TR&D3

University Of Michigan At Ann Arbor, Ann Arbor MI

Investigators

Abstract

D Cai and K Najarian will serve as Project Lead and co-Lead, respectively, for TRD3. Live fluorescence imaging is challenged by high levels of autofluorescence background and potential for photobleaching and phototoxicity. Faster frame rates are needed to capture subtle cellular behaviors, but an increase in imaging speed may result in a reduction of signal. TRD3 will develop properly trained machine-leaning models that can effectively denoise images captured with at least 5 times higher frame rates. Importantly, a faster frame rate can significantly mitigate motion artifacts, which is critical to interpret images collected in vivo and track biological processes. Environmental perturbations and movement can blur individual images, and substantially degrade the usefulness of video streams. Microendoscopy is especially sensitive to motion because the image field-of-view is typically only a few hundred microns in dimensions. Image segmentation remains one of the most challenging tasks in biomedical image processing, in particular, at the cellular level. Live imaging normally results in uneven intensity distributions, noise patterns, weak edges, and frequently incomplete features. Rapid advances in the deep learning field has dramatically improved the accuracy of feature extraction from in vivo fluorescence images using complex illumination and noise conditions, and have increased in popularity with the development of massive parallel computing architectures. These algorithms can enable precise image recognition based on high-dimensional hierarchical image features. Deep learning algorithms can be trained using validated clinical datasets to provide real time digital readouts. In vivo cellular activity visualized using microendoscopy will be correlated with their molecular properties via postmortem tissue imaging. This approach allows for the dynamic behaviors of live cells visualized in vivo, and compared with mRNA and protein expression to identify upregulated signaling pathways. Valuable insights into cellular function and behaviors can be used to elucidate physiological mechanisms. Morphological profiles that are versatile descriptors of biological systems can be generated and used to predict in vitro and in vivo drug effects. New drugs can be developed, and their potential impact can be better understood. CP/SPs are currently performing live imaging and molecular profiling separately using different sets of animals. Scientific conclusions are drawn from statistical averages calculated from different cohorts. The number of experimental animals needed to draw significant conclusions can be substantially reduced, and uncertainties introduced by individual differences in statistical analysis can be minimized. TRD3 will support CP/SPs to advance their scientific discoveries with higher accuracy and confidence. Synergy: TRD3 will use in vivo optical sections collected by TRD2 to develop machine learning models for automated digital pathology, and will quantify tumor dimensions to assess therapeutic efficacy using photoacoustic images collected with microsystems sensors developed in TRD1.