ECE PhD Thesis Defense: Qianwan Yang

ECE PhD Thesis Defense: Qianwan Yang

Title: Advancing Computational Multi-aperture Fluorescence Microscopy via Model- and Learning-based Techniques

Presenter: Qianwan Yang

Advisor: Professor Lei Tian

Chair: TBA

Committee: Professor Vivek Goyal, Professor Jerome Mertz, Professor Janusz Konrad, Professor Lei Tian.

Google Scholar Link: https://scholar.google.com/citations?user=R9vLRnYAAAAJ&hl=en

Abstract: Traditional fluorescence microscopy faces inherent trade-offs among resolution, field of view (FOV), and system complexity. Computational Miniature Mesoscope (CM2) addresses this by placing a microlens array (MLA) before a single CMOS sensor: each microlens captures a sub-FOV, and intentional overlap maximizes usable FOV, encoding extended high-resolution scenes on a compact device. This optical simplicity shifts complexity to computation: the periodic MLA induces ill-conditioned multiplexed measurement, and the system exhibits highly non-local, spatially varying PSFs. This dissertation explicitly models that spatial variance (SV) process and solves the associated inverse problem, enabling uniform, wide FOV reconstructions from miniature hardware.

First, I established a low-rank LSV forward model. Sparse 3D PSFs are factorized via truncated singular value decomposition (TSVD) to yield a compact basis, and the corresponding coefficient maps are smoothly interpolated to capture field-dependent variation. Based on this forward model, I pose the inverse problem as a regularized least-squares objective and develop a model-based LSV ADMM solver that improves peripheral fidelity and overall reconstruction quality compared with a linear shift invariant (LSI) ADMM solver. The LSV ADMM solver serves as the reconstruction baseline, while the LSV forward model provides a high-fidelity simulator to generate diverse measurements for training and benchmarking learning-based methods. Building on this foundation, I introduce CM2Net, a multi-stage deep learning framework that combines view demixing, view synthesis, and refocusing-enhancement to perform single-shot 3D particle localization. Trained purely on simulated measurements, CM2Net generalizes to experiments, achieving ∼7 mm FOV over ∼0.8 mm depth with ∼6 µm lateral and ∼25 µm axial resolution, while delivering ∼1400× faster runtime and ∼19× lower peak memory than the iterative baseline. To overcome the locality bias and implicit shift-invariance of conventional CNNs and to generalize reconstruction to extended biological structures, I propose SVFourierNet, which performs global, low-rank deconvolution in the Fourier domain using learned filters that approximate the inverse of the LSV forward model. SVFourierNet attains uniform 7.8 µm resolution across a 6.5 mm FOV and shows interpretable agreement between its effective receptive fields and the physical PSFs. Finally, I introduce SV-CoDe, a scalable SV coordinate-conditioned deconvolution framework. By modulating compact convolutional features with coordinate-driven coefficients predicted by a lightweight MLP, SV-CoDe improves peripheral fidelity and decouples parameter and memory costs from FOV size, enabling SV reconstruction with scalable computation.

Together, the calibrated low-rank LSV model with an LSV–ADMM baseline, CM2Net, SV-FourierNet, and SV-CoDe unify physics and learning to resolve view multiplexing and spatially varying blur, enabling uniform, wide-FOV, high-throughput fluorescence imaging on compact hardware.