A module and method for identifying cellular aggregates using digital holographic microscopy

Abstract

This disclosure describes a computing module and method for identifying cellular aggregates in real-time based on digital holographic images that are captured using digital holographic microscopy (DHM) techniques. In particular, the computing module is configured to use artificial intelligence models and data analysis techniques to analyze the captured DHM images. The analyzed DHM images are then classified and quantified to identify cellular patterns which are then statistically analyzed to reveal patterns linking the cellular patterns to associated diseases.

Description

A MODULE AND METHOD FOR IDENTIFYING CELLULAR AGGREGATES USING DIGITAL HOLOGRAPHIC MICROSCOPYCROSS REFERENCE TO RELATED APPLICATION

[0001] This application claims the benefit of priority to Singapore patent application no. 10202401892U which was filed on 26 June 2024, the contents of which are hereby incorporated by reference in its entirety for all purposes.TECHNICAL FIELD

[0002] This application relates to a computing module and method for identifying cellular aggregates in real-time based on digital holographic images that are captured using digital holographic microscopy (DHM) techniques. In particular, the computing module is configured to use artificial intelligence models and data analysis techniques to analyze DHM images. The analyzed DHM images are then classified and quantified to identify cellular patterns which are then statistically analyzed to reveal patterns linking the cellular patterns to associated diseases.BACKGROUND

[0003] Blood tests in the field of hematology constitute a well-established and highly automated diagnostic modality and remain among the most commonly performed laboratory assessment. In addition to standard hematology analyzers, a variety of digital imaging systems for automated blood smear analysis have been developed. These systems typically combine automated microscopy with digital cameras to capture images of individual blood cells from peripheral blood smears, often employing May-Grunwald Giemsa staining. The acquired images are then subjected to computer-aided pre-classification based on morphological features.

[0004] However, existing methods lack sufficient resolution to detect subtle changes in the biophysical morphology of cells and cellular aggregation. This limitation reduces their effectiveness in predicting disease progression and early pathological changes. Furthermore, these systems are unable to detect cellular aggregates without the use of labeling techniques. Since cell aggregates are biologically unstable, the delay introduced by transport and processing times further complicates the accurate quantification of these cells, making timely clinical assessment challenging.

[0005] A technique widely utilized in research and clinical settings for analyzing cell surface and intracellular molecule expressions is fluorescence flow cytometry (FCM). FCM enables simultaneous multi -parameter analysis of single cells and is commonly used to characterize heterogeneous cell populations, assess purity of isolated subpopulations, analyze cell size and volume, and perform cell sorting through fluorescence-activated cell sorting (FACS). A flow cytometry test is performed by measuring the fluorescent intensity produced by fluorescently labeled antibodies specific to proteins on or in cells or ligands that bind to specific cell associated molecules. However, FCM is labor-intensive as it involves making single-cell suspensions from cell culture or tissue samples, that are incubated with unlabeled or fluorophore-labeled antibodies to the actual analysis of the samples by trained technicians. While FCM may detect abnormalities in white blood cell differentials, it often fails to reliably classify abnormal or immature cells, necessitating manual morphological review through blood smear microscopy.

[0006] Another technique that is adopted by those skilled in the art is the imaging flow cytometry (IFC) method which combines the capabilities of traditional flow cytometry imaging with the analysis of labeled cellular populations, enabling concurrent acquisition of morphological and biochemical data from thousands of cells per second IFC supports detailed single-cell analysis and facilitates identification of rare cell populations and cellular heterogeneity within cell groups. Additionally, IFC allows for the investigation of cellular interactions and dynamics, capturing both spatial and temporal characteristics of cellular processes. Further development of automated image analysis algorithms allows for the classification of cells into distinct categories based on morphological and phenotypic criteria while advances in machine learning, and artificial intelligence have further enhanced IFC’s analytical capabilities, improving classification accuracy and predictive performance in complex cellular analyses.

[0007] Yet another technique that is adopted by those skilled in the art is fluorescence microscopy (FM) and this is a powerful tool for visualizing targeted cellular and molecular components using fluorescent probes, providing high-contrast, detailed images of cellular structures. FM also supports real-time imaging, allowing the dynamic biological processes in the living cells to be monitored. In addition to FM techniques, plate readers, or microplate photometers, including platforms for enzyme-linked immunosorbent assays (ELISA), cellbased assays, and nucleic acid quantification, may be employed to facilitate the detection andmeasurement of absorbance, fluorescence, luminescence, and other optical signals across multiple samples in a high-throughput manner.

[0008] Another technique that is commonly adopted by those skilled in the art are automated hematology analyzers (HA) and this tool offers rapid, precise, and high-throughput blood tests, including complete blood counts (CBC) that measures key parameters such as red blood cell (RBC) count, white blood cell (WBC) count, platelet (PLT) count, hemoglobin concentration, and hematocrit levels. These analyzers typically process small blood samples using advanced techniques such as flow cytometry, electrical impedance, or fluorescent flow cytometry to differentiate and count various blood cell types and analyze their morphological characteristics, such as their size, shape and hemoglobin content.

[0009] Despite advancements in defect detection, existing solutions still require some type or form of cell labelling and chemical pre-treatment before disease classification may be accurately performed. Hence, those skilled in the art are constantly looking for a label-free phase imaging-based flow cytometry platform that integrates digital holographic imaging technology with machine learning to deliver real-time functional diagnostics at the point-of- care.SUMMARY

[0010] In one aspect, the present application discloses a computing module for identifying cellular aggregates in real-time based on digital holographic microscopy (DHM). The disclosed computing module comprises a processing unit, and a non-transitory media readable by the processing unit. The media stores instructions that when executed by the processing unit causes the processing unit to acquire digital holographic images of a sample captured using a digital holographic microscopy module, wherein the digital holographic images comprise spatial- domain interference patterns generated by an object beam and at least one reference beam. The instructions also cause the processing unit to generate, for each of the digital holographic images, a corresponding amplitude image and phase image using an off-axis hologram reconstruction module, the off-axis hologram reconstruction module comprising: at least one Fourier imager head model, each Fourier imager head model corresponding to a respective reference beam and configured to perform frequency-domain filtering on a frequency -domain representation of the digital holographic image to obtain an isolated object beam componentassociated with the corresponding reference beam; and a complex valued network model configured to convert the isolated object beam components into corresponding amplitude and phase image components. The instructions also cause the processing unit to detect, using an object detection model, cellular structures within each of the digital holographic images based on the corresponding amplitude and phase image components, and perform, using an aggregation analysis model, cellular aggregate analysis on the detected cellular structures within each of the digital holographic images to identify and classify cellular aggregates.

[0011] In embodiments of this one aspect, each Fourier imager head model may comprise a set of trained two-dimensional matrices configured to perform frequency-domain filtering through matrix multiplications and element-wise multiplication.

[0012] In embodiments of this one aspect, the set of trained two-dimensional matrices may comprise shift matricesconfigured to translate an origin of the frequency domainrepresentation to a center region of the frequency-domain representation, additional matrices, M and M* configured to crop and shift a target frequency region to the center of the frequency domain-representation, shift-back matrices, M2and M2configured to translate shifted regions of the frequency spectrum back to their original frequency domain-representations, and a trained spectral mask matrix Mmaskconfigured to perform pixel-wise spectral filtering by attenuating high-frequency components in the cropped and shifted frequency domainrepresentation.

[0013] In embodiments of this another aspect, a method for identifying cellular aggregates in real-time based on digital holographic microscopy (DHM) using a computing module is disclosed. The disclosed method comprises the steps of acquiring digital holographic images of a sample captured using a digital holographic microscopy module, wherein the digital holographic images comprise spatial-domain interference patterns generated by an object beam and at least one reference beam; generating, for each of the digital holographic images, a corresponding amplitude image and phase image using an off-axis hologram reconstruction module, the off-axis hologram reconstruction module comprising: at least one Fourier imager head model, each Fourier imager head model corresponding to a respective reference beam and configured to perform frequency-domain filtering on a frequency-domain representation of the digital holographic image to obtain an isolated object beam component associated with thecorresponding reference beam; and a complex valued network model configured to convert the isolated object beam components into corresponding amplitude and phase image components. The method also comprises the steps of detecting, using an object detection model, cellular structures within each of the digital holographic images based on the corresponding amplitude and phase image components, and performing, using an aggregation analysis model, cellular aggregate analysis on the detected cellular structures within each of the digital holographic images to identify and classify cellular aggregatesBRIEF DESCRIPTION OF THE DRAWINGS

[0014] Various embodiments of the present disclosure are described below with reference to the following drawings:Figure 1 illustrates a block diagram of components or modules that are provided within a computing module for identifying cellular aggregates in real-time based on a digital holographic microscopy (DHM) setup in accordance with embodiments of the present disclosure;Figure 2 illustrates orientations of reference beams and object beams in a DHM setup in accordance with embodiments of the present disclosure;Figure 3 illustrates an experimental DHM setup in accordance with embodiments of the present disclosure,Figure 4a illustrates a block diagram of components or modules that are provided within an off-axis hologram reconstruction module in accordance with embodiments of the present disclosure;Figure 4b illustrates initial and fine-tuned weights for the spectral mask matrix Mmaskin the frequency domain for the respective Fourier image head (FIH ) models;Figure 5 illustrates frequency spectra that have been Fourier filtered with a circular filter and phase corrected using a hologram reconstruction procedure;Figure 6 illustrates an exemplary raw hologram image and its corresponding spectrum after the raw hologram image has been Fourier transformed and frequency shifted in accordance with embodiments of the disclosure;Figure 7a illustrates a comparison between a target phase / amplitude and a predicted phase / amplitude as inferred by the off-axis hologram reconstruction module when the module has a medium phase prediction error;Figure 7b illustrates a comparison between a target phase / amplitude and a predicted phase / amplitude as inferred by the off-axis hologram reconstruction module under the conditions of the highest phase prediction error;Figure 8 illustrates a full data processing workflow of a YOLO-Graph model for analyzing cellular aggregates based on data obtained the DHM setup shown in Figure 3 in accordance with embodiments of the present disclosure;Figure 9 illustrates a block diagram of a processing system for performing embodiments of the present disclosure;Figure 10 illustrates a flow chart showing the process for identifying cellular aggregates in realtime based on digital holographic microscopy (DHM) in accordance with embodiments of the present disclosure;Figure 11 illustrates a schematic representation of a DHM setup in accordance with embodiments of the present disclosure;Figure 12a illustrates an image obtained based on digital holographic microscopy and the outcome from the cellular detection and aggregation analysis process in accordance with embodiments of the present disclosure when no cellular aggregates were detected;Figure 12b illustrates an image obtained based on digital holographic microscopy and the outcome from the cellular detection and aggregation analysis process in accordance with embodiments of the present disclosure when WBC-WBC aggregates are detected, andFigure 12c illustrates an image obtained based on digital holographic microscopy and the outcome from the cellular detection and aggregation analysis process in accordance with embodiments of the present disclosure when WBC-PLT aggregates are detected.DETAILED DESCRIPTION

[0015] The following detailed description is made with reference to the accompanying drawings, showing details and embodiments of the present disclosure for the purposes of illustration. Features that are described in the context of an embodiment may correspondingly be applicable to the same or similar features in the other embodiments, even if not explicitly described in these other embodiments. Additions and / or combinations and / or alternatives as described for a feature in the context of an embodiment may correspondingly be applicable to the same or similar feature in the other embodiments.

[0016] In the context of various embodiments, the articles “a”, “an” and “the” as used with regard to a feature or element include a reference to one or more of the features or elements.

[0017] In the context of various embodiments, the term “about” or “approximately” as applied to a numeric value encompasses the exact value and a reasonable variance as generally understood in the relevant technical field, e.g., within 10% of the specified value.

[0018] As used herein, the term “and / or” includes any and all combinations of one or more of the associated listed items.

[0019] As used herein, “comprising” means including, but not limited to, whatever follows the word “comprising”. Thus, use of the term “comprising” indicates that the listed elements are required or mandatory, but that other elements are optional and may or may not be present.

[0020] As used herein, “consisting of’ means including, and limited to, whatever follows the phrase “consisting of’. Thus, use of the phrase “consisting of’ indicates that the listed elements are required or mandatory, and that no other elements may be present.

[0021] One skilled in the art will recognize that certain functional units in this description have been labelled as modules throughout the specification. The person skilled in the art will also recognize that a module may be implemented as circuits, logic chips or any sort of discrete component. Still further, one skilled in the art will also recognize that a module may be implemented in software which may then be executed by a variety of processor architectures In embodiments of the disclosure, a module may also comprise computer instructions or executable code that may instruct a computer processor to carry out a sequence of events based on instructions received. The choice of the implementation of the modules is left as a design choice for a person skilled in the art and does not limit the scope of the claimed subject matter in any way.

[0022] Digital Holographic Microscopy (DHM) is an imaging modality that enables label- free visualization of cells by detecting variations in the refractive index within cellular structures, thereby capturing digital hologram images. These holograms may subsequently be processed to reconstruct both amplitude and phase images, producing high-contrast visualizations of transparent biological samples. This facilitates the identification of various leukocyte subtypes and potential pathogens, including leukemia cells and microaggregates associated with infectious or cardiovascular conditions. By eliminating the need for celllabeling, Digital Holographic Microscopy (DHM) enables the rapid extraction of clinically relevant information compared to conventional imaging techniques Nevertheless, the efficiency of DHM remains constrained by image processing and analysis workflows, which currently constitute major bottlenecks in its widespread clinical adoption. While the present disclosure is described with reference to DHM, it will be appreciated by those skilled in the art that the described computational framework is broadly applicable to other types of quantitative phase imaging (QPI) techniques, such as transport-of-intensity imaging, diffraction phase microscopy, and spatial light interference microscopy, without departing from the scope of the disclosure and that the images captured by these respective QPI techniques may be treated or processed in the same manner as the digital hologram images described in this disclosure.

[0023] A block diagram of components or modules that are provided within a computing module for identifying cellular aggregates in real-time based on a digital holographic microscopy (DHM) setup in accordance with embodiments of the present disclosure is illustrated in Figure 1. Specifically, computing module 100 integrates flow cytometry, digital holographic microscopy, and Al-powered machine learning to identify innovative biomarkers in cellular diagnostics suitable for bedside point-of-care (POC) applications.

[0024] Computing module 100 offers a distinct advantage by enabling label -free cell analysis under conditions closely resembling the cells' natural state, without requiring specific sample preparation. This approach removes constraints on blood volume, making it particularly suitable for automated processing and the detection of rare cellular biomarkers. Not limited to a single diagnostic application, the label-free flow cytometry technology implemented by computing module 100 is applicable across a broad range of diagnostic fields beyond infectious diseases and is capable of supporting differential diagnostics in both routine and patientspecific clinical settings.

[0025] As illustrated in Figure 1, computing module 100 is configured to acquire digital holographic images 102 of a sample captured using a digital holographic microscopy module (not shown). Digital holographic images 102 may comprise, but are not limited to, spatial- domain interference patterns generated by an object beam and one or more reference beams.

[0026] Computing module 100 comprises off-axis hologram reconstruction module 104 that is configured to receive digital holographic images 102. Computing module 100 further comprises object detection module 106 having an input communicatively coupled to off-axis hologram reconstruction module 104 and an output communicatively coupled to aggregation analysis module 107. Outputs generated by aggregation analysis module 107 may be optionally stored in data storage 108. In embodiments of the disclosure, data storage 108 may comprise, but is not limited to, non-transitory computer-readable media such as hard disk drives, solid- state drives, flash memory devices, optical storage media, or cloud-based storage systems configured to store digital holographic images, processed amplitude and phase images, detected cellular structures, aggregation analysis results, and corresponding diagnostic reports.

[0027] In particular, off-axis hologram reconstruction module 104 is configured to generate amplitude and phase images for each of digital holographic images 102. Object detection module 106 then utilizes the amplitude and phase image components generated by off-axis hologram reconstruction module 104 to detect cellular structures within the corresponding digital holographic images. Aggregation analysis module 107 is then configured to perform cellular aggregate analysis on the detected cellular structures to identify and classify cellular aggregates within each of digital holographic images 102.

[0028] As further illustrated in Figure 1, computing module 100 may be optionally communicatively coupled with large language model (LLM) module 110, which is configured to receive user instructions 111 and provide results 112 based on the user’s requests. In embodiments of the disclosure, LLM module 1 10 may provide operational control instructions to computing module 100, such as initiating hologram acquisition, triggering object detection and aggregation analysis, and generating automated diagnostic reports based on the stored results in data storage 108.

[0029] In embodiments of the disclosure, digital holographic images 102 may be acquired and provided to computing module 100 through a streamlined sample preparation and DHM imaging process. A minimal sample volume, as low as 10 pl, may be used for the analysis, conserving biological material and enabling continuous monitoring of disease progression and treatment response. It should be noted that this sample preparation process eliminates the needfor fluorophores or stains, thereby simplifying handling procedures and reducing the risk of sample contamination.

[0030] To preserve cell integrity and prevent unintended cell activation or hemolysis, the blood sample may be introduced into a polymethyl methacrylate (PMMA) microfluidics cartridge compatible with DHM. The microfluidics cartridge may incorporate multiple sheath inlets, e.g. four sheath inlets, to establish a well-defined depth of field, enabling viscoelastic focusing of cells at a single imaging plane. This configuration ensures consistent image quality across various leukocyte types and minimizes optical distortion due to interactions with the channel walls of the microfluidics cartridge.

[0031] Digital holographic images 102 may then be captured using a DHM system, which provides label-free imaging of cells by detecting refractive index variations within the cellular structures The holograms captured by the DHM system represent both inter-cellular and intracellular information and these are processed by computing module 100 to facilitate the identification of different leukocyte subtypes and potential pathogenic indicators, such as leukemia cells or microaggregates associated with infectious or cardiovascular diseases.

[0032] The DHM system that was utilized for acquiring digital holographic images 102 employs an off-axis interferometric configuration comprising an object beam and one or more reference beams Figure 2 illustrates reference beams and an object beam of the DHM system in accordance with embodiments of the present disclosure. As illustrated, object beam 0 is oriented perpendicular to charge-coupled device (CCD) sensor 214 and propagates towards CCD sensor 214 along the direction of the z-axis.

[0033] The DHM system also includes reference beam Ry(depicted as dashed lines) and reference beam Rx(depicted as solid lines) to facilitate spatial separation of interference components in the Fourier domain. As can be seen from the isometric view 202, right side view 204 and top view 206 of the DHM system, reference beam Rxis tilted along the x-axis while the reference beam Ryis tilted along the y-axis. Reference beam Rxmay be defined as reference beam Rx(x,y) and reference beam Rymay be defined as reference beam / y( , y) with CCD sensor 214 being configured to record an intensity distribution IH(x, y) whichrepresents the interference pattern of object beam 0, which may be defined as O(x,y) Hence, the intensity distribution / H(x,y) recorded by CCD sensor 214 may be defined as:...equation (1) where * denotes the complex conjugate of the equation.

[0034] When the intensity distribution is recorded with a sample, it may be defined as IH sand the corresponding object beam may be defined as: ...equation (2)where XO sis the amplitude of the object beam 0(x, y), <p0is the object beam’s phase, and A< / >(x,y) is the phase shift caused by the refractive index variations of the sample. When the intensity distribution IHis recorded in the absence of a sample, it may be denoted as IH b, and the corresponding object beam may be defined as:...equation (3)As can be seen from equation (3), the corresponding object beam equation Ob(x,y) does not include a sample-induced phase shift.

[0035] The sample’s signal is separated by filtering out the reference wave components. As a result, the reference beam terms remain unchanged in both the intensity distribution with a sample, IH sand the intensity distribution without a sample, IH b. The respective reference beams are then defined as: ...equation (4) ...equation (5) ...equation (6)where A is defined as the wavelength of the laser, andand ARyare the amplitudes of the references waves which are uniformed over CCD sensor 214. The phase shift — fcx sin O is caused by the tilted angle 0 between the phase plane of reference beam Rxand CCD sensor214. Similarly, the phase shift — ky sin 77 is caused by the tilted angle 17 between the phase plane of reference beam Ryand CCD sensor 214. When equations (2), (4) and (5) are substituted into equation (1) for an intensity distribution of a sample IH s, this results in:... equation (7)

[0036] When these terms undergo a Fourier transformation process (denoted as T), the terms become:y yff(y)

[0037] Equation (9) shows that F(ORf ) are shifted by sin j along the x-axis from the origin. Similarly, all the other terms are shifted along different directions in the Fourier spectrum except for the zero-order term ?(RXRX+ RyRf + 00*) as defined in equation (8). For completeness, Figure 2 illustrates the propagation of reference beams Rxand Ry, and object beam 0 from an isometric view 202, from a right-side view 204 (in the ZY plane) and from a top view 206 (in the ZX plane) with object beam O propagating along the z-axis.

[0038] Figure 3 illustrates a DHM system that may be used to capture digital holographic images of biological samples. The DHM system includes super-luminescent diode (SLED) 302 which acts as a partially coherent light source for the DHM system. The optical beam emitted by SLED 302 is then directed through pinhole 304, which focuses the optical beam ontomicrofluidic chip 306 which holds the biological samples that are to be analysed. After the optical beam has passed through the microfluidic chip, the optical beam is then directed through high-magnification 40x objective lens 308, which collects the transmitted optical beam and forms an enlarged image of the sample region. The enlarged image then passes through diffraction grating 310, which introduces angular separation between the object beam and reference beam components, enabling their spatial separation at the detector for off-axis holographic recording. Finally, the separated beams pass through optical wedge 312 that introduce controlled angular deviations, producing the required off-axis configuration for hologram formation. Object beam 314 and reference beam 316 then interfere at CMOS sensor 318, causing the generation of a holographic interference pattern of the sample at CMOS sensor 318. This holographic interference pattern is captured as a digital hologram directly at CMOS sensor 318, where the optical interference is converted into a digital signal for further processing in accordance with embodiments of the disclosure. It is useful to note that this spatial-domain interference pattern may contain both amplitude and phase information, which can later be reconstructed using various computational algorithms to recover quantitative phase images of the sample.

[0039] In embodiments of the disclosure, the spatial separation of the Fourier spectrum may be utilized to extract the real image components (either ORX* or ORy) directly from the digital hologram. Frequency shifts introduced by the tilt angles of the reference beams, specificallymay be corrected by repositioning the corresponding spectral components back to the center of the Fourier spectrum in accordance with embodiments of the disclosure. The detailed procedure is described in a later section with reference to Figure 5

[0040] Following the frequency-domain operations and application of the inverse Fourier transformation, the real images may then be recovered in the spatial domain as follows:... equation (15)... equation (16) where(Fs(x,y) and ' / 7 / ,(x, y) are derived from lH sand IH b, respectively, ^(x, y) could be further demodulated by dividing it by (x, y), resulting in:... equation (17)where A< >(x,y) provides the 3D information of the sample, and AO sx, y) / Ao b(x,y) denotes the relative intensity changes induced by the sample.

[0041] A block diagram of components or modules that are provided within an off-axis hologram reconstruction module (e.g., module 104 as shown in Figure 1) in accordance with embodiments of the present disclosure is illustrated in Figure 4a.

[0042] As illustrated in Figure 4a, off-axis hologram reconstruction module 104 is configured to receive a digital hologram 401 of a sample and a corresponding background digital hologram 402, both of which may be acquired using a DHM system. Fourier transformation module 403 is then used to convert digital holograms 401 and 402 into corresponding frequency-domain representations. Following the Fourier transformation process, the frequency-domain data is processed by the one or more Fourier Image Head (FIH) models 404. In embodiments of the disclosure, each of FIH models 404 may be configured to perform frequency-domain filtering to remove undesired signals in the frequency-domain data and to correct frequency shifts caused by the tilted angles between the reference beams and the object beam.

[0043] The filtered frequency -domain data is then provided to inverse Fourier transformation module 406, where an inverse Fourier transformation is performed to convert the processed data back to the spatial domain, yielding complex-valued object wave representations. These complex -valued object wave representations are then provided to complex valued network (CVN) model 408, which separates the complex data into distinct phase 409 and amplitude 410 image components. It should be noted that when multiple reference beams are employed in the DHM setup, an isolated object beam component may be derived from each reference beam. CVN model 408 may further be configured to merge the multiple isolated object beam components into a single combined object component for final image reconstruction.

[0044] In embodiments of the disclosure, OR* separation and shifting in the Fourier frequency domain may be performed by two matrix multiplications (denoted as X ) and one element-wise matrix multiplication (denoted as O) in FIH models 404 and may be defined as:where T is the Fourier transformation, F is the output of a FIH model - the corresponding Fourier frequency of O / ? that has been separated and shifted. The matrices MltMr, and Mmaskcomprise trainable two-dimensional matrices with each set of matrices Mt, Mr, and Mmaskbeing grouped as one Fourier Imager Head model that is associated with a reference beam, e.g. reference beam Rxor RyBased on the example illustrated in Figure 2, this implies that two FIH models are required to distill ORXand ORy, respectively.

[0045] In embodiments of the disclosure, H may be kept at high resolution with an original size of 1536 x 2048 pixels. The matrices Mt, Mr, and Mmaskwhich are two-dimensional trainable matrices, may have matrix dimensions of 384 x 1536, 2048 x 512, and 384 * 512, respectively, while F is the output of FIH that is the corresponding Fourier frequency of OR*.

[0046] In embodiments of the disclosure, the initialization of the coefficients in matrices Mt, Mr, and Mmaskmay be based on a predefined shape based filter such as, but not limited to, the circular shape filter, e.g., a spatial filter in the Fourier domain that allows frequencies within a certain radius from the center, with the mathematical details set out in Table 1 below. The initialization of matrices Mt, Mr, and Mmaskmay be defined as:where Utare random weights generated by Pytorch to eliminate zeroes in the matrices which are undesirable for parameter training.

[0047] As set out in equation (18), each FIH involves three trainable matrices: Mt, Mr, and Mmask, which are multiplied with the Fourier-transformed hologram FOn) .Spectral mask matrix Mmaskfunctions as a filter in the frequency domain, with the corresponding initial and fine-tuned weights shown in Figure 4b. Specifically, plot 420 shows the initial weight of spectral mask matrix Mmaskwhen it is implemented as a predefined shape based filter having a circular shape filter in the frequency domain. This circular low-pass filter acts as a binary or tapered mask to attenuate high-frequency components during the early stages of training. Plots 422 and 424 show the fine-tuned versions of spectral mask matrix Mmaskfor the two FIH models after they have been trained. Unlike the initial weights shown in Figure 4a, these fine-tuned masks display non-circular, asymmetric patterns, reflecting adaptive learning from data as they are configured to selectively retain or suppress frequency components beyond the initial low-pass design.

[0048] As can be seen from plots 422 and 424, the fine-tuned spectral mask matrix Mmaskof both FIH models are distinct from the initial circular shape shown in plot 420 and exhibit interesting patterns not reported before. The initial weights of matricesand Mrconsist only of 0 or 1 and are solely responsible for cropping and shifting the Fourier spectrum. After being fine-tuned as trainable parameters, matrices Mtand Mrpartially contribute to frequency filtration and improved model performance.

[0049] During training, the weights of matrices Mi, Mr, and Mmask, are not strictly restricted within the range of [0, 1], but most of the fine-tuned values (approximately 99%) fall within this range. Table A below sets out the detailed statistics of the weights in matrices Mj, Mr, and Mmas.

[0050] Tn embodiments of the disclosure, during the training of matrices Mt, Mr, and Mmask(as part of the end-to-end supervised learning process), each matrix may be initialized based on a predefined shape based filter such as, but not limited to, a circular shape filter (as described in equations (19)(a) - (21)(a)) and optimized using gradient-based backpropagation. The entire model is trained end-to-end using a supervised learning approach, where the loss is computed between the reconstructed image and a ground truth reference image. Through iterative updates via an optimizer (e.g., Adam), the matrices adaptively learn frequency-space transformations that enhance reconstruction performance, effectively evolving from interpretable initial filters to task-specific frequency operators.Table ATable 1

[0051] Table 1 describes a sequential frequency-domain processing workflow that may be implemented by each of FIH models 404. Specifically, Table 1 illustrates that each step of the spatial filtering process in Figure 5 may be converted into a matrix multiplication step, which could be further simplified as:M(= Mt Mi Mi ...equation 19(b)...equation 20(b)Mma.sk — Mi MmaskMr... equation 21(b) where Mt, Mrand Mmaskmay be determined manually based on equations 19(b)-21(b) and then loaded into the neural network as initial weights. In embodiments of the disclosure, equations 19(a)-21 (a) may also be used to manually determine the values of M(, Mrand Mmaskand this is left as a design choice to one skilled in the art. For completeness, it should be noted that equations 19(b)-21(b) are similar to that of equations 19(a)-21(a) with the random weights U„ being removed for ease of computation.

[0052] With reference to the example illustrated in Figure 5, Figure 5A illustrates an original digital holographic image as acquired by a DHM system in accordance with embodiments of the disclosure. Figure 5B illustrates the corresponding frequency spectrum of the digital holographic image in Figure 5A after this image has been processed by a 2D Fourier transformation process, where the zero-frequency components are located at the corners of the spectrum.

[0053] In the first step shown in Table 1, the origin of the frequency domain is converted from the image corners to the center. This step corresponds to the transition from Figure 5B to 5C. In embodiments of the disclosure, this step may be performed by applying left and right transformation matrices,and M?, which effectively shift the zero-frequency (DC) component to the center of the spectrum. As illustrated in Figure 5C, the target frequency component corresponding to the object beam is now visually indicated within box 502, centred for easier extraction.

[0054] The second step corresponds to the transition from Figure 5C to 5D. In this step, the target frequency region containing the object beam component is extracted, and non-target components are masked out. The extracted region is then repositioned to the center of the frequency spectrum to correct for phase shifts caused by the tilted reference beams. This is illustrated as box 504 in Figure 5D. In embodiments of the disclosure, this is accomplished by applying additional transfonnation matrices, M / and M), which perform both cropping and frequency shifting operations.

[0055] In the third step, shown as the transition from Figure 5D to 5E, high-frequency noise is further suppressed. This is achieved by applying an element-wise spectral mask 506 using a trainable matrix Mmask. The mask matrix selectively attenuates undesired frequency components while preserving the information relevant to the object beam, thereby enhancing the quality of the reconstructed images

[0056] Finally, the last step corresponds to the transition from Figure 5E to 5F, where the frequency-domain origin is shifted from the center back to the corners in preparation for the inverse Fourier transformation process. This is performed by incorporating additional transformation matrices M2and M2to the earlier matrix multiplications. The resulting frequency-domain representation, as shown in Figure 5F, is then subjected to an inverse 2D Fourier transform to produce the final phase and amplitude images. The reconstructed phase image is illustrated in Figure 5G, and the amplitude image is shown in Figure 5H.

[0057] In embodiments of the disclosure, the sequential frequency-domain processing workflow that may be implemented by each of FIH models 404 as shown in Table 1 may also be described by the following workflow.

[0058] Shift matrices M;°are configured to translate the origin of the frequency domain to the center of the frequency spectrum. This functionality corresponds to the operation step 5B 5C in Table 1. The subsequent step of cropping and shifting a target frequency region to the center of the frequency spectrum (step 5C 5D in Table I) is performed using additional matrices. The frequency components are then returned to their original positions using shift-back matrices M2and M2as described in step 5E —> 5F in Table 1. A trained spectral mask matrix Mmaskis applied during filtering to attenuate high-frequency components in the spectrum, as shown in step 5D — ► 5E in Table 1. Table 1 further supports these configurations by confirming that the matrices Mt, Mr, and Mmaskare initialized with binary values in the range [0, 1], reflecting their intended roles in shifting, cropping, and filtering. After training, the matrix values ofand Mmaskremain largely within [—0.1, 1], indicating that the matrices retain their structural function throughout the optimization process. It should be noted that matrices M;° , M°, M, ,, and M2are not trained separately and that they are used to generate the initial value ofMr, and Mmask, as shown in equations 19(b)-21 (b). Once Mt, Mr, and Mmaskhave been initialized, they are then directlytrained during model training. If matrices M°, M , M*, M , M , and M were to be trained, this would be mathematically equivalent to trainingand Mmask, and this will introduce unnecessary computation complexity.

[0059] Figure 6 illustrates an exemplary visualization of a raw hologram image and its corresponding amplitude-contrast Fourier spectrum following Fourier transformation and frequency shifting, in accordance with embodiments of the disclosure Specifically, image 602 shows the raw hologram as directly captured by a DHM system. Image 604 illustrates the resulting Fourier spectrum after application of the mathematical operations defined by equations (8) through (14). Plot 610 highlights the zero-order term of the spectrum after the raw hologram shown in image 602 has been Fourier transformed according to equation (8) above.

[0060] The other plots illustrate the spatially shifted frequency components corresponding to various interference terms. In particular, plot 611 shows the spatially shifted frequency component after it has been shifted according to equation (10), plot 612 shows the spatially shifted frequency component after it has been shifted according to equation (12), plot 613 represents the spatially shifted frequency component after it has been shifted according to equation (13), plot 614 shows the spatially shifted frequency component after it has been shifted according to equation (14), plot 615 represents the spatially shifted frequency component after it has been shifted according to equation (1 1 ) and plot 616 shows the spatially shifted frequency component after it has been shifted according to equation (9). In summary, these shifted frequency components represent the spatial separation of object-reference beam interference terms in the Fourier domain.

[0061] Returning to Figure 4a, the outputs from FIH models 404 comprising ORXand ORy are further demodulated to remove Rxand Ryvia element-wise division to generate two object waves O having minor differences. These two object waves O are then processed using an inverse Fourier transformation process back to the spatial domain by inverse Fourier transform module 406. These complex-valued object wave representations O are then provided to CVN model 408 to be processed into distinct unwrapped phase 409 and amplitude 410 image components.

[0062] In this embodiment of the disclosure, as it was previously disclosed that the object wave 0 can be derived using either of the reference beams Rxor Ry, two distinct object waves will be generated by FIH models 404 and subsequently provided as inputs to CVN model 408.

[0063] In embodiments of the disclosure, CVN model 408 may be configured to perform three primary processing tasks: (1) converting each complex-valued object wave O into corresponding phase and amplitude representations, (2) merging multiple object waves O into a single unified object wave, and (3) performing phase unwrapping on the merged phase information to produce unwrapped phase image component 409. In some embodiments, the initial conversion of the object waves O from their complex form to phase and amplitude representations is performed without the use of trainable parameters, while the merging of multiple object waves O and phase unwrapping may be performed using trained convolutional neural network (CNN) layers.

[0064] In alternative embodiments, the order of the three primary processing tasks within CVN model 408 may be rearranged, wherein the multiple object waves 0 are first merged while still in their complex-valued form using the complex-valued trained CNN layers, followed by conversion into the phase and amplitude representations. The phase unwrapping task is then performed on these phase and amplitude representations using the trained CNN layers. However, experimental evaluation has indicated no significant difference in reconstruction performance between the two strategies. Accordingly, the first approach where complex-to-phase-amplitude conversion precedes merging may be preferred, as it eliminates the need for complex-valued computation within the trained CNN layers, thereby reducing computational overhead. This approach also facilitates acceleration of the processing pipeline using advanced inference optimization techniques such as NVIDIA TensorRT.

[0065] As described in the previous sections, it can be seen that FIH model 404 replaces complex iterative procedures with a series of efficient matrix computations. This design enables the processing of digital holograms in a single forward pass without requiring iterative refinement. By avoiding iterative methods, the system eliminates two primary drawbacks: (1) reconstruction time is no longer linearly dependent on the number of iterations, thus significantly accelerating the overall processing speed, and (2) the system achieves consistentperformance across different holograms without requiring adaptive iteration counts, facilitating efficient batch processing and parallel computation.

[0066] Additionally, the architecture of FIH model 404 minimizes computational overhead by utilizing only a limited number of CNN layers. This is in contrast to conventional models that often employ complex U-Net architectures. In embodiments of the disclosure, once trained, these trained CNN layers may be further optimized through a re-parameterization technique known as the RepVGG method, which simplifies the network structure and reduces inference complexity. The exact details of this RepVGG method are omitted for brevity and it is described in the publication titled “ RepVGG: Making VFF-style CorrvNets Great Again" by the authors Xiaohan Ding et. al. Furthermore, all computations for FIH model 404 may performed directly on the GPU, fully leveraging parallel processing capabilities and eliminating the need for frequent data transfers between the GPU and CPU.

[0067] In embodiments of the disclosure, the matrix coefficients used in the matrices in FIH models 404 and the CNN layers within the CVN model 408 may be jointly trained using an end-to-end supervised learning framework using ground truth data comprising a target phase reconstruction of the hologram. Specifically, the training data or ground truth data may comprise of, but are not limited to, digital holograms recorded in the presence of biological samples and corresponding background holograms acquired without the presence of a sample. In scenarios where the biological samples are sparsely distributed and randomly positioned, as illustrated in Figures 7a and 7b, the background hologram may alternatively be approximated by averaging multiple sample holograms, with negligible impact on reconstruction accuracy In embodiments of the disclosure, the target images or ground truth data may comprise the phase (< >) and amplitude (A) images reconstructed from sample hologram images using methods where Fourier transformation and spatial filtering in the frequency domain are applied. For example, target images may be generated using Ovizio, which is an API software provided by the DHM manufacturer.

[0068] Specifically, Figure 7a illustrates a comparison between the target phase and the predicted phase as inferred by off-axis hologram reconstruction module 104 (as shown in Figure 1) for a typical sample exhibiting medium phase prediction error while Figure 7b illustrates a situation corresponding to the highest phase prediction error observed during avalidation step. The Structural Similarity Index Measure (SSIM) values, displayed alongside each example, quantitatively assess the reconstruction quality, with values approaching 1.00 indicating high similarity between predicted and ground truth images. As shown, even under high error conditions in phase prediction (SSIM = 0.85), the amplitude predictions maintain near-perfect accuracy (SSIM = 1.00). In embodiments of the disclosure, the ground truth phase and amplitude images for supervised training may be generated using validated external reconstruction methods, such as a OsOne software provided by the DHM system manufacturer.

[0069] To address the inherent data imbalance between background pixels and sample regions, a weighted LI loss function is employed during training. This loss function places greater emphasis on regions containing meaningful sample information, ensuring that the trained matrices and CNN layers focus on accurately reconstructing critical phase and amplitude features. The loss function is defined as:. . . equation (22) where A and < / > are the target amplitude and phase respectively, while A and (p are the predicted amplitude and phase respectively, j and k are defined as the pixel coordinates, w and h are the width and height of the image, respectively, i is used to indicate the Ithsample and n is the batch size. W is the pixel weight function and may be defined as:...equation (23) where clamp(x, a, b) = max (a, min(x, b)) , grad(_(pj:k) is defined as the gradient of the target phase image which may be derived by the Sobel filter followed by a Gaussian blur.

[0070] Based on the above, it can be seen that the trainable matrices within FIH models 404, including the various left transformation matrices (Mj), right transformation matrices (Mr), and spectral mask matrix (Mmask), may be optimized using the supervised learning approach integrated within the end-to-end training framework During the training process, thesematrices receive frequency-domain representations of digital holographic images as input, following a 2D Fourier transformation. The matrices are applied through sequential matrix multiplications and element-wise operations as described above (and in Table 1) to perform frequency-domain filtering, to isolate the object beam components associated with each reference beam. The filtered outputs are then evaluated based on their contribution to the accuracy of the final reconstructed phase and amplitude images produced by the CVN model 408. To guide the optimization of these matrices, the weighted LI loss function may be employed, emphasizing regions with high phase gradients and critical morphological features.

[0071] Similarly, the CNN layers within CVN model 408 may be trained (as part of the end- to-end supervised learning process) using a supervised learning approach based on a dataset comprising digital holographic images and corresponding ground truth amplitude and phase images. These ground truth images are generated through established hologram reconstruction methods, such as those provided by commercial software tools, which produce highly accurate amplitude and phase representations for training purposes. During training, the CNN layers receive complex-valued object waves or their corresponding amplitude and phase components as input and are optimized to minimize the difference between its predicted outputs and the ground truth labels. The weighted LI loss function as described in equation (22) may be similarly employed during CNN training to prioritize important structural details and enhance the accuracy of the reconstructed images.

[0072] In addition, the phase unwrapping CNN layers of CVN model 408 may be trained (as part of the end-to-end supervised learning process) to reduce phase discontinuities and eliminate wrapping artifacts, using synthetic and experimentally acquired datasets that include known wrapped and unwrapped phase patterns. The CNN layers responsible for merging multiple object beams and performing phase unwrapping may be trained using standard optimization algorithms, such as the Adam or stochastic gradient descent (SGD) optimizers, with learning rate scheduling to ensure stable convergence. In embodiments of the disclosure, the networks in this disclosure may be trained using the Adam optimizer with a constant learning rate optimized by grid search. Training of each network was halted when the validation loss did not improve over 200 successive epochs, and the model with the lowest validation loss was selected for further testing To further enhance computational efficiency and support realtime diagnostic applications, the trained CNN models may be subjected to inference acceleration using platforms such as NVIDIA TensorRT.

[0073] With reference to Figure 1, it can be seen that once the amplitude and phase image components corresponding to each of the digital holographic images have been generated by off-axis hologram reconstruction module 104, object detection module 106 then utilizes this information to detect cellular structures within each of these digital holographic images.

[0074] In embodiments of the disclosure, object detection module 106 may comprise, but is not limited to, a YOLOv8x-p2 model. The YOLOv8x-p2 model improves the detection of small objects, such as platelets, compared to the original Y0L0v8x model, which is the largest model configuration in the YOLOv8 series. This enhanced capability is achieved by incorporating additional feature pyramid levels into its architecture, allowing the model to better capture fine-grained details essential for identifying small cellular structures within phase images. At the core of its architecture is an advanced neural network backbone specifically designed for effective feature extraction from input phase images. This backbone is coupled with a detection head responsible for accurately predicting object bounding boxes and associated class probabilities.

[0075] In addition, the YOLOv8x-p2 model employs an anchor-free detection mechanism, eliminating the reliance on pre-defined anchor boxes traditionally used for object proposal generation. Instead, the model directly predicts bounding box locations from learned image features, thereby simplifying the model architecture and reducing computational complexity. The model also implements a multi-scale prediction strategy, enabling it to detect objects across various spatial resolutions within the network. This multi-scale approach ensures that both small and large cellular structures can be accurately detected within a single inference pass, making the model particularly well-suited for analysing biological samples with varying cell sizes. The detailed workings of the YOLOv8x-p2 model are omitted in this disclosure for brevity.

[0076] Once the cellular structures have been detected, the detected cellular structures are then provided to aggregation analysis module 107. Aggregation analysis module 107 is configured to perform cellular aggregate analysis on the detected cellular structures within each of the digital holographic images to identify and classify cellular aggregates within the digital holographic images.

[0077] In embodiments of the disclosure, aggregation analysis module 107 employs a graph model to perform aggregate analysis of the cellular structures. Following object detection using the YOLOv8x-p2 model, aggregation analysis module 107 receives the detected cellular structures which may include the cell bounding boxes and associated class predictions. The spatial relationships between these detected cells are then analysed by computing the centroid distances between bounding boxes to assess the presence of cellular aggregates and any relevant aggregation patterns.

[0078] The aggregation analysis process proceeds through five distinct computational steps. First, a graph is initialized, where each detected cell is represented as a node. Second, nodes are assigned class labels corresponding to their detected cell types. Third, the module evaluates all pairwise combinations of detected cells, identifying potential aggregation candidates. Fourth, edges are added between pairs of nodes only if both belong to target cell classes, for example leukocytes or platelets, and if the spatial distance between their centroids falls within a predefined threshold. Finally, in the fifth step, the graph is pruned by discarding all nodes representing non-target cells, such as erythrocytes, and removing edges that do not satisfy the aggregation criteria. The resulting graph structure thus represents the final set of detected cellular aggregates. As illustrated in the exemplary case shown in Figure 8, aggregates are confirmed when adjacent leukocytes or platelets are identified, while aggregates containing only erythrocytes are disregarded, as these are not relevant to the targeted aggregation analysis in this context. In particular, Figure 8 illustrates the full data processing pipeline based on the YOLO-Graph model, i.e., the combination of object detection module 106 and aggregation analysis module 107 in accordance with embodiments of the disclosure.

[0079] As shown in Figure 8, the amplitude and phase components 805 are provided to backbone network 801 of the YOLOv8x-p2 model. Backbone network 801 is responsible for extracting multi-scale features from the input phase images and does so by progressively downsampling the input through a hierarchical structure (Pl to P5), allowing for the capture of both fine and coarse features. These extracted features are then passed to the model’s detection head 802, which includes a combination of submodules or blocks each configured to perform operations such as Conv2Fusion 83 for feature aggregation, upsampling 814 for resolution enhancement, concatenation 812 for feature fusion, detection 816 for generating predictions of bounding boxes, and convolution 815 for feature transformation.

[0080] During the training phase of the model, the outputs generated by the model’s detection head 802 are provided to loss model 803, which computes loss values to guide the optimization of network parameters. Loss model 803 comprises two primary loss components: Bounding Box (BBox) loss 821 and classification loss 822. These losses are calculated based on the difference between the model’s predicted outputs and corresponding ground truth labels and are subsequently used to optimize the weights of backbone network 801 and detection head 802 via backpropagation.

[0081] BBox loss 821 utilizes a regression-based approach, where each detected object is represented by parameters corresponding to the object’s spatial coordinates and dimensions. The notation "4><reg_max" indicates that a high-precision regression method is employed to optimize the bounding box predictions. This ensures that the detected bounding boxes closely align with the true object locations and sizes in the input phase images.

[0082] Classification loss 822 evaluates the accuracy of the model’s predictions in assigning the correct class labels to the detected objects This component may employ a cross-entropy loss function or a focal loss function to address class imbalance, ensuring that rare but clinically significant cell types, are accurately detected despite being underrepresented in the dataset. The "Adding Non-zero Classes" block ensures that only relevant detected objects contribute to the final loss calculation.

[0083] F oilowing obj ect detection and under the assumption that the model has been trained, the workflow proceeds to graph-based aggregation analysis model 804. The detected cellular structures are classified into categories such as white blood cells (WBC), platelets (PLT), and red blood cells (RBC) and are represented as nodes in a graph structure The system evaluates all pairwise combinations of detected cells to identify potential aggregation candidates. Edges are added between nodes if both cells belong to target classes (PLT or WBC) and if their centroid distances fall below a predefined threshold. Non-target cell types such as RBCs are excluded from the final aggregation analysis. The completed graph representation is then analysed to return aggregate result 806, identifying clinically relevant cell aggregates based on spatial proximity and cell type.

[0084] A pseudo algorithm showing the processes performed by graph-based aggregation analysis model 804 is set out in Algorithm 1 below in accordance with embodiments of the disclosure.

[0085] In accordance with embodiments of the present disclosure, a block diagram representative of components of processing system 900 that may be provided within computing module 100, and / or any of the modules shown in Figure 1 to carry out the computing and processing functions in accordance with embodiments of the disclosure is shown in Figure 9. One skilled in the art will recognize that the exact configuration of each processing system provided within these modules may be different and the exact configuration of processing system 900 may vary and the arrangement illustrated in Figure 9 is provided by way of example only.

[0086] Tn embodiments of the disclosure, processing system 900 may comprise controller 901 and user interface 902. User interface 902 is arranged to enable manual interactions between a user and the computing module as required and for this purpose includes theinput / output components required for the user to enter instructions to provide updates to each of these modules. A person skilled in the art will recognize that components of user interface 902 may vary from embodiment to embodiment but will typically include one or more of display 940, keyboard 935 and optical device 936.

[0087] Controller 901 is in data communication with user interface 902 via bus 915 and includes memory 920, processing unit or processor 905 mounted on a circuit board that processes instructions and data for performing the method of this embodiment, an operating system 906, an input / output (I / O) interface 930 for communicating with user interface 902 and a communications interface, in this embodiment in the form of a network card 950. Network card 950 may, for example, be utilized to send data from these modules via a wired or wireless network to other processing devices or to receive data via the wired or wireless network. Wireless networks that may be utilized by network card 950 include, but are not limited to, Wireless-Fidelity (Wi-Fi), Bluetooth, Near Field Communication (NFC), cellular networks, satellite networks, telecommunication networks, Wide Area Networks (WAN) etc.

[0088] Memory 920 and operating system 906 are in data communication with processor 905 via bus 910. The memory components include both volatile and non-volatile memory and more than one of each type of memory, including Random Access Memory (RAM) 923, Read Only Memory (ROM) 925 and a mass storage device 945, the last comprising one or more solid-state drives (SSDs). One skilled in the art will recognize that the memory components described above comprise non-transitory computer-readable media and shall be taken to comprise all computer-readable media except for a transitory, propagating signal. Typically, the instructions are stored as program code in the memory components but can also be hardwired. Memory 920 may include a kernel and / or programming modules such as a software application that may be stored in either volatile or non-volatile memory.

[0089] Herein the term “processor” or “processing unit” is used to refer generically to any device or component that can process such instructions and may include: a microprocessor, a processing unit, a microcontroller, a programmable logic device or other computational device. That is, processor 905 may be provided by any suitable logic circuitry for receiving inputs, processing them in accordance with instructions stored in memory and generating outputs (for example to the memory components or on display 940). In this embodiment, processor 905may be a single core or multi-core processor with memory addressable space. In one example, processor 905 may be multi-core, comprising — for example — an 8 core CPU. In another example, it could be a cluster of CPU cores operating in parallel to accelerate computations.

[0090] A flowchart which sets out the process for identifying cellular aggregates in realtime based on digital holographic images acquired using a digital holographic microscopy (DHM) setup using a computing module in accordance with embodiments of the present disclosure is illustrated in Figure 10. In embodiments of the disclosure, process 1000 as illustrated in Figure 10 may be performed by computing module 100 or any combination of modules described in the sections above.

[0091] Process 1000 begins at step 1002 with process 1000 acquiring digital holographic images of a sample captured using a digital holographic microscopy module, wherein the digital holographic images comprise spatial -domain interference patterns generated by an object beam and at least one reference beam. At step 1004, process 1000 then generates, for each of the digital holographic images, a corresponding amplitude image and phase image using an off-axis hologram reconstruction module. In embodiments of the disclosure, the off- axis hologram reconstruction module may comprise at least one Fourier imager head model, each Fourier imager head model corresponding to a respective reference beam and configured to perform frequency-domain filtering on a frequency-domain representation of the digital holographic image to obtain an isolated object beam component associated with the corresponding reference beam and a complex valued network model configured to convert the isolated object beam components into corresponding amplitude and phase image components. Process 1000 then proceeds to detect, using an object detection module, cellular structures within each of the digital holographic images based on the corresponding amplitude and phase image components. This takes place at step 1006. At step 1008, process 1000 then performs, using an aggregation analysis module, cellular aggregate analysis on the detected cellular structures within each of the digital holographic images to identify and classify cellular aggregates.

[0092] In accordance with embodiments of the disclosure, each Fourier imager head model comprises a set of trained two-dimensional matrices configured to perform frequency-domain filtering through matrix multiplications and element-wise multiplication.

[0093] In further embodiments, the frequency-domain filtering performed by the Fourier imager head model comprises process 1000 applying a frequency -domain origin conversion operation to shift a zero-frequency component of the frequency-domain representation of the digital holographic image from a comer region to a center region of the frequency-domain representation and extracting a target frequency region corresponding to an object beam component from the frequency-domain representation and perform a frequency-domain shifting operation to reposition the extracted target frequency region to the center region of the frequency-domain representation. Process 1000 then applies an element-wise spectral masking operation to the frequency-domain representation using a trainable spectral mask matrix to attenuate undesired frequency components and enhance isolation of the object beam component and applies a frequency-domain origin reversion operation to shift the zerofrequency component of the frequency-domain representation from the center region back to the corner region.

[0094] In further embodiments, the converting of the isolated object beam components by the complex valued network model comprises process 1000 performing an inverse Fourier transformation on each of the isolated object beam components received from each of the Fourier imager head models to obtain corresponding spatial-domain object waves and converting each complex-valued spatial-domain object wave into corresponding amplitude and phase image formats. Process 1000 then merges the multiple amplitude and phase image formats into a single combined object wave using a first set of trained convolutional neural network (CNN) layers and applies phase unwrapping to the single combined object wave using a second set of trained CNN layers to generate unwrapped phase image components.

[0095] In further embodiments, the converting of the isolated object beam components by the complex valued network model comprises process 1000 performing an inverse Fourier transformation on each of the isolated object beam components received from each of the Fourier imager head models to obtain corresponding spatial -domain object waves and merging the complex-valued spatial-domain object waves into a single combined object wave using a first set of trained convolutional neural network (CNN) layers. Process 1000 then converts single combined object wave into a corresponding amplitude and phase image format and applies phase unwrapping to the corresponding amplitude and phase image format using a third set of trained CNN layers to generate unwrapped phase image components.

[0096] In further embodiments, the performing of the cellular aggregate analysis on the detected cellular structures by the aggregation analysis module comprises process 1000 constructing a graph-based representation of the detected cellular structures, wherein nodes represent detected cells and edges represent spatial proximity relationships and identifying cellular aggregates by evaluating connected components in the graph-based representation based on a predefined distance threshold and target cell type combinations.

[0097] In other embodiments of the disclosure, a quantitative phase imaging (QPI) system may be used for identifying cellular aggregates in real-time. The QPI system may have a computing module that is configured to acquire quantitative phase images of a sample captured using a QPI module, wherein the quantitative phase images comprise spatial-domain interference patterns generated by an object beam and at least one reference beam. The computing module may also be configured to generate, for each of the quantitative phase images, a corresponding amplitude image and phase image using an off-axis hologram reconstruction module. In embodiments, the off-axis hologram reconstruction module may have at least one Fourier imager head model where each Fourier imager head model corresponds to a respective reference beam and is configured to perform frequency -domain filtering on a frequency-domain representation of the quantitative phase image to obtain an isolated object beam component associated with the corresponding reference beam. The system may also include a complex valued network module configured to convert the isolated object beam components into corresponding amplitude and phase image components. The system also includes an object detection module that is configured to detect cellular structures within each of the quantitative phase images based on the corresponding amplitude and phase image components, and to perform, using an aggregation analysis model, cellular aggregate analysis on the detected cellular structures within each of the quantitative phase images to identify and classify cellular aggregates.

[0098] In embodiments of the disclosure, the frequency-domain filtering performed by the Fourier imager head model comprises the Fourier imager head model being configured to apply a frequency-domain origin conversion operation to shift a zero-frequency component of the frequency-domain representation of the quantitative phase image from a corner region to a center region of the frequency-domain representation, extract a target frequency region corresponding to an object beam component from the frequency-domain representation andperform a frequency-domain shifting operation to reposition the extracted target frequency region to the center region of the frequency-domain representation, apply an element-wise spectral masking operation to the frequency-domain representation using a trainable spectral mask matrix to attenuate undesired frequency components and enhance isolation of the object beam component, and apply a frequency-domain origin reversion operation to shift the zerofrequency component of the frequency-domain representation from the center region back to the corner region.

[0099] In embodiments of the disclosure, the converting of the isolated object beam components by the complex valued network module comprises the complex valued network module being configured to perform an inverse Fourier transformation on each of the isolated object beam components received from each of the Fourier imager head models to obtain corresponding spatial-domain object waves; convert each complex-valued spatial-domain object wave into corresponding amplitude and phase image formats; merge the multiple amplitude and phase image formats into a single combined object wave using a first set of trained convolutional neural network (CNN) layers; and apply phase unwrapping to the single combined object wave using a second set of trained CNN layers to generate unwrapped phase image components.

[0100] In embodiments of the disclosure, the converting of the isolated object beam components by the complex valued network module comprises the complex valued network module being configured to perform an inverse Fourier transformation on each of the isolated object beam components received from each of the Fourier imager head models to obtain corresponding spatial-domain object waves; merge the complex-valued spatial-domain object waves into a single combined object wave using a first set of trained convolutional neural network (CNN) layers; convert the single combined object wave into a corresponding amplitude and phase image format; and apply phase unwrapping to the corresponding amplitude and phase image formats using a third set of trained CNN layers to generate unwrapped phase image components.

[0101] In further embodiments, the performing of the cellular aggregate analysis on the detected cellular structures by the aggregation analysis model comprises the aggregation analysis model being configured to construct a graph-based representation of the detected cellular structures, wherein nodes represent detected cells and edges represent spatial proximityrelationships; and identify cellular aggregates by evaluating connected components in the graph-based representation based on a predefined distance threshold and target cell type combinations.

[0102] In still further embodiments, the QPI module may comprise a digital holographic microscopy (DHM) module, a diffraction phase microscopy (DPM) module, a transport-of- intensity imaging module, or a spatial light interference microscopy module.

[0103] Example of a Digital Holographic Microscopy (DHM) setup

[0104] A schematic representation of a DHM setup in accordance with embodiments of the present disclosure is illustrated in Figure 11. As shown, DHM system is integrated with a microfluidic channel designed for high-contrast visualization of cellular structures. The side view of the microfluidic channel shows the sample flow region centrally positioned between two sheath flow regions, which provide hydrodynamic focusing of cells. This design ensures that cells are aligned and pass through a well-defined optical focal plane, enhancing imaging consistency. Coherent light source 1104 is directed through the microfluidic channel, producing an optical phase difference as it interacts with the cells. The resulting phase shifts capture detailed morphological information essential for the identification of various leukocyte subtypes and disease-related structures.

[0105] The interaction between light source 1104 and the flowing sample is depicted with the cross section of the reconstructed image derived from the optical phase difference being illustrated as plot 1102. Specifically, the refractive indices of the different regions, denoted as ni, n2, and n3, highlight the optical path differences encountered by coherent light source 1104. Shear forces within the sample channel facilitate the alignment and separation of cellular aggregates, such as, but are not limited to, leukocyte-platelet aggregates and bacterial contaminants, allowing the DHM system to capture detailed holographic images without requiring fluorescent labels or dyes. This label-free imaging approach preserves the natural state of the cells, supporting real-time analysis with minimal sample preparation.

[0106] High-contrast reconstructed images 1106 of various cellular aggregates (as processed by computing module 100) are also illustrated in Figure 11. Images 1106 demonstrate the DHM system and computing module 100’s ability to differentiate betweenleukocyte-leukocyte aggregates, platelet-leukocyte aggregates, and platelet aggregates, each clearly resolved with micrometre-scale precision. The accompanying scale bars confirm the system’s capacity to visualize fine cellular structures at a resolution of 4 pm.

[0107] Simulated outputs from an Aggregation Analysis module

[0108] Figures 12a, 12b and 12c illustrate simulated results of aggregate detection performed using the graph-based aggregation analysis module described in this disclosure. Specifically, Figure 12a demonstrates a case where no cellular aggregates were detected despite multiple individual cells being identified. Figure 12b shows the successful detection of a leukocyte-leukocyte (WBC-WBC) aggregate, highlighted by bounding boxes indicating spatial proximity and cell classification. Figure 12c illustrates the detection of a leukocyte-platelet (WBC-PLT) aggregate, again indicated by bounding boxes, showcasing the system’s capability to classify cell types and identify clinically relevant aggregation patterns.

[0109] Numerous other changes, substitutions, variations, and modifications may be ascertained by the skilled in the art and it is intended that the present application encompass all such changes, substitutions, variations, and modifications as falling within the scope of the appended claims.

Claims

CLAIMS1. A computing module for identifying cellular aggregates in real-time based on digital holographic microscopy (DHM), the computing module comprising: a processing unit; and a non-transitory media readable by the processing unit, the media storing instructions that when executed by the processing unit causes the processing unit to: acquire digital holographic images of a sample captured using a digital holographic microscopy module, wherein the digital holographic images comprise spatial-domain interference patterns generated by an object beam and at least one reference beam; generate, for each of the digital holographic images, a corresponding amplitude image and phase image using an off-axis hologram reconstruction module, the off-axis hologram reconstruction module comprising: at least one Fourier imager head model, each Fourier imager head model corresponding to a respective reference beam and configured to perform frequency-domain filtering on a frequency-domain representation of the digital holographic image to obtain an isolated object beam component associated with the corresponding reference beam; a complex valued network model configured to convert the isolated object beam components into corresponding amplitude and phase image components; detect, using an object detection model, cellular structures within each of the digital holographic images based on the corresponding amplitude and phase image components; and perform, using an aggregation analysis model, cellular aggregate analysis on the detected cellular structures within each of the digital holographic images to identify and classify cellular aggregates.

2. The computing module according to claim 1, wherein each Fourier imager head model comprises a set of trained two-dimensional matrices configured to perform frequency-domain filtering through matrix multiplications and element-wise multiplication.

3. The computing module according to claim 2, wherein the set of trained two- dimensional matrices comprises shift matrices M? configured to translate an origin of the frequency domain-representation to a center region of the frequency-domain representation, additional matrices,and M* configured to crop and shift a target frequency region to the center of the frequency domain-representation, shift-back matrices, M2and M2configured to translate shifted regions of the frequency spectrum back to their original frequency domainrepresentations, and a trained spectral mask matrix Mmaskconfigured to perform pixel-wise spectral filtering by attenuating high-frequency components in the cropped and shifted frequency domain-representation.

4. The computing module according to claims 2 or 3, wherein matrix-coefficients in the set of trained two-dimensional matrices are initialized based on properties of a predefined shape based filter.

5. The computing module according to claim 2, wherein the frequency-domain filtering performed by the Fourier imager head model comprises instructions for directing the processing unit to: apply a frequency-domain origin conversion operation to shift a zero-frequency component of the frequency-domain representation of the digital holographic image from a comer region to a center region of the frequency-domain representation; extract a target frequency region corresponding to an object beam component from the frequency-domain representation and perform a frequency-domain shifting operation to reposition the extracted target frequency region to the center region of the frequency-domain representation; apply an element-wise spectral masking operation to the frequency-domain representation using a trainable spectral mask matrix to attenuate undesired frequency components and enhance isolation of the object beam component; and apply a frequency-domain origin reversion operation to shift the zero-frequency component of the frequency-domain representation from the center region back to the comer region.

6. The computing module according to claim 5, wherein the converting of the isolated object beam components by the complex valued network model comprising instructions for directing the processing unit to: perform an inverse Fourier transformation on each of the isolated object beam components received from each of the Fourier imager head models to obtain corresponding spatial-domain object waves; convert each complex -valued spatial -domain object wave into corresponding amplitude and phase image formats; merge the multiple amplitude and phase image formats into a single combined object wave using a first set of trained convolutional neural network (CNN) layers; and apply phase unwrapping to the single combined object wave using a second set of trained CNN layers to generate unwrapped phase image components.

7. The computing module according to claim 5, wherein the converting of the isolated object beam components by the complex valued network model comprising instructions for directing the processing unit to: perform an inverse Fourier transformation on each of the isolated object beam components received from each of the Fourier imager head models to obtain corresponding spatial -domain object waves; merge the complex -valued spatial -domain object waves into a single combined object wave using a first set of trained convolutional neural network (CNN) layers; convert the single combined object wave into a corresponding amplitude and phase image format; and apply phase unwrapping to the corresponding amplitude and phase image format using a third set of trained CNN layers to generate unwrapped phase image components.

8. The computing module according to claim 1, wherein the performing of the cellular aggregate analysis on the detected cellular structures by the aggregation analysis model comprises instructions for directing the processing unit to: construct a graph-based representation of the detected cellular structures, wherein nodes represent detected cells and edges represent spatial proximity relationships; andidentify cellular aggregates by evaluating connected components in the graphbased representation based on a predefined distance threshold and target cell type combinations.

9. The computing module according to claim 1, wherein the object detection model comprises a You-Only-Look-Once (YOLO)-based neural network configured to detect cellular structures based on the amplitude and phase image components.

10. The computing module according to claims 6 or 7 wherein the set of two-dimensional matrices, and the CNN layers are jointly trained in an end-to-end supervised learning process using ground truth data comprising a target phase reconstruction of the hologram.

11. A method for identifying cellular aggregates in real-time based on digital holographic microscopy (DHM) using a computing module, the method comprising: acquiring digital holographic images of a sample captured using a digital holographic microscopy module, wherein the digital holographic images comprise spatial-domain interference patterns generated by an object beam and at least one reference beam; generating, for each of the digital holographic images, a corresponding amplitude image and phase image using an off-axis hologram reconstruction module, the off-axis hologram reconstruction module comprising: at least one Fourier imager head model, each Fourier imager head model corresponding to a respective reference beam and configured to perform frequencydomain filtering on a frequency-domain representation of the digital holographic image to obtain an isolated object beam component associated with the corresponding reference beam; a complex valued network model configured to convert the isolated object beam components into corresponding amplitude and phase image components; detecting, using an object detection model, cellular structures within each of the digital holographic images based on the corresponding amplitude and phase image components; and performing, using an aggregation analysis model, cellular aggregate analysis on the detected cellular structures within each of the digital holographic images to identify and classify cellular aggregates.

12. The method according to claim 11, wherein each Fourier imager head model comprises a set of trained two-dimensional matrices configured to perform frequency-domain filtering through matrix multiplications and element-wise multiplication.

13. The method according to claim 12, wherein the set of trained two-dimensional matrices comprises shift matricesconfigured to translate an origin of the frequency domainrepresentation to a center region of the frequency-domain representation, additional matrices, configured to crop and shift a target frequency region to the center of the frequency domain-representation, shift-back matrices, M2and M2configured to translate shifted regions of the frequency spectrum back to their original frequency domain -representations, and a trained spectral mask matrix Mmaskconfigured to perform pixel-wise spectral filtering by attenuating high-frequency components in the cropped and shifted frequency domainrepresentation.

14. The method according to claims 12 or 13, wherein matrix-coefficients in the set of trained two-dimensional matrices are initialized based on properties of a predefined shape based filter.

15. The method according to claim 12, wherein the frequency-domain filtering performed by the Fourier imager head model comprises the steps of: applying a frequency-domain origin conversion operation to shift a zero-frequency component of the frequency-domain representation of the digital holographic image from a corner region to a center region of the frequency-domain representation; extracting a target frequency region corresponding to an object beam component from the frequency-domain representation and perform a frequency-domain shifting operation to reposition the extracted target frequency region to the center region of the frequency-domain representation; applying an element-wise spectral masking operation to the frequency-domain representation using a trainable spectral mask matrix to attenuate undesired frequency components and enhance isolation of the object beam component; and applying a frequency-domain origin reversion operation to shift the zero-frequency component of the frequency-domain representation from the center region back to the comer region.

16. The method according to claim 15, wherein the converting of the isolated object beam components by the complex valued network model comprises the steps of: performing an inverse Fourier transformation on each of the isolated object beam components received from each of the Fourier imager head models to obtain corresponding spatial-domain object waves; converting each complex-valued spatial-domain object wave into corresponding amplitude and phase image formats; merging the multiple amplitude and phase image formats into a single combined object wave using a first set of trained convolutional neural network (CNN) layers; and applying phase unwrapping to the single combined object wave using a second set of trained CNN layers to generate unwrapped phase image components.

17. The method according to claim 15, wherein the converting of the isolated object beam components by the complex valued network model comprises the steps of: performing an inverse Fourier transformation on each of the isolated object beam components received from each of the Fourier imager head models to obtain corresponding spatial-domain object waves; merging the complex-valued spatial -domain object waves into a single combined object wave using a first set of trained convolutional neural network (CNN) layers; converting the single combined object wave into a corresponding amplitude and phase image formats; and applying phase unwrapping to the corresponding amplitude and phase image formats using a third set of trained CNN layers to generate unwrapped phase image components.

18. The method according to claim 11, wherein the performing of the cellular aggregate analysis on the detected cellular structures by the aggregation analysis model comprises the steps of: constructing a graph-based representation of the detected cellular structures, wherein nodes represent detected cells and edges represent spatial proximity relationships; and identifying cellular aggregates by evaluating connected components in the graph-based representation based on a predefined distance threshold and target cell type combinations.

19. The method according to claim 11, wherein the object detection model comprises a You-Only-Look-Once (YOLO)-based neural network configured to detect cellular structures based on the amplitude and phase image components.

20. The method according to claims 16 or 17 wherein the set of two-dimensional matrices, and the CNN layers are jointly trained in an end-to-end supervised learning process using ground truth data comprising a target phase reconstruction of the hologram.

21. A system for identifying cellular aggregates in real-time based on quantitative phase imaging (QPI), the system comprising: a computing module configured to: acquire quantitative phase images of a sample captured using a QPI module, wherein the quantitative phase images comprise spatial-domain interference patterns generated by an object beam and at least one reference beam; generate, for each of the quantitative phase images, a corresponding amplitude image and phase image using an off-axis hologram reconstruction module, the off-axis hologram reconstruction module comprising: at least one Fourier imager head model, each Fourier imager head model corresponding to a respective reference beam and configured to perform frequency-domain filtering on a frequency-domain representation of the quantitative phase image to obtain an isolated object beam component associated with the corresponding reference beam; a complex valued network module configured to convert the isolated object beam components into corresponding amplitude and phase image components; an object detection module configured to: detect cellular structures within each of the quantitative phase images based on the corresponding amplitude and phase image components; and perform, using an aggregation analysis model, cellular aggregate analysis on the detected cellular structures within each of the quantitative phase images to identify and classify cellular aggregates.

22. The system according to claim 21, wherein each Fourier imager head model comprises a set of trained two-dimensional matrices configured to perform frequency-domain filtering through matrix multiplications and element-wise multiplication.

23. The system according to claim 22, wherein the set of trained two-dimensional matrices comprises shift matricesconfigured to translate an origin of the frequency domainrepresentation to a center region of the frequency-domain representation, additional matrices, configured to crop and shift a target frequency region to the center of the frequency domain-representation, shift-back matrices, M2and M2configured to translate shifted regions of the frequency spectrum back to their original frequency domain -representations, and a trained spectral mask matrix Mmaskconfigured to perform pixel-wise spectral filtering by attenuating high-frequency components in the cropped and shifted frequency domainrepresentation.

24. The system according to claims 22 or 23, wherein matrix-coefficients in the set of trained two-dimensional matrices are initialized based on properties of a predefined shape based filter.

25. The system according to claim 22, wherein the frequency-domain filtering performed by the Fourier imager head model comprises the Fourier imager head model being configured to: apply a frequency-domain origin conversion operation to shift a zero-frequency component of the frequency-domain representation of the quantitative phase image from a corner region to a center region of the frequency-domain representation; extract a target frequency region corresponding to an object beam component from the frequency-domain representation and perform a frequency-domain shifting operation to reposition the extracted target frequency region to the center region of the frequency -domain representation; apply an element-wise spectral masking operation to the frequency-domain representation using a trainable spectral mask matrix to attenuate undesired frequency components and enhance isolation of the object beam component; andapply a frequency-domain origin reversion operation to shift the zero-frequency component of the frequency-domain representation from the center region back to the comer region.

26. The system according to claim 25, wherein the converting of the isolated object beam components by the complex valued network module comprises the complex valued network module being configured to: perform an inverse Fourier transformation on each of the isolated object beam components received from each of the Fourier imager head models to obtain corresponding spatial-domain object waves; convert each complex -valued spatial -domain object wave into corresponding amplitude and phase image formats; merge the multiple amplitude and phase image formats into a single combined object wave using a first set of trained convolutional neural network (CNN) layers; and apply phase unwrapping to the single combined object wave using a second set of trained CNN layers to generate unwrapped phase image components.

27. The system according to claim 25, wherein the converting of the isolated object beam components by the complex valued network module comprises the complex valued network module being configured to: perform an inverse Fourier transformation on each of the isolated object beam components received from each of the Fourier imager head models to obtain corresponding spatial-domain object waves; merge the complex -valued spatial -domain object waves into a single combined object wave using a first set of trained convolutional neural network (CNN) layers; convert the single combined object wave into a corresponding amplitude and phase image format; and apply phase unwrapping to the corresponding amplitude and phase image format using a third set of trained CNN layers to generate unwrapped phase image components.

28. The system according to claim 21, wherein the performing of the cellular aggregate analysis on the detected cellular structures by the aggregation analysis model comprises the aggregation analysis model being configured to:construct a graph-based representation of the detected cellular structures, wherein nodes represent detected cells and edges represent spatial proximity relationships; and identify cellular aggregates by evaluating connected components in the graph-based representation based on a predefined distance threshold and target cell type combinations29. The system according to claim 21 , wherein the obj ect detection model comprises a You- Only-Look-Once (YOLO)-based neural network configured to detect cellular structures based on the amplitude and phase image components.

30. The system according to claims 26 or 27 wherein the set of two-dimensional matrices, and the CNN layers are jointly trained in an end-to-end supervised learning process using ground truth data comprising a target phase reconstruction of the hologram.

31. The system according to any one of claims 21 to 30, wherein the QPI module comprises a digital holographic microscopy (DHM) module, a diffraction phase microscopy (DPM) module, a transport-of-intensity imaging module, or a spatial light interference microscopy module.

Images