Accelerated High-Resolution 3D Refractive Index Reconstruction Using Holographic Incoherent-Light-Source QPI and Deep Learning

Accelerated High-Resolution 3D Refractive Index Reconstruction Using Holographic Incoherent-Light-Source QPI and Deep Learning

Paper in proceedings (conference paper)

Quantitative Phase Imaging (QPI) offers 2D label-free live cell observations. To satisfy the burgeoning need for expanding 2D QPI for the ability of 3D refractive index distribution (RID) reconstruction, an approach known as Holographic Tomography (HT) has been developed. Our work proposes an alternative 3D RID reconstruction method, utilizing a z-stack of phase images obtained by the Holographic Incoherent-light-source QPI (hiQPI). Precise reconstruction of a 3D RID from z-stacked hiQPI phase images represents an inverse problem. This inverse problem can be solved either by physics-driven iterative algorithms or by a dataset-driven approach, i.e. by leveraging trained neural networks. Though the physics-driven algorithms are more established, the dataset-driven algorithms can significantly reduce the reconstruction time. We present a rapid dataset-driven 3D reconstruction algorithm utilizing a U-net-based convolutional neural network (CNN). The CNN is trained on a dataset comprising various simulated red blood cells (RBC) and corresponding simulated hiQPI z-stacks. RBCs are generated with varying parameters and refractive indices. Furthermore, to enlarge the dataset, the RBCs are augmented with affine transformations, including rotation, elastic deformation, Gaussian noise insertion, and blur. The hiQPI z-stacks are simulated employing the multi-slice beam propagation method in conjunction with the underlying hiQPI theory. This study demonstrates a novel alternative approach to the 3D RID reconstruction method, utilizing a z-stack of hiQPI phase images and a fast, high-quality reconstruction algorithm based on supervised deep learning. However, the results should be thoroughly validated against physics-based approaches in the future.

Quantitative Phase Imaging (QPI) offers 2D label-free live cell observations. To satisfy the burgeoning need for expanding 2D QPI for the ability of 3D refractive index distribution (RID) reconstruction, an approach known as Holographic Tomography (HT) has been developed. Our work proposes an alternative 3D RID reconstruction method, utilizing a z-stack of phase images obtained by the Holographic Incoherent-light-source QPI (hiQPI). Precise reconstruction of a 3D RID from z-stacked hiQPI phase images represents an inverse problem. This inverse problem can be solved either by physics-driven iterative algorithms or by a dataset-driven approach, i.e. by leveraging trained neural networks. Though the physics-driven algorithms are more established, the dataset-driven algorithms can significantly reduce the reconstruction time. We present a rapid dataset-driven 3D reconstruction algorithm utilizing a U-net-based convolutional neural network (CNN). The CNN is trained on a dataset comprising various simulated red blood cells (RBC) and corresponding simulated hiQPI z-stacks. RBCs are generated with varying parameters and refractive indices. Furthermore, to enlarge the dataset, the RBCs are augmented with affine transformations, including rotation, elastic deformation, Gaussian noise insertion, and blur. The hiQPI z-stacks are simulated employing the multi-slice beam propagation method in conjunction with the underlying hiQPI theory. This study demonstrates a novel alternative approach to the 3D RID reconstruction method, utilizing a z-stack of hiQPI phase images and a fast, high-quality reconstruction algorithm based on supervised deep learning. However, the results should be thoroughly validated against physics-based approaches in the future.

3D quantitative phase imaging | deep learning | digital holographic microscopy | holographic tomography | neural networks-based inverse problem solver | reconstruction algorithms for 3D QPI

3D quantitative phase imaging | deep learning | digital holographic microscopy | holographic tomography | neural networks-based inverse problem solver | reconstruction algorithms for 3D QPI

MICHÁLKOVÁ, I.; CHMELÍK, R.; ĎURIŠ, M.

2026

01.01.2025

SPIE

9781510684065

Progress in Biomedical Optics and Imaging Proceedings of SPIE

Proceedings of SPIE

United States of America

10

https://www.spiedigitallibrary.org/conference-proceedings-of-spie/13329/133290H/Accelerated-high-resolution-3D-refractive-index-reconstruction-using-holographic-incoherent/10.1117/12.3041129.full