Applications of Computer Vision on the Biology of the Inner Ear

Abstract

The hair cells of the sensory epithelium of the peripheral auditory system are precisely tuned for separating frequency, power, and phase from acoustic signals. Enabled by modern imaging techniques, these cells are imaged over large spatial areas and commonly analyzed by hand. Manual analysis is the gold standard in terms of accuracy however, as datasets increase in size the time and labor costs become significant. With careful application, deep learning techniques have demonstrated the ability to capture high level of biological complexity and can aid in automating image analysis. These techniques have potential to save considerable time and drastically expand the amount of imaging data able to be analyzed, leading to less biased research outcomes and new insights. With this work, I have taken considerable steps in maturing the use of deep learning for image analysis of the biology of the inner ear, creating applications and tools across imaging modalities for varied analysis tasks. I have compiled the first open dataset for detection and classification of inner ear hair cells imaged with light microscopy. Using this dataset, I trained the first generalizable deep learning model for hair cell detection and demonstrated its accuracy and usefulness. I have created and validated a novel approach at three-dimensional instance segmentation, demonstrating its usefulness in assessing outer hair cell mitochondria health in normal and pathologic conditions. Finally, I have developed a framework easing the future development and validation of machine learning tools. With these works, I have created a foundation for deep learning image analysis in the auditory field, demonstrated best principles for training, evaluating, and validating deep learning performance, and have created new approaches which have the potential to impact all fields of biology.