomedical images in recent decades. These methods have greatly advanced our understanding of medical and biological structures at all scales, from the macromolecular level to the anatomical systems level. However, these tools remain niche in their use in scientific applications due to their high supervisory requirements, as many challenges still exist that computer scientists have only begun to address.

 

This dissertation proposes three major improvements to two tasks that concern distinct scales of biomedical imaging. In the microscopy scale, specifically, for cryo-electron microscopy -- the state of the art method for obtaining near-atomic resolution images of biological micromolecules -- this work presents a deep learning method for automating grid screening, the process of imaging samples at different resolutions while deciding where to image next. Even though our work focuses on the lowest magnification level, the techniques can be applied to higher resolutions given ample data, resulting in a prototype to fully automating the grid screening pipeline. If similar techniques are implemented and refined within microscope software, it can significantly increase imaging throughput and decrease the cost of imaging macromolecular structures, precipitously accelerating advancement in foundational biomedical research.

 

For the other task of particle based statistical shape modeling -- the process of modeling shape variation in populations of anatomies through automatically placed surface particles -- this work adds a way to model arbitrary regions of interest with minimal overhead, provides the ability to provide full adaptivity of particle spread to underlying surfaces in both coerce and fine-grained detail, and explores the use of better optimization losses to facilitate building effective models for biomedical research. These contributions vastly expand the capabilities of studying arbitrary surfaces shapes and accelerate the efficiency of running shape modeling pipelines by reducing particle budget without losing fidelity.

 

Ultimately, my work explores and expands improvements to computational and machine learning approaches used for cutting edge biomedical discovery, looking to increase the tools availability and democratize these modern techniques to new users and applications.