Giga-voxel multidimensional fluorescence imaging combining single-pixel detection and data fusion

Data fusion concept. From Fig.1 in the manuscript. Do you want a 4D reconstruction? Just take several 2D/3D objects and merge them in a clever way.

Some time ago I wrote a short post about using Data Fusion (DF) to perform some kind of Compressive Sensing (CS). We came with that idea when tackling a common problem in multidimensional imaging systems: the more you want to measure, the harder it gets. It is not only the fact that you need a system that is sensitive to many different physical parameters (wavelength, time, polarization, etc.), but also the point of having huge datasets that you need to record and store. If you try to measure a scene with high spatial resolution, in tens or hundreds of spectral channels, and with video frame rates (let’s say 30 or 60 frames per second), you generate gigabytes of data every second. This will burn through your hard drives in a moment, and if you want to send your data to a different lab/computer for analysis, you will need to wait ages for the transmission to end.

While there have been many techniques trying to solve these problems, there is not a really perfect solution (and, in my honest opinion, there cannot be a single solution that will solve all the problems that different systems will face) that allows you to obtain super high quality pictures in many different dimensions. You always need to live with some tradeoffs (for example, doing low spatial resolution but high frame rate, or gathering a low number of spectral bands with good image quality).

Data fusion results, from Fig.3 in the manuscript. Here you can see that the initial single-pixel datasets have low spatial resolution, but the DF results have high spatial resolution AND both spectral and temporal resolution.

However, there are cool ideas that can help a lot. In our last paper, we show how, by borrowing ideas from remote sensing and/or autonomous driving, you can obtain high resolution, multispectral, time-resolved images of fluorescent objects in a simple and effective manner. We use a single-pixel imaging system to build two single-pixel cameras: one that measures multispectral images, and another that obtains time-resolved measurements (in the ps range). Also, we use a conventional pixelated detector to obtain a high spatial resolution image (with no temporal or spectral resolution). The key point here is that we have multiple systems working in parallel, each one doing its best to obtain one specific dimension. For example, the single-pixel spectral camera obtains a 3D image (x,y,lambda) with a very good spectral resolution, but with very low spatial resolution. On the other hand, the pixelated detector acquires a high spatial resolution image, but neither spectral nor time resolved. After obtaining the different datasets, DF allows you to merge all the information in a final multidimensional image, where all the dimensions have been sampled at high resolution (so, our final 4D object has high spatial, temporal, and spectral resolution).

So, what about the compression? The cool thing here is that we only obtain three different datasets: the high resolution picture from the camera, and the two multispectral/time-resolved images from the single-pixel cameras. However, after the reconstruction we obtain a full 4D dataset that amounts for about 1 Gigavoxel. In the end, if you compare the number of voxels we measure versus the number of voxels we retrieve, we have a compression ratio higher than 99.9% (which is quite big if you ask me).

As a sample of the technique, we show the time-resolved fluorescence decay of a simple scene with three different fluorophores (each one of the letters you see on the following figures), where the species are excited and the fluorescence process takes place in less than 25 ns (woah!). You can see the live reconstruction here, and a short talk I made a while ago after the info of the paper, where you can see all the details about the system, the reconstruction algorithm, and so.

Giga-voxel multidimensional fluorescence imaging combining single-pixel detection and data fusion

F. Soldevila, A. J. M. Lenz, A. Ghezzi, A. Farina, C. D’Andrea, and E. Tajahuerce, on Optics Letters (and the arxiv version)

Abstract: Time-resolved fluorescence imaging is a key tool in biomedical applications, as it allows to non-invasively obtain functional and structural information. However, the big amount of collected data introduces challenges in both acquisition speed and processing needs. Here, we introduce a novel technique that allows to acquire a giga-voxel 4D hypercube in a fast manner while measuring only 0.03% of the dataset. The system combines two single-pixel cameras and a conventional 2D array detector working in parallel. Data fusion techniques are introduced to combine the individual 2D and 3D projections acquired by each sensor in the final high-resolution 4D hypercube, which can be used to identify different fluorophore species by their spectral and temporal signatures.

Data fusion as a way to perform compressive sensing

Some time ago I started working on some kind of data fusion problem where we have access to several imaging systems working in parallel, each one gathering a different multidimensional dataset with mixed spectral, temporal, and/or spatial resolutions. The idea is to perform 4D imaging at high spectral, temporal, and spatial resolutions using some single-pixel/multi-pixel detectors, where each detector is specialized on measuring one dimension in high detail while subsampling the others. Using ideas from regularization/compressive sensing, the goal is to merge all the information we acquire individually in a way that makes sense, and while doing so, achieve very high compression ratios.

Looking for similar approaches, I stumbled with a series of papers from people doing remote sensing that basically do the same thing. While the idea is fundamentally the same, their fusion model relies on a Bayesian approach, which is something I have never seen before, and seems quite interesting. They try to estimate a 4D object that maximizes the coherence between the data they acquire (the low-resolution 2D/3D projections) and their estimation. This is quite close to what we usually do in compressive sensing experiments on imaging, but with a minimization based on a sparsity prior.

An Integrated Framework for the Spatio–Temporal–Spectral Fusion of Remote Sensing Images

By Huanfeng Shen ; Xiangchao Meng , and Liangpei Zhang, at IEEE Transactions on Geoscience and Remote Sensing

Abstract:

Remote sensing satellite sensors feature a tradeoff between the spatial, temporal, and spectral resolutions. In this paper, we propose an integrated framework for the spatio-temporal-spectral fusion of remote sensing images. There are two main advantages of the proposed integrated fusion framework: it can accomplish different kinds of fusion tasks, such as multiview spatial fusion, spatio-spectral fusion, and spatio-temporal fusion, based on a single unified model, and it can achieve the integrated fusion of multisource observations to obtain high spatio-temporal-spectral resolution images, without limitations on the number of remote sensing sensors. The proposed integrated fusion framework was comprehensively tested and verified in a variety of image fusion experiments. In the experiments, a number of different remote sensing satellites were utilized, including IKONOS, the Enhanced Thematic Mapper Plus (ETM+), the Moderate Resolution Imaging Spectroradiometer (MODIS), the Hyperspectral Digital Imagery Collection Experiment (HYDICE), and Systeme Pour l’ Observation de la Terre-5 (SPOT-5). The experimental results confirm the effectiveness of the proposed method.

Operation principle of the technique. Multiple images with different resolutions are combined to obtain a high-resolution multidimensional reconstruction of the data. Extracted from Fig.1 of the manuscript.

As sensors evolve, the amount of information we can gather grows at an alarming rate: we have gone from just hundreds or thousands of pixels to millions in just a few decades. Also, now we gather hundreds of spectral channels at hundreds or thousands of frames per second. This implies that we usually suffer from bottlenecks in the acquisition and storage of multidimensional datasets. Using approaches like this, one can make it possible to obtain very good object estimations while measuring very little data in a fast way, which is always a must.

Bonus: Of course, they have also been doing the same but using Machine Learning ideas:

Spatial–Spectral Fusion by Combining Deep Learning and Variational Model

By Huanfeng Shen, Menghui Jiang, Jie Li, Qiangqiang Yuan, Yanchong Wei, and Liangpei Zhang, at IEEE Transactions on Geoscience and Remote Sensing

Inverse Scattering via Transmission Matrices: Broadband Illumination and Fast Phase Retrieval Algorithms

Interesting paper by people at Rice and Northwestern universities about different phase retrieval algorithms for measuring transmission matrices without using interferometric techniques. The thing with interferometers is that they provide you lots of cool stuff (high sensibility, phase information, etc.), but also involve quite a lot of technical problems that you do not want to face every day in the lab: they are so sensitive that it is a pain in the ass to calibrate and measure without vibrations messing everything up.

Using only intensity measurements (provided by a common sensor such as a CCD) and algorithmic approaches can provide the phase information, but at a computational cost that sometimes makes things not very useful. There is more info about all of this (for the coherent illumination case) in the Rice webpage (including a dataset and an implementation of some of the codes).

Inverse Scattering via Transmission Matrices: Broadband Illumination and Fast Phase Retrieval Algorithms

by Sharma, M. et al., at IEEE Transactions on Computational Imaging 

Abstract:

When a narrowband coherent wavefront passes through or reflects off of a scattering medium, the input and output relationship of the incident field is linear and so can be described by a transmission matrix (TM). If the TM for a given scattering medium is known, one can computationally “invert” the scattering process and image through the medium. In this work, we investigate the effect of broadband illumination, i.e., what happens when the wavefront is only partially coherent? Can one still measure a TM and “invert” the scattering? To accomplish this task, we measure TMs using the double phase retrieval technique, a method which uses phase retrieval algorithms to avoid difficult-to-capture interferometric measurements. Generally, using the double phase retrieval method re- quires performing massive amounts of computation. We alleviate this burden by developing a fast, GPU-accelerated algorithm, prVAMP, which lets us reconstruct 256^2×64^2 TMs in under five hours.

After reconstructing several TMs using this method, we find that, as expected, reducing the coherence of the illumination significantly restricts our ability to invert the scattering process. Moreover, we find that past a certain bandwidth an incoherent, intensity-based scattering model better describes the scattering process and is easier to invert.

Single-pixel imaging with sampling distributed over simplex vertices

Last week I posted a recently uploaded paper on using positive-only patterns in a single-pixel imaging system.

Today I just found another implementation looking for the same objective. This time the authors (from University of Warsaw, leaded by Rafał Kotyński) introduce the idea of simplexes, or how any point in some N-dimensional space can be located using only positive coordinates if you choose the correct coordinate system. Cool concept!

Fig.1 extracted from “Single-pixel imaging with sampling distributed over simplex vertices,”
Krzysztof M. Czajkowski, Anna Pastuszczak, and Rafał Kotyński, Opt. Lett. 44, 1241-1244 (2019)

Single-pixel imaging with sampling distributed over simplex vertices

by Krzysztof M. Czajkowski et al., on Optics Letters

Abstract:

We propose a method of reduction of experimental noise in single-pixel imaging by expressing the subsets of sampling patterns as linear combinations of vertices of a multidimensional regular simplex. This method also may be directly extended to complementary sampling. The modified measurement matrix contains nonnegative elements with patterns that may be directly displayed on intensity spatial light modulators. The measurement becomes theoretically independent of the ambient illumination, and in practice becomes more robust to the varying conditions of the experiment. We show how the optimal dimension of the simplex depends on the level of measurement noise. We present experimental results of single-pixel imaging using binarized sampling and real-time reconstruction with the Fourier domain regularized inversion method.

Handling negative patterns for fast single-pixel lifetime imaging

A group of researchers working in France and USA, leaded by N. Ducros, has uploaded an interesting paper this week.

When doing single-pixel imaging, one of the most important aspects you need to take into account is the kind of structured patters (functions) you are going to use. This is quite relevant because it is greatly connected with the speed you are going to achieve (as the number of total measurements needed for obtaining good images strongly depends on the set of functions you choose). Usually, the go-to solution for single-pixel cameras is to either choose random functions, or a set (family) of orthogonal functions (Fourier, DCT, Hadamard, etc.).

The problem with random functions is that they are not orthogonal (it is very hard to distinguish between two different random functions, all of them are similar), so you usually need to project a high number of them (which is time consuming). Orthogonal functions that belong to a basis are a better choice, because you can send the full basis to get “perfect” quality (i.e., without losing information due to undersampling). However, usually these functions have positive and negative values, which is something you cannot directly implement in lots of Spatial Light Modulators (for example, in Digital Micromirror Devices). If you want to implement these patterns, there are multiple workarounds. The most common one is to implement two closely-related patterns sequentially in the SLM to generate one function. This solves the negative-positive problem, but increases the time it takes to obtain an image in a factor two.

What Lorente-Mur et al. show in this paper is a method to generate a new family of positive-only patterns, derived from the original positive-negative family. This makes it possible to obtain images with a reduced number of measurements when compared to the dual or splitting approach I mentioned earlier, but still with high quality. Nice way to tackle one of the most limiting factors of single-pixel architectures.

Working principle visualization of the generalization method to measure with positive-only patterns in single-pixel imaging setups. Figure extracted from Lorente-Mur et al., ”
Handling negative patterns for fast single-pixel lifetime imaging,” at https://hal.archives-ouvertes.fr/hal-02017598

Handling negative patterns for fast single-pixel lifetime imaging

by Antonio Lorente Mur et al., at https://hal.archives-ouvertes.fr/hal-02017598

Abstract:

Pattern generalization was proposed recently as an avenue to increase the acquisition speed of single-pixel imaging setups. This approach consists of designing some positive patterns that reproduce the target patterns with negative values through linear combinations. This avoids the typical burden of acquiring the positive and negative parts of each of the target patterns, which doubles the acquisition time. In this study, we consider the generalization of the Daubechies wavelet patterns and compare images reconstructed using our approach and using the regular splitting approach. Overall, the reduction in the number of illumination patterns should facilitate the implementation of compressive hyperspectral lifetime imaging for fluorescence-guided surgery.

Realization of hybrid compressive imaging strategies

Recently I have been reading a lot about Compressive Sensing strategies. One of the things we always want when we work in a single-pixel architecture is to project the lowest possible number of masks, because the projecting process is the longest in all the acquisition procedure (and it gets longer and longer when you increase the spatial resolution of your images).

In the past, several strategies haven been implemented to reduce that number of projections. From going fully random to partially scan a basis at random and at the low frequency region, each approach presents some benefits and more or less speed gain.

In this work by the group of K.F. Kelly, they explored a different approach. Instead of chosing one measurement basis and design a sensing strategy (picking random elements, or centering around the low frequency part of the basis, or a mix), they create a measurement basis by merging different functions. They call it hybrid patterns. The basic idea is to chose a low number of patterns which work well for recovering low frequency content of natural images, and also some other patterns which are good to recover high frequency content. The novel thing here is that they do not require the patterns to belong to the same orthogonal basis, thus being able to carefully design its measurement basis. This provides very good quality results with a low number of projections.

Another thing I liked a lot was the Principal Component Analysis (PCA) part of the paper. Basically, they gathered a collection of natural images and they generated an orthogonal basis by using PCA. This leads me to think of PCA as a way of obtaining orthogonal bases where objects have their sparsest representation (maybe I am wrong about that).

Realization of hybrid compressive imaging strategies,

Y.Li et al, at Journal of the Optical Society of America A

(featured image exctracted from Fig.2 of the manuscript)

Abstract:

The tendency of natural scenes to cluster around low frequencies is not only useful in image compression, it also can prove advantageous in novel infrared and hyperspectral image acquisition. In this paper, we exploit this signal model with two approaches to enhance the quality of compressive imaging as implemented in a single-pixel compressive camera and compare these results against purely random acquisition. We combine projection patterns that can efficiently extract the model-based information with subsequent random projections to form the hybrid pattern sets. With the first approach, we generate low-frequency patterns via a direct transform. As an alternative, we also used principal component analysis of an image library to identify the low-frequency components. We present the first (to the best of our knowledge) experimental validation of this hybrid signal model on real data. For both methods, we acquire comparable quality of reconstructions while acquiring only half the number of measurements needed by traditional random sequences. The optimal combination of hybrid patterns and the effects of noise on image reconstruction are also discussed.

Fig3_Kelly
Really nice to see that PCA gives something very similar to DCT functions. This means that compressing images with DCT is really a good choice.