Shunan Wu
BME PhD Proposal Presentation

Date: 2024-05-23
Time: 05/23 8:00 PM-9:00 PM (EST, UTC-4) / 05/24 8:00 AM-9:00 AM (Beijing Time, UTC+8)
Location / Meeting Link: https://gatech.zoom.us/j/93242241953?pwd=MGt6T1k3SnFOWlFWRTJJRjlnUFFuQT09

Committee Members:
Xunbin Wei, PhD (Advisor); Shu Jia, PhD (Co-Advisor); Peng Xi, PhD; Changhui Li, PhD; Jiajia Luo, PhD


Title: Self-Supervised High Dynamic Range Imaging Reconstruction for Fluorescence Microscopy

Abstract:
Fluorescence microscopy is a powerful tool for studying biological systems, but its utility can be limited by the dynamic range of the imaging system. The intra-scene dynamic range (IDR) of many biological samples often exceeds the capabilities of traditional microscopy techniques, resulting in over- or under-exposed regions and compromising image quality. Additionally, when the fluorescent light intensity is very low or the exposure time is set to be very short for ultrafast capturing, the quantization precision will be limited, further reducing the effective dynamic range and leading to loss of detail in dim regions. High dynamic range (HDR) imaging techniques offer a solution by merging multiple exposures to capture a wider range of intensities, enabling the visualization of both bright and dim features within the same field of view. However, existing HDR methods can introduce ghosting artifacts when applied to dynamic live-cell imaging due to sample motion between exposures. To this end, this work aims to utilize a self-supervised deep learning approach called SelfHDR for HDR reconstruction in fluorescence microscopy. SelfHDR learns to merge multi-exposure images and handle sample motion without introducing the ground-truth HDR data into the model, making it more practical for live-cell imaging applications. In aim 1, we will develop the SelfHDR method for HDR reconstruction in fluorescence microscopy. Aim 2 focuses on establishing a training strategy for the SelfHDR model and validating its performance using simulated phantom datasets. In Aim 3, we will apply the trained SelfHDR model to simulated fluorescence microscopy datasets generated from open-source software. The successful completion of these aims is expected to provide a powerful new tool for HDR imaging in fluorescence microscopy, enabling researchers to overcome the limitations of low dynamic range and high IDR, and capture a wider range of biological phenomena with improved detail and accuracy. By eliminating the need for specialized hardware or complex calibration procedures, SelfHDR will make HDR imaging more accessible to researchers, ultimately leading to new insights and discoveries in the field of life sciences.