Speckle imaging with blind source separation and total variation deconvolution

Randy Bartels, Olivier Pinaud, Maxine Varughese · 2024

Read Speckle imaging with blind source separation and total variation deconvolution by Randy Bartels, Olivier Pinaud, Maxine Varughese (2024) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2412-06755.txt.

Introduction

Seeing through murky tissue with light is hard because the light scatters again and again, washing out the clean beam that standard scans rely on. Randy Bartels, Olivier Pinaud, and Maxine Varughese propose a way to use the scattered light itself in a 2024 paper. Their setting is a microscope scan where each shot of laser light comes out fully speckled.

The method first untangles the mixed speckle into the fields scattered by each point of the sample. It then uses a known quirk, the memory effect of speckle, to judge how far apart those points sit. A sharpening step with an edge-keeping filter cleans the result. Computer tests show clear images at the detail limit, for both lone points and solid shapes, where older methods fail.

Catalog record

CREATORRandy Bartels, Olivier Pinaud, Maxine Varughese
YEAR2024
COLLECTIONPAPERS
FORMATTXT
FILE SIZE102 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2412-06755

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Speckle imaging with blind source separation and total variation deconvolution Randy Bartels, Olivier Pinaud, Maxine Varughese Abstract This work is concerned with optical imaging in strongly diffusive environments. We consider a typical setting in optical coherence tomography where a sample is probed by a collection of wavefields produced by a laser and propagating through a microscope. We operate in a scenario where the illuminations are in a speckle regime, namely fully randomized. This occurs when the light propagates deep in highly heterogeneous media. State-of-the-art coherent techniques are based on the ballistic part of the wavefield, that is the fraction of the wave that propagates freely and decays exponentially fast. In a speckle regime, the ballistic field is negligible compared to the scattered field, which precludes the use of coherent methods and different approaches are needed. We propose a strategy based on blind source separation and total variation deconvolution to obtain images with diffraction-limited resolution. The source separation allows us to isolate the fields diffused by the different scatterers to be imaged, while the deconvolution exploits the speckle memory effect to estimate the distance between these scatterers. Our method is validated with numerical simulations and is shown to be effective not only for imaging discrete scatterers, but also continuous objects. 1 Introduction We address in this work the optical imaging of objects in strongly diffusive environments. As light propagates in a heterogeneous medium, a fraction of the light behaves as if the medium were homogeneous, while the remaining part is scattered by the inhomogeneities. The first contribution is referred to as the ballistic part of the wave, and is exponentially damped as the light travels deeper into the medium. [ 20 , 48 ] Many state-of-the-art imaging techniques, such as confocal microscopy, [ 51 , 32 , 2 , 55 ] nonlinear microscopy, [ 10 , 26 , 36 , 25 , 22 , 62 ] optical coherence tomography (OCT), [ 31 , 24 , 14 ] SMART-OCT [ 8 ] and closed-loop accumulation of single scattering (CLASS) [ 33 , 63 ] , utilize the ballistic part only. The use of ballistic light with high numerical aperture (NA) objective lenses ensures that the images obtained from these techniques reach an imaging resolution that is close to the diffraction limit. In situations where the objects to be imaged are buried deep into the medium, however, the ballistic field is extremely small compared to the scattered field and it becomes a challenge to extract the ballistic wave from the total signal. This is achieved in SMART-OCT and CLASS by exploiting coherence effects that are present in the ballistic wave and not in the scattered wave. In biological tissues, ballistic methods perform well up to depths ranging from about tens of microns up to a few millimeters, depending on the imaging modality. [ 7 , 43 ] Beyond such depths, the ballistic light is unexploitable and a different approach is needed. Such imaging strategies must rely on the scattered field since it overwhelmingly dominates the measured ballistic signal. The crucial difficulty is then to leverage the information carried by the scattered part, [ 4 , 44 , 57 , 58 , 17 ] while this is fairly direct with the ballistic part once it is extracted. For example, wave-imaging techniques like ultrasound and seismic imaging [ 5 , 13 ] rely on methods that isolate the single-scattered field from the multiply scattered field to obtain high resolution optical image information in the first-Born approximation. It is also unclear what imaging resolution can be achieved in such a situation. While deeper imaging is possible with diffuse optical techniques, [ 34 , 3 ] imaging resolution is limited to a transverse spatial resolution on the order of the depth of imaging within the tissue, [ 39 ] and consequentially other models for image estimation are required to recover diffraction-limited imaging resolution performance. We propose in this work a strategy allowing us to image objects with diffraction-limited resolution in a configuration where the scattered field is in a speckle regime , namely the field can be modeled by a fully randomized wave. There is no ballistic part in such a regime. Our method is based on the framework of the reflection matrix , see e.g. [ 45 ] , and our imaging configuration follows the standard experimental setting of OCT: the sample to be imaged is placed at one conjugate plane of a microscope illuminated by a laser, and a camera placed at the other conjugate plane collects the reflected intensity. Interferometry techniques then allow one to recover the whole field and not just its intensity. [ 37 , 61 , 53 , 64 , 30 , 23 ] In addition, a time gating procedure makes it possible to select a particular depth in the sample, with an axial resolution inversely proportional to the bandwidth of the laser (the axial resolution is typically on the order of a few micrometers, with a laser central wavelength between 500 nanometers and 1 micrometer). [ 19 , 41 ] The measured field is therefore a combination of scattered fields emanating from the same slab at a given depth, ignoring most of the signal scattered in slabs with different travel times. In this strategy, the sample is probed by a series of random illuminations, and the associated measured fields are stored in a matrix \mathcal{R} (after flattening of the 2D fields into a vector), where a column of \mathcal{R} corresponds to a measured field for one illumination. Our method is inspired by that of [ 66 ] for fluorescence microscopy, and is conceptually quite similar to the decomposition of the time reversal operator (DORT) method [ 46 , 47 ] , even though the latter operates in a completely different regime where propagation is mostly ballistic. There are two steps. First, the fields emanating from different scatterers are separated, which gives us access to the Green’s functions associated with different point scatterers. This separation is the most difficult part. In the DORT method, a standard singular value decomposition (SVD) is enough, provided the scatterers are sufficiently far apart and have sufficiently different intensities. The scatterers separation assumption ensures the associated Green’s functions (actually simple plane waves in DORT) are nearly orthogonal, but also limits the applicability of the method. To achieve higher resolution and expand the class of objects that can be imaged, we need a better separation method. The problem falls into the category of blind source separation (BSS) problems and is solved using independent component analysis (ICA) techniques. Not only do ICA methods improve on the separation provided by the SVD, but they also separate fields when the SVD completely fails to, which occurs when e.g. the scatterers are too close to each other. In particular, there is a resonance effect, which we refer to as the speckle bond resonance in reference to the chemical bond resonance in the physics literature, [ 56 , 60 ] that prevents the SVD separation of two scatterers with comparable intensities even though the associated Green’s functions are nearly orthogonal. The ICA method we are using is immune to this resonance effect. The second step exploits these individual Green’s functions to locate

Read the complete text