| Abstract |
Fluorescence microscopy is a pivotal technology in biological research, enabling high-resolution imaging of cellular and subcellular structures and processes. Modern imaging modalities, such as light-sheet microscopy, are able to acquire images with high spatial and temporal resolution over large areas or long durations, leading to the routine generation of terabyte-sized datasets. Diverse image processing algorithms are required to extract useful information from these image datasets, but their application is impeded by the vast data size. This “data bottleneck” often limits the throughput and scalability of fluorescence microscopy studies and leads to under-utilization of the information contained in the images. This thesis addresses the computational challenges of processing large image volumes by leveraging the Adaptive Particle Representation (APR). The APR is a multi-resolution image representation that optimally adapts the local sampling density to the image contents, thereby reducing redundancies in the representation of sparse images typical of fluorescence microscopy. We build upon the APR and its previously existing software to enable a wide range of image processing methods, from basic filtering operations to advanced deep-learning techniques, to leverage the data-efficient APR format for enhanced computational efficiency and greatly reduced memory requirements on parallel computer architectures. We demonstrate in real large-scale imaging applications that this can provide a comprehensive solution to the data bottleneck in fluorescence microscopy. First, we present data structures and algorithms that enable efficient and native processing on APR images using multi-core CPU and GPU parallelization. We define an adaptation of discrete convolutions, which are essential for many image processing tasks, and strategies for defining scale-adaptive filters that exploit the varying spatial scales of the APR. We demonstrate the viability of this approach in the task of image deconvolution on synthetic and real images, and quantify the computational efficiency of our implementation compared to pixel convolutions on evenly sampled data. Second, we demonstrate the practical utility of APR-based image processing in large-scale neurohistology applications. We present methods that enable complete APR-native pipelines, including automatic APR conversion, multi-tile stitching, segmentation, visualization, and atlas registration. Applied in imaging experiments on an entire mouse brain and a large section of human brain tissue, our pipeline exhibits substantially increased efficiency over established voxel methods, achieving 115-fold reduced storage requirements and 71 times accelerated processing, enabling acquisition-rate processing on a modest workstation CPU. Finally, we adapt convolutional neural networks (CNNs) to operate natively on the APR, resulting in APR-CNNs that leverage the APR data structures to reduce their memory footprint and computational burden. Given that computationally intensive CNNs have emerged as the state of the art across a wide range of image processing tasks, this adaptation greatly broadens the applicability of APR-based processing. We evaluate the performance of APR-CNNs in instance segmentation on real microscopy data, showing that they can achieve comparable segmentation accuracy to traditional pixel CNNs despite significantly reduced input data size. This thesis demonstrates the potential of APR-native image processing as a transformative tool for fluorescence microscopy. By developing and optimizing data structures, algorithms, and pipelines tailored for data-efficient APR images, this work paves the way toward comprehensive solutions to the data bottleneck in large-scale imaging through a combination of data reduction and parallel processing. In particular, the adaptation of deep-learning methods has broad applicability, potentially leading to more efficient and scalable bioimaging workflows that can accelerate and reduce the cost of scientific discovery. |