Lensless holographic microscopy is a simple and compact imaging technique which comprises a coherent light source and an imaging sensor. A sample is usually placed as close as possible to the imaging sensor, on which an interference pattern between the reference and object electromagnetic wave is formed. The interference pattern is also commonly known as a hologram. The hologram encapsulates information about the amplitude and phase of the electromagnetic wave, which can be numerically reconstructed. In contrast to the conventional microscopy, the numerical reconstruction enables refocusing at desired focus planes effectively providing three-dimensional information about the scene. For a correct object reconstruction, a focus plane must be selected, which corresponds to the exact distance between the object and imaging sensor. The focus plane can be selected manually with visual inspection or automatically using an autofocusing method.
Autofocusing algorithms are evaluated on synthetic and experimentally acquired holograms. Synthetic holograms are commonly modelled with the angular spectrum method, which is the same numerical propagation method as used for the reconstruction of holograms. Unfortunately, this may conceal some errors, that stem from the presumptions of the propagation method. In addition, angular spectrum method cannot model holograms of truly three-dimensional objects. In contrast, experimentally acquired holograms can be affected by noise and artefacts resulting from mismatch between parameters of the experimental setup and parameters of the propagation model, such as source wavelength and pixel size. Furthermore, the exact distance between the object and the imaging sensor is not known.
In this work, we objective evaluate autofocusing algorithms on holograms modelled by Mie theory and T-matrix method. Both methods can model holograms of truly three-dimensional spherical objects. We implemented different autofocusing algorithms, which were objectively evaluated and compared according to the accuracy of the estimated focal plane and computational cost. Subsequently, we presented a proof-of-concept real-time implementation of the iterative autofocusing algorithm based on the PyOpenCL framework for execution on graphics processing units. Our best implementation resulted in an average absolute error of 1.61 µm, while the computational time for 1024×1024 holograms was 330 µs per iteration. This allows processing of approximately 20 holograms per second. The results provide promising starting point for use in real-time microfluidic applications for tracking and analysis of size and refractive index of microscopic particles.
Furthermore, we studied different deep learning architectures for predicting the focus distance, diameter and refractive index of microscopic particles. Models trained on synthetic holograms modelled with Mie theory allowed estimation of the focus distance with a mean absolute error of 1.60 µm. The mean relative errors for the estimation of diameter and refractive index were less than 0.5% and 0.05%, respectively. The estimation time for a hologram of size 150×150 was approximately 0.1 ms.
|
