Abstract
Three-dimensional (3D) localization of point sources is an important task and an indispensable part of applications in different fields. The applications include 3D single-molecule localization microscopy (SMLM), and detecting and quantifying space debris in the vicinity of a space asset. Point spread function (PSF) engineering is a promising technique to solve this 3D localization problem, which encodes the depth information by a selected phase pattern.On top of existing model-based methods, the thesis focuses on high-resolution imaging and localization problem of 3D point source recovery from 2D data using supervised deep learning methods under the Poisson noise model. A specific technique invented by S. Prasad is considered, which obtains the depth position of a single point source from the amount of rotation of a single lobe point spread function. First, we formulate the physics model of the single lobe rotating PSF, and Cramer-Rao lower bound (CRLB) analysis is used to estimate the minimum variance of the dimensionless source coordinates. We use it as a criterion for adjusting the number of Fresnel zones L for rotating PSF and performance evaluation. Based on adjusted rotating PSF, we have applied two supervised convolution neural network (CNN)-based approaches. In the localization network (LocNet), the outputs and labels of the network are all in discretized forms. The distribution of point sources is discretized on a cubical lattice where the indices of nonzero entries represent the 3D locations of the point sources. Like the model-based methods, it needs a post-processing step to cluster the predictions from the network before getting the final locations of point sources. The DECODE type uses a different structure, with multiple channels, to predict various information of the input images, including the probability of the existence of a point source prediction, x, y, z coordinates, and flux values. In addition, we explore multiple ways to improve these results, for example, using a hard sample strategy in the training set preparation and regularization strategies in the loss function. The two approaches with adjustment strategy are applied to space debris detecting problem setting and super-resolution imaging using microscopy. To preserve the interpretability of the model-based method, we develop another new network architecture by applying the unrolling approach to the KL-NC optimization model.
Experimental results demonstrate the efficiency of the deep learning-based methods compared with the state-of-the-art performance, when sufficient training data is available. The 3D localization algorithms and strategies can also be readily applied to other applications, such as microscopy settings and other depth-related PSFs. In the future directions and perspective, we consider applying our approaches to the imaging and localization problems of fluorophores in microscopy settings for both simulated and real datasets. Besides, we also consider the interpretability of the algorithms using other tools, such as physics-informed neural networks and loss terms, and the multi-classification of point sources after finishing the stage of localization of point sources.
| Date of Award | 20 Jul 2023 |
|---|---|
| Original language | English |
| Awarding Institution |
|
| Supervisor | Hon Fu Raymond CHAN (Supervisor) |