Ai-powered digital scanning optical microscope

Abstract

An imaging system is disclosed for automatically classifying and / or analyzing a sample through motion-blurred images. The system includes an imager with a light source that directs light onto a stage that contains the sample and an image sensor, wherein the stage and / or the imager components are configured to move continuously in a lateral (x, y) direction. Image processing software executed by a computing device acquires motion-blurred video of the sample while moving the stage and / or the imager in the lateral direction, wherein the image processing software is further configured to stitch motion-blurred video image frames or optionally generated motion-deblurred video image frames into a larger field-of-view (FOV) image containing the one or more samples. A trained machine learning model is executed by the computing device configured to receive the larger (FOV) image or portion(s) thereof and output a classification and / or an analysis of the one or more samples.

Description

AI-POWERED DIGITAL SCANNING OPTICAL MICROSCOPERelated Applications

[0001] This Application claims priority to U.S. Provisional Patent Application No.63 / 710,581 filed on October 22, 2024 and U.S. Provisional Patent Application No.63 / 768,477 filed on March 7, 2025, which are hereby incorporated by reference in its entirety7. Priority' is claimed pursuant to 35 U.S.C. § 119 and any other applicable statute.Technical Field

[0002] The technical field generally relates to microscopes and in particular automated microscopes for imaging or scanning of samples. This includes, for example, devices and methods for the automatic scanning and analysis of immunohistochemically (IHC) stained tissue sections. In one particular example, the device was used to automatically assign human epidermal grow th factor receptor 2 (HER2) scores on IHC-stained breast tissue sections, achieving concordant results with those obtained from a high-end digital scanning microscope.Statement Regarding Federally SponsoredResearch and Development

[0003] This invention was made with government support under EB032840 awarded by the National Institutes of Health. The government has certain rights in the invention.Background

[0004] The advent of digitization in the field of pathology has drastically aided the medical workflow of histological and cellular investigations. Pathologists can now handle greater volumes of patient data with higher precision, ease and throughput. The digitization of biopsy tissue slides, for instance, has led to numerous favorable outcomes, with ameliorations in remote assessments, file transfers, research, analysis ergonomics, and overall patient care. Cyber advantages notwithstanding, a set of drawbacks accompany these enhancements, namely the speed, cost, and size of imaging hardware. State-of-the-art pathology scanners have speed metrics constrained by multiple factors, such as camera frame rate, stage stability, and slide exchange processes. The speed of conventional microscopes, irrespective of various efforts at acceleration, e.g.. through illumination manipulation, linescanning, multifocal plane imaging or time-delay integration (TDI), remains stunted by mechanical and optical complexities. Additional issues include the high cost of microscopes, with ranges averaging more than $200K, and cumbersome dimensions of top-notch digital pathology scanner systems, making them difficult to acquire in resource-limited institutions or in modest, short-staffed clinics. Additionally, to ensure continuous operation and backup during potential shutdowns, at least two digital scanners are required for each pathology department. To address this disparity, new cost-effective and compact microscopy solutions are necessary to democratize access to advanced pathology gear.Summary

[0005] Here, an artificial intelligence (Al)-powered, cost-effective, and compact digital scanning optical microscope is disclosed. The microscope, which was given the name “BlurryScope” and is sometimes referred to by this name operates as a scanning digital microscope. The microscope is precisely devised to achieve rapid scans of tissue slides at a markedly reduced cost and form factor than standard commercial alternatives. To rigorously ascertain the pragmatic boundaries of BlurryScope’s viability, the important and ambitious target of human epidermal growth factor receptor 2 (HER2) tissue classification was selected as a test case of the BlurryScope’s potential. Breast cancer (BC) remains one of the most common cancers globally, the most prevalent among women, and a leading cause of cancer-related deaths. Accurate histological diagnostics, including determining HER2 status, are essential for effective BC management. HER2 expression levels are crucial for assessing the aggressiveness of BC and guiding treatment decisions. However, traditional manual HER2 evaluation is time-consuming and prone to variability. The application of BlurryScope in the automated classification of HER2 scores on immunohistochemically (IHC) stained breast tissue sections thus serves as a challenging test for the BlurryScope platform, and as a potentially impactful auxiliary tool for medical practices. Therefore, as part of a proof-of-concept test, automated, deep learning-based HER2 score classification w as explored using BlurryScope. which contrasts notably with the conventional pathology pipeline, involving large, costly microscopy equipment.

[0006] According to one embodiment, the BlurryScope automates the entire workflow, encompassing scanning and stitching to cropping regions of interest and finally classifying the HER2 score of each tissue sample. To evaluate BlurryScope’s performance, tissue samples were used from tissue microarrays (TMAs) containing >1400 cores corresponding todifferent patients, which were split into 1144 cores used for training / validation and 284 cores used for bling testing. Images of each tissue sample were collected through a continuous scanning speed of 5,000 pm / s, which introduces bidirectional motion blur artifacts. These compromised images were then used to train the Fourier-transform-based neural network models for automated HER2 score classification through blurred images. A blinded test set of 284 unique patient cores was used to assess the accuracies of 4-class (HER2 scores: 0, 1+, 2+, 3+) and 2-class (HER2 scores: 0 / 1+, 2+ / 3+) classification networks, achieving 79.3% and 89.7% classification accuracy, respectively. Each patient's specimen was scanned three times to rigorously evaluate the reliability of the BlurryScope system, which revealed that the overall HER2 score consistency across all the tested cores was 86.2%, indicating a high level of repeatability in BlurryScope's classification performance, despite random orientations of the tested slides in each run. The results and analyses highlight BlurryScope's ability to supplement existing pathology7systems, reduce the time required for diagnostic evaluations, and advance the accuracy and efficiency of cancer detection, categorization, and treatment in resource-limited settings.

[0007] In one embodiment, an imaging system is disclosed for automatically classifying and / or analyzing one or more samples through motion-blurred images recorded with continuous relative movement between an imager and the one or more samples. The system includes an imager that includes one or more light sources configured to direct light along an optical path; an image sensor disposed along the optical path; a stage having a sample holder disposed along the optical path between the one or more light sources and the image sensor, wherein the stage and / or the imager is configured to move continuously in a lateral direction (x and / or y); image processing software executed by a computing device that is configured to acquire motion-blurred video image frames of the one or more samples while moving the stage and / or the imager in the lateral direction, wherein the image processing software is further configured stitch motion-blurred video image frames or optionally generated motion-deblurred video image frames into a larger field-of-view (FOV) image containing the one or more samples; and one or more trained machine learning model(s) executed by the computing device configured to receive the larger (FOV) image or portion(s) thereof and output a classification and / or an analysis of the one or more samples.

[0008] In another embodiment, a method of automatically classifying one or more samples with an imager includes providing a one or more samples on a stage located along an optical path of the imager between one or more light sources and an image sensor; acquiringmotion-blurred video image frames of the one or more samples while continuously moving the stage and / or the imager in one or more lateral directions (x, y) to generate motion-blurred video image frames; optionally deblurring the motion-blurred video image frames to generate motion-deblurred video image frames; stitching either the motion-blurred video image frames or the optional motion-deblurred video image frames into a larger field-of-view (FOV) image containing the one or more samples; inputting the larger FOV image or portion(s) thereof into one or more trained machine learning model(s) configured to output a classification and / or an analysis of the one or more samples.

[0009] In another embodiment, an imaging system is disclosed for automatically classifying and / or analyzing one or more samples through motion-blurred images recorded with relative movement between an imager and the one or more samples. The system includes an imager with one or more light sources configured to direct light along an optical path; an image sensor disposed along the optical path; a stage having plurality of individually actuatable electromagnets disposed in or adjacent to the stage; a sample holder or microscope slide holding the one or more samples disposed on the stage along the optical path between the one or more light sources and the image sensor, the sample holder or microscope slide containing a plurality of permanent magnets disposed thereon; image processing software executed by a computing device that is configured to acquire motion-blurred video image frames of the one or more samples while moving the sample holder or microscope slide relative to the stage, wherein the image processing software is further configured to stitch motion-deblurred video image frames or optionally generated motion-deblurred video image frames into a larger field-of-view (FOV) image containing the one or more samples; and one or more trained machine learning model(s) executed by the computing device configured to receive the larger FOV image or portion(s) thereof configured to output a classification and / or an analysis of the one or more samples.

[0010] In another embodiment, a method of automatically classifying one or more samples with an imager includes providing an imager comprising one or more samples on a microscope slide or sample holder disposed on a stage located along an optical path of the imager between one or more light sources and an image sensor, wherein the stage comprises a plurality of electromagnets disposed in or adjacent to the stage and wherein the microscope slide or sample holder contains a plurality of permanent magnets disposed thereon; acquiring motion-blurred video image frames of the one or more samples while moving the microscope slide or sample holder in one or more lateral directions (x, y) in response to actuation of oneor more of the plurality of electromagnets; optionally deblurring video image frames of the motion-blurred video image frames to generate motion-deblurred video image frames; stitching either the motion-blurred video image frames or the optional motion-deblurred video image frames into a larger field-of-view (FOV) image containing the one or more samples with image processing software; and inputting the larger FOV image or portion(s) thereof into one or more machine learning model(s) trained to output a classification and / or an analysis of the one or more samples.

[0011] In another embodiment, an imaging system is disclosed for automatically classifying and / or analyzing one or more samples through motion-blurred images recorded with relative movement between an imager and the one or more samples. The system includes an imager that has one or more light sources configured to direct light along an optical path; an image sensor disposed along the optical path; a stage comprising a plurality of holes or perforations coupled to a pressurized air or gas source; a sample holder or microscope slide holding the one or more samples disposed on the stage along the optical path between the one or more light sources and the image sensor; image processing software executed by a computing device that is configured to acquire motion-blurred video image frames of the one or more samples while moving the sample holder or microscope slide relative to the stage, wherein the image processing software is further configured to stitch motion-blurred video image frames or optionally generated motion-deblurred video image frames into a larger field-of-view (FOV) image containing the one or more samples; and one or more trained machine learning model(s) executed by the computing device configured to receive the larger FOV image or portion(s) thereof configured to output a classification and / or an analy sis of the one or more samples.

[0012] In another embodiment, an imaging system is disclosed for automatically classifying and / or analyzing one or more samples through motion-blurred images recorded with relative movement between an imager and the one or more samples. The system includes an imager that has one or more light sources configured to direct light along an optical path; an image sensor disposed along the optical path; a stage for holding a sample holder or microscope slide; a plurality of propellers and / or drones are configured to transport or move the sample holder or microscope slide relative to the stage; image processing software executed by a computing device that is configured to acquire motion-blurred video image frames of the one or more samples while moving the sample holder or microscope slide relative to the stage, wherein the image processing software is further configured tostitch motion-blurred video image frames or optionally generated motion-deblurred video image frames into a larger field-of-view (FOV) image containing the one or more samples; and one or more trained machine learning model(s) executed by the computing device configured to receive the larger FOV image or portion(s) thereof configured to output a classification and / or an analysis of the one or more samplesBrief Description of the Drawings

[0013] FIG. 1 illustrates one embodiment of an imaging system for automatically classifying and / or analyzing one or more samples or objects contained within the samples is disclosed. The imaging system automates the entire workflow from image acquisition to HER2 classification (as one example) by integrating specialized optical hardware and deep learning algorithms to rapidly process motion-blurred images with deep learning-based HER2 classification, offering an efficient and cost-effective alternative for task-specific inference in resource limited settings.

[0014] FIG. 2 illustrates the stitching algorithm used to generate larger FOV images that combines two techniques, correlation analysis and square-wave fitting, to align and merge image frames obtained from the video during the sane process accurately, even when the movement of the stage is not perfectly smooth.

[0015] FIG. 3 illustrates an enhanced Fourier Imager Network (eFIN) that is used for the one or more machine learning model(s) to perform classification of samples. This particular machine learning model operates in the Fourier (spatial frequency) domain.

[0016] FIGS. 4A and 4B illustrate a deblurring neural network that is configured as a generative adversarial network (GAN) that is used to de-blur video image frames during a deblurring operation. The generator network is illustrated in FIG. 4A while the discriminator network is illustrated in FIG. 4B.

[0017] FIG. 5 A illustrates a top-dowTi view of an embodiment that uses magnetic forces to levitate and move a sample holder holding a microscope slide. The stage of the imager includes electromagnets that are selectively actuated to move the sample holder and / or slide which has permanent magnets located thereon.

[0018] FIG. 5B illustrates a side view of an embodiment that uses magnetic forces to levitate and move a microscope slide. Electromagnets are illustrated in the stage while permanent magnets are disposed on the microscope slide.

[0019] FIG. 5C illustrates a side view of an embodiment that uses magnetic forces to levitate and move a microscope slide. Electromagnets are illustrated in the stage while permanent magnets are disposed on a sample holder that holds the microscope slide.

[0020] FIG. 6A illustrates a side view of another embodiment that uses magnetic forces along with air or pressurized gas to levitate and move a microscope slide according to one embodiment.

[0021] FIG. 6B illustrates a top-down view of the imager stage illustrated in FIG. 6A. The electromagnets, air holes or apertures, and air plenum are seen. A pump is connected to the air plenum.

[0022] FIG. 6C illustrates the manual manipulation of a microscope slide that floats above the stage of FIGS. 6A and 6B. The hand of a user physically manipulates a microscope slide that is disposed atop the stage. This embodiment illustrates both electromagnets and holes or perforations for reducing friction between the microscope slide and the stage (and to aid in levitating the same).

[0023] FIG. 7 schematically illustrates the 3D movement of the microscope slide or sample holder according to one embodiment.

[0024] FIG. 8 illustrates input and output cartridges disposed adjacent to the imaging stage of a microscope or other imaging system. Microscope slides are moved from the input cartridge over the stage for imaging or scanning operations followed by storage in the output cartridge.

[0025] FIG. 9 illustrates an ambodiment of the imager having a stage like that illustrated in FIGS. 6A-6C that inludes a camera system that identifies the positioning of the microscope slide or sample holder as well as the user’s hands.

[0026] FIG. 10A illustrates an embodiment of an imager that includes a display incorporated therein.

[0027] FIG. 10B illustrates an embodiment of an imager that includes a touchscreen display incorporated therein.

[0028] FIG. 11 illustrates an embodiment of an imager that includes a display or projector incorporated therein that projects a 3D hologram of the sample.

[0029] FIG. 12 illustrates an embodiment of an imaging system in a fluorescent microscopy mode that integrates a rectangular or circular tray of fluorescent emission and excitation fdter discs that are moving continuously (in vertical or horizontal movements in their respective planes) while the stage is also moving continuously.

[0030] FIG. 13 illustrates an embodiment of the system that can undergo near-instantaneous whole slide imaging by capturing multiple overlapping frames from different angles using an array of cameras and dichroic mirrors that selectively transmit and reflect light. The levitating sample is illuminated from multiple light sources, with light bouncing through the system via controlled reflections and refractions, ensuring comprehensive optical coverage. Each camera records a slightly different perspective, creating an extensive dataset with significant frame overlap.

[0031] FIG. 14 illustrates an embodiment of a propeller-based drone system that enables controlled 3D movement of a microscope slide and / or sample holder, cuvette or tubule containing solid or fluid specimens by employ ing propeller drive drones.

[0032] FIG. 15 illustrates another embodiment of a propeller-based drone system. In this embodiment, the drones are disposed on a sample holder which itself holds the microscope slide.

[0033] FIG. 16 illustrates another embodiment of a propeller-based drone system. In this embodiment, drones with magnets are used to interface with magnets located on the microscope slide (or sample holder).

[0034] FIG. 17 schematically illustrates how the propellers and / or drones may be used to move the optics, camera(s), and light source(s) in one direction and other propellers and / or drones move the microscope slide or sample holder in an opposing direction for faster scanning.

[0035] FIG. 18 illustrates an embodiment that may be linked to an inventory facility containing optical components, including mirrors, filters, waveguides, polarization elements, prisms, diffusive structures, digital micromirror devices, spatial light modulators, and apertures. These components can be integrated into the system through robotic mechanisms, propeller-based or drone-based transport, or magnetically controlled carriers that attach magnetic stickers or new propeller carriers to these items, enabling dynamic adjustments to the optical modalities to control the configuration of the imager.

[0036] FIG. 19 illustrates an embodiment that can be connected to a facility in proximity with a plurality of 3D printers capable of fabricating deep learning-designed optical components using plastics, polymers, composites, metals, glass, concrete, biomaterials, electronic materials, and conductive materials. These components are Al-designed to finetune, simplify, accelerate, or enhance imaging performance by integrating seamlessly into the optical pathway.

[0037] FIG. 20 illustrates that an additional facility can be integrated into this selfoptimizing optical system, incorporating 3D / laser printing technology to fabricate AI-designed sample holders using materials optimized for feasibility and performance.

[0038] FIG. 21 illustrates a detachable microscopy eyepiece that can be connected to the optical system, allowing the user to observe the specimen directly.

[0039] FIG. 22 illustrates that the eyepiece can be integrated with such a dedicated photonic deep learning training unit that utilizes optical processing components.

[0040] FIGS. 23A and 23B illustrate a comparison of a traditional digital pathology scanner and the BlurryScope system. FIG. 23A illustrates a traditional pathology scanner that uses a "stop-and-stare" method, producing high-resolution images with no motion blur, as shown by the sharp tissue core image inset. FIG. 23B illustrates how the Bluny Scope captures images continuously at -5,000 pm / s. introducing bidirectional motion blur artifacts, as depicted in the blurry core inset.

[0041] FIGS. 24A and 24B illustrate images of tissue cores with different HER2 scores. FIG. 24A shows images of tissue specimens with HER2 scores (0, 1+, 2+, 3+) obtained using a traditional digital pathology scanner, showing clear and well-defined cores for each HER2 score. FIG. 24B shows same as FIG. 24A, except the images are obtained using BlurryScope, where the images exhibit smudged details.

[0042] FIG. 25 illustrates the BlurryScope data processing pipeline according to one embodiment. FIG. 25 illustrates the data processing workflow of BlurryScope begins with the continuous video output of the scanned slides, followed by automated stitching and labeling. Images are then cropped and concatenated into a stack of subsampled patches. The image patches are then processed by deep learning-based classification netw orks (4-class and 2-class are illustrated). Scale bar 200 pm.

[0043] FIGS. 26A-26F illustrate testing accuracy as a function of the confidence threshold. FIG. 26A shows testing accuracy^ and indeterminate percentage for the 4-class HER2 classification system with 3N samples. FIG. 26B shows testing accuracy and indeterminate percentage for the 4-class HER2 classification system with the highest CI. FIG.26C illustrates testing accuracy and indeterminate percentage for the 4-class system with the Ci-weighted method. FIG. 26D illustrates testing accuracy vs. indeterminate percentage for the 2-class system with 3N samples. FIG. 26E shows testing accuracy and indeterminate percentage for the 2-class system with the highest CI. FIG. 26F illustrates testing accuracyand indeterminate percentage for the 2-class system with the Ci-weighted method. Grey- dashed lines refer to a 15% indeterminate rate.

[0044] FIGS. 27A-27F illustrate confusion matrices and classification accuracy. FIG. 27A shows a confusion matrix for the 4-class HER2 classification network for all scans. FIG. 27B shows a confusion matrix for the 4-class HER2 classification network with the highest CI scores. FIG. 27C shows a confusion matrix for the 4-class HER2 classification network with average CI scores. FIG. 27D shows a confusion matrix for the 2-class HER2 classification network for all scans. FIG. 27E illustrates a confusion matrix for the 2-class HER2 classification network with the highest CI scores. FIG. 27F shows a confusion matrix for the 2-class network with the average CI scores.

[0045] FIG. 28 illustrates prediction consistency for each core: the prediction consistency across all scanned cores using the BlurryScope imaging system for HER2 classification was assessed. The bar graph in FIG. 28 shows consistency levels for each core, defined as the proportion of the predictions that match the most frequent prediction (mode) across three scans per core. The results indicate high consistency across the majority of the cores (0.67 corresponds to two out of three matched scores), with some variability due to potential operational factors, such as slide placement or scan misalignments.

[0046] FIG. 29 illustrates ROC curves for 2-class HER2 cases: ROC cur es for the BlurryScope 2-class HER2 classification network were evaluated. Four methods of confidence interval (CI) integration — total scans, highest CI, average (absolute, no CI weights) and weighted average CI — are compared along with their area under the curve (AUC) values. These curves demonstrate the trade-offs between sensitivity- and specificity- for each classification method. The maximum CI method achieves the highest AUC, indicating a superior balance between correctly identifying true positives and minimizing false positives. The weighted average and average CI methods perform comparably but lag slightly behind in overall performance.

[0047] FIG. 30 illustrates the microscope performing the zigzag scanning operation which generates the smeared or blurry images that are then deblurred input to the trained neural network for classification. The sample is first focused using the axial motor for (z) direction adjustment. The lateral scan then takes place using lateral motors for scanning in the (x) and (y) directions in lateral movements. Still images are generated at the edges of the slide. The moveable stage stops at edges to move down and change direction.Detailed Description of Illustrated Embodiments

[0048] In one embodiment, and with reference to FIG. 1, an imaging system 10 for automatically classifying and / or analyzing one or more samples 100 or objects contained within the samples 100 is disclosed. The imaging system 10 includes an imager 12 and a computing device 14. The computing device 14 may include a personal computer or PC, laptop, tablet PC, server. Smartphone, a microcontroller or the like. In a preferred embodiment, the imager 12 is a microscope. The imager 12 may include optionally include one or more lenses 16 or lens sets therein. FIG. 1 illustrates a configuration that uses an objective (OBJ) lens 16 and a condenser (CON) lens 16. Alternatively, the imager 12 of microscope may include a lens-free configuration. In some embodiments the computing device 14 is integrated into the imager 12. however, in other embodiments the computing device 14 is separate from the imager 12. For example, the computing device 14 may be a separate or even remote computing device 14 (e.g., server) that performs image processing functionality and the machine learning operations discussed herein. The imager 12 includes one or more light sources 18 configured to direct light along an optical path 20 (FIG. 1). The one or more light sources 18 may include coherent or incoherent light sources. The one or more light sources 18 may include, for example, one or more light emitting diodes (LEDs), laser diodes, and the like. An image sensor 22, which may also be implemented as part of a camera in some embodiments, is disposed along the optical path 20 as seen in FIG. 1. The image sensor 22 may include an RGB CMOS image sensor typically found in a camera but other image sensors 22 may also be used. Examples include a CCD image sensor 22, monochrome CMOS image sensor 22, or a single-line scanning image sensor 22 or camera. For a single-line scanning image sensor 22 or camera, it may be configured for time delay and integration (TDI).

[0049] The imager 12 includes a stage 24 that is configured to hold the microscope slide 102 (as seen in FIG. 1), a sample holder 26 (e.g., FIG. 5A) or the sample 100 itself and is disposed along the optical path 20 between the one or more light sources 18 and the image sensor 22. The imager 12 may function in transmission mode in which case the microscope slide 102 and / or the sample holder 26 is / are optically transparent and light passes through the sample 100 and / or sample holder 26 to the image sensor 22. The sample holder 26 may also include an open region (e.g., a center region) that also allows light to pass through. If a separate sample holder 26 is not used, the light may also just pass through the microscope slide 102 or the sample 100. The imager 12 may also operate in reflection mode where lightreflected from the sample 100 is captured by the image sensor 22. The stage 24 is moveable in at least the x and / or y (lateral or horizontal) directions which is in the in-plane direction of the sample 100 or sample holder 26 for scanning. In some embodiments, the stage 24 may also move in the z-direction for, for example, focusing on the sample 100. Focusing may also be accomplished by z-direction movement of one or more lenses 16 (e.g., objective lens). The stage 24, in some embodiments, may use stepper motors for each axis of motion (x, y) along with lead screws and couplings used to drive respective movements of the stage 24. The computing device 14 may include instructions, program, or a script for controlling movement of the stage 24 and / or the imager 12 or components thereof in an automated fashion using control software 31. For example, Thonny (v4.14, Aivar Annamaa) software may be used to controlling movement of the stage 24 and / or the imager 12 or components thereof. This control software 31 may be part of or separate from image processing software 30 as shown in FIG. 1.

[0050] The computing device 14 or the functionality for controlling movement of the stage 24 may also be integrated into the imager 12 such as an on-board or integrated microcontroller or other processor 32. In a preferred embodiment, the stage 24 moves in both the x and y directions (e.g., left and right and in some embodiments up or down in the z-direction). In another embodiments, z movement is imparted on a lens 16 or set of lenses (e.g., objecting lens 16) to focus on the sample 100 or provide for volume scanning of the sample 100. In other embodiments, the stage 24 may also be moveable in the (axial) z-direction which is used to focus images the sample 100 that are captured by the image sensor 22. The sample 100 may be focused initially before laterally scanning but may also be focused during lateral scanning in alternative embodiments.

[0051] As an alternative embodiment, the imager 12 may move relative to a stationary stage 24 that holds the one or more samples 100. For example, the one or more light sources 18 and optics (e.g., lens(es) 16) of the imager 12 may move relative to the stage 24 that holds the sample holder 26 containing the sample 100 or the sample 100 itself. In yet another implementation both the stage 24 and imager 12 may move relative to each other. For example, as explained herein, movement of the stage 24 in one direction and movement of the imager 12 or components thereof in opposing directions can increase the overall scan speed of the sample 100 (FIG. 17). All that is required is relative movement between the imager 12 and the sample 100 which is used to capture the motion-blurred video as explained herein. In a preferred embodiment, the stage is continuously moveable in the lateral direction(x and / or y) at a speed of up to 10-100 mm / s. It should be appreciated that the speed of the relative movement may slow or even fully stop in some instances as the scanning approaches the edge of the sample 100 as a different portion of the sample 100 is imaged (e.g., moving to a different row). The continuous relative movement between the imager 12 and the one or more samples 100 may in some embodiments, include non-zero movement in an x-direction or y-direction during all or a portion of the scan. As noted herein, the continuous relative movement may occur at different speeds. For example, relative movement may be higher in the center region of the sample 100 while imaging at the edge(s) of the sample 100 may be slower or even stop in some instances.

[0052] The sample 100 may include tissue sample including immunohistochemically (IHC) stained tissue as described herein although other embodiments may scan unstained tissue. The sample 100 (e.g., tissue) may also include other stains, reporter molecules, or dyes that are used to stain features of the sample 100. In other embodiments, the one or more samples 100 include a blood smear, an environmental sample, a biological sample, biomedical sample or, materials, glasses, polymers, plastics, liquid, or an article of manufacture (e.g., wafers, fabric, cloth). The sample 100 may include small objects therein that are characterized and / or analyzed by the imaging system 10. This may include organisms such as bacteria, yeast, parasites, etc. The sample 100 may also environmental organisms, particles or the like that are observed by the imager 12.

[0053] The optical path 20 may include one or more lenses 16 or lens sets in the optical path 20 before and / or after the sample 100. In the embodiment disclosed herein, there is a 10X objective lens 16 prior to the sample 100 and a condenser lens 16 located after the sample 100. The objective lens 16 is mechanically connected to a motor and moveable in the orthogonal (z) direction at up to 10 mm / s. In some embodiments of the imaging system 10, the one or more light sources 18, the lenses 16 or lens sets, the stage 24, and the image sensor 22 are contained within a housing 28 such as illustrated in FIG. 1. The housing 28 may be printed from a polymer material using an additive (e.g., 3D) printer or any other technique such as molding of components which are then assembled into the imager 12. It should be appreciated that other materials and manufacturing processes may be used for the housing 28. The overall imaging system 10 is compact and cost effective. As seen in FIG. 1, the exemplary imager 12 is 35 cm in height and weight around 5 lbs. or less. Note that conventional stop-and-stare microscopes are much larger and may weigh -60-120 lbs. The imager 12 may have a smaller size and / or weight as the design is further optimized.

[0054] The computing device 14 that forms part of the system 10 includes image processing software 30 executed thereby that is configured to acquire a motion-blurred video of the one or more samples 100 while moving the stage 24 in lateral directions (or moving imager 12 or components as discussed herein). The image processing software 30, which is executed by one or more processors 32, is further configured to stitch motion-blurred video image frames of the acquired video into a larger field-of-view (FOV) image (in some cases a whole slide image) containing the one or more samples 100.

[0055] The stitching algorithm, illustrated schematically in FIG. 2, combines two techniques, correlation analysis and square-wave fitting, to align and merge image frames obtained from the video during the sane process accurately, even when the movement of the stage 24 is not perfectly smooth. With reference to FIG. 2, in operation 300 video is obtained of the sample during the imager scanning operation with relative movement between the imager 12 and the sample 100. Image frames are extracted from the video as seen in operation 310. The algorithm then determines the moving or static states based on the correlation between image frames and an optional motion-model as seen in operation 320. In this process, the algorithm first calculates the correlation between consecutive image frames to determine whether each image frame represents motion (“moving”) or stillness (“static”). Groups of highly correlated frames indicate a stable scan line. In the scanning sequence, each horizontal line includes a short pause before and after horizontal scanning (about 0.5 seconds, or 30 frames) and a vertical jump to the next line (about 3 frames). These pauses and jumps can be detected through correlation data.

[0056] When the sample area lacks distinct image features, correlation alone may not clearly indicate motion. To handle such cases, the algorithm uses a predefined model of the zigzag scan pattern. This model alternates between “moving” and “not moving” states consistent with the zigzag pattern, represented mathematically as a square-wave function, matching the known timing of scanning and pausing. Using this combined correlation and motion-model data, the algorithm identifies the start and end of each scan line with high accuracy and reconstructs the larger FOV image of the sample from the video image frames as seen in operation 330 in FIG. 2. In some cases, the larger FOV image 106 (FIG. 1) may a whole slide image. The entire stitching process is performed in MATLAB and requires about 2 minutes per slide (7.8 mm2 / s) when executed on a GeForce RTX 4090 GPU. As noted herein, the video image frames that are stitched together may be motion-blurred image framesor they may optionally be motion-deblurred image frames generated using a deblurring operation described herein.

[0057] The stitching operation was implemented with MATLAB using several built-in functions to automate video-to-image reconstruction. Frames are extracted from the recorded video using VideoReader and read, and their similarity is analyzed with corr2 or normxcorr2 to determine whether each frame represents motion or stillness. The algorithm then fits a square-wave model to the correlation data using fit or square to represent the zigzag scan pattern (FIG. 30) of moving and paused frames. Transition points between motion and stillness are detected with diff, identifying the start and end of each scan line. Stable frame regions are aligned and merged using the imregcorr, imwarp, and imfuse functions to create the larger FOV image or the final stitched whole-slide image. GPU acceleration with gpuArray and performance timing with tic / toc enable efficient processing, allowing each slide to be reconstructed in about two minutes on a GeForce RTX 4090

[0058] With reference to FIG. 1, the computing device 14 further executes one or more machine learning model(s) 34. which in some implementations includes a trained neural network, that is / are configured to receive the larger FOV image 106 or portions thereof and trained to output a classification and / or analysis of the one or more samples 100. The categories may include a qualitative output (e.g., low, intermediate, or high, or “positive"’ or “negative”) or it may be a quantitative output such as a number, number range, or the like. In the specific example used herein, the output includes HER2 scores of IHC-stained breast tissue (i.e., 0, 1+, 2+, 3+ or 0&1+ and 2+&3+). The classification and / or analysis may also include the detection and / or quantification of one or more biomarkers in the sample 100 (e.g., tissue sample).

[0059] The one or more machine learning model(s) 34 may include an enhanced Fourier Imager Network (eFIN) that operates in the Fourier (spatial frequency) domain. This model or network 34 analyzes images not just in normal pixel space but also in the frequency domain, where patterns and textures are easier to distinguish. Inside eFIN, as shown in FIG.3, the image is first converted into its frequency components using a Fourier transform 164. A lightweight U-Net subnetwork 166 then adjusts these frequency features dynamically before they are brought back into image space using an inverse Fourier transform 168. The processed image then passes through convolution layer 170 and activation (PReLU) layer 172 to extract features for classification. For HER2 analysis, the original eFIN model 34 was modified in two ways: (1) a global average pooling (GAP) layer 174 was added at the end toreduce each image to a 4-value output corresponding to the four HER2 score classes (0, 1+, 2+. 3+); and (2) the model was trained using cross-entropy loss, which ensures the predicted class probabilities align with verified pathologist scores. The model 34 was trained using the AdamW optimizer with adaptive learning rates and executed on a GeForce RTX 3090 GPU, classifying each image stack in about 0.85 seconds. These adjustments keep eFIN’s original strengths, the frequency-domain precision and efficiency, while tailoring it for accurate medical image classification. Additional details regarding the eFIN model 34 may be found in Chen et al., eFIN: Enhanced Fourier Imager Network for Generalizable Autofocusing and Pixel Super-Resolution in Holographic Imaging, IEEE Journal of Selected Topics in Quantum Electronics, vol. 29, no. 4: Biophotonics, pp. 1-10, July-Aug. 2023, Art no.6800810. which is incorporated by reference herein.

[0060] The imaging system 10 operates to automatically classify and / or analyze one or more samples 100. An exemplary workflow for the imager f2 includes providing one or more samples 100 on an optically transparent substrate (e g., glass microscope slide) disposed on the stage 24 located along the optical path 20 of the imager 12 between one or more light sources 18 and an image sensor 22. The optically transparent slide may be secured with a sample holder 26 or it may just rest atop the stage 24. The sample 100 may be brought into focus using the stage 24 to adjust the z-direction or through adjusting the z-direction of the lens(es) 16 such as the objective lens. A video is then acquired of the one or more samples 100 while moving the stage 24 and the sample 100 contained thereon in lateral directions (x. y). The movement of the lateral scan may proceed in a zigzag direction as illustrated in FIG.30. In some embodiments, the zigzag pattern is adjustable in terms of total width and length dimensions and / or the amount of row and / or column overlap. Other patterns of scanning include a spiral pattern, a raster pattern, or a beveled pattern.

[0061] The motion-blurred video includes image frames that collectively make up video are extracted as seen in operation 210 of FIG. 2. The video image frames are then stitched together with image processing software 30 to generate a larger field-of-view' (FOV) image containing the one or more samples 100. Note that in some embodiments, the video image frames that are stitched in to the larger FOV image 106 are motion-blurred video image frames. In other embodiments, the motion-blurred video image frames are optionally electronically de-blurred and then stitches into the larger FOV image 106. The larger FOV image 106 may include a whole slide image in some embodiments. The larger FOV image 106 or portions thereof are then input into one or more machine learning model(s) 34executed by the computing device 14 to output a classification and / or analysis of the one or more samples 100. For example, portions of the larger FOV image 106 may include image patches or cropped regions of the larger FOV image 106. In another embodiment, the larger FOV image 106 is first created from motion-blurred video image frames and then the larger FOV image 106 or portion(s) thereof are then subject to the deblurring operation.

[0062] In one specific embodiment, a method of automatically classifying one or more samples 100 with an imager 12 includes providing a one or more samples 100 on a stage 24 (or sample holder 26) located along an optical path 20 of the imager 12 between one or more light sources 18 and an image sensor 22. Motion-blurred video is then acquired of the one or more samples 100 while continuously moving the stage 24 and / or the imager 12 (or components thereof) in one or more lateral directions (x, y). This continuous movement of the stage 24 and / or the imager 12 (or components thereof) may follow a pre-programmed pattern as described herein. In some embodiments, the acquired motion-blurred video image frames that form the video are then then stitched into a larger FOV image 106 containing the one or more samples 100. Alternatively, the acquired motion-blurred video image frames are subject to a deblurring operation to convert the same to motion-deblurred video image frames. The motion-blurred video image frames or the motion-deblurred video image frames are then stitched into a larger FOV image 106 containing the one or more samples 100. In some embodiments, the larger FOV image 106 which was generated by stitched motion-blurred video image frames may be subject to a deblurring operation itself on the entire image or portions thereof. The larger FOV image or portion(s) thereof are then input into one or more trained machine learning model (s) 34 configured to output a classification and / or an analysis of the one or more samples 100. The one or more trained machine learning model(s) 34 may include a convolutional neural network, a recurrent neural network, a Fourier-transform-based neural network (as illustrated in FIG. 3), a fully connected neural network, or an artificial neural network as explained herein.

[0063] To de-blur the motion-blurred image frames acquired during the scanning process (or the larger FOV image or portion(s) thereof), a deblurring neural network 180 configured as a generative adversarial network (GAN) is used. In this embodiment, image-to-image supervised network training is performed using training images or image patches that represent fast (more blurred) and slow (less blurred) images using a Pearson correlation coefficient. The GAN that is trained includes a generator network 182 and a discriminator network 184 as seen in FIGS. 4A and 4B. The generator network 182 takes as an input ablurred image 186, which may be a motion-blurred video image frame and outputs a deblurred image 188 (e.g., a motion-deblurred video image frame). The discriminator network 184 takes the output, namely the de-blurred image 188, of the generator network 182 and compares this to a ground truth image and then determines whether the de-blurred, sharp image is true (real) or false (fake). The results of the discriminator netw ork 184 output are used to adjust and optimize the parameters of the generator network 182 during the training operation. Specifically, the generator network 182 is trained to produce outputs that cannot be distinguished from ground truth images by the trained discriminator network 184. Once the training is complete, the generator network 182 w ith the updated parameters / w eights is then established and functions as the deblurring neural netw ork. The generator network 182 is based on the known U-Net, which has eight layers of encoding and decoding, respectively. The discriminator network 184 has four (4) stages of encoding before making a decision on the authenticity of the data as real or fake. In the specific implementation used herein, the generator network 182 and the discriminator network 184 w ere trained with batches of images of size 256 x 256 x 1-3 pixels, cropped from 600x800 x 1-3 images from a training set. The batch size was set to 2. The learning rate was set to 0.0002 and the maximum number of epochs was set to 200. The model w as implemented using MATLAB. The training w as performed on a NVIDIA GeForce GTX 30900 GPU with 24 GB of memory. Additional details regarding using a GAN-based deblurring neural netw ork may be found in Fanous et al.. GANscan: continuous scanning microscopy using deep learning deblurring, Light: Science & Applications (2022) 11 :265, which is incorporated by reference herein.

[0064] In one alternative embodiment, and with reference to FIGS. 5A-5C, the imager 12 employs magnets to impart relative movement between the imager 12 and the sample 100. In this embodiment, the stage 24 of the imager 12 which is used to hold the sample 100 or a sample holder 26 that holds the sample 100 employs magnetic forces for lateral manipulation of the slide 102 (FIG. 5B) or sample holder 26 (FIG. 5C) across the stationary stage 24. With reference to FIG. 5 A, in this embodiment, there is a grid (e.g., array) of electromagnets 40 disposed in or adjacent (below) the upper stage surface of the stage 24. The electromagnets 40 are individually actuatable to control movement of a microscope slide 102 or a sample holder 26 that contains the sample 100. The microscope slide 102 may, as explained below, be contained in a sample holder 26 as illustrated in FIG. 5C that contains permanent magnets 42 therein or thereon that are used to impart movement of the sample holder 26 and microscope slide 102 loaded in the sample holder 26. The permanent magnets 42 mayinclude by way of example, rare-earth magnets such as Neodymium (Nd-Fe-B) and Samarium Cobalt (SmCo). Alternatively, the permanent magnets 42 may be disposed on the microscope slide 102 itself as illustrated in FIG. 5B in which case a separate sample holder 26 may be omitted as the slide 102 functions as a sample holder. For example, magnetic materials can be integrated onto the microscope slide 102 via an adhesive or as magnetic stickers that may be secured to the microscope slide 102. Permanent magnets 42 mas also be secured to the microscope slide 102 or the sample holder 26. Sequential activation of the electromagnets 40 in the stage 24 enables smooth, contactless x- and y-axis translation of the microscope slide 102 or sample holder 26 (with microscope slide 102) in any direction at speeds ranging 1-5000 mm / s. That is to say, the sample holder 26 and / or the microscope slide 102 are moved laterally with respect to the stage 24 using only magnetic force.

[0065] By modulating the relative power of adjacent electromagnets 40, the magnetic force is finely tuned, allowing precise positioning of the slide 102 or sample holder 26 at any point within the grid or array as illustrated in FIG. 5A. In this embodiment, the sample holder 26 and / or slide 102 is able to float on or is in gliding contact with the stage 24 with very little faction much like a puck floats or glides on the surface of an air hockey table when actuated. In one embodiment, the electromagnets 40 are actuated according to a script or program to perform zigzag (or another pattern as described herein) scanning of the microscope slide 102 as illustrated in FIG. 30. The particular sequence or modulation of power of the electromagnets 40 may be controlled via control software 31 as described previously. In other embodiments, as explained herein, the microscope slide 102 or sample holder 26 (with microscope slide 102) may be manually manipulated by the user to move the microscope slide 102 and / or sample holder 26 to various orientations on the stage 24. For example, magnetic forces may lift the slide 102 and / or sample holder 26 to hover over the surface of the stage 24. The user can then manually move the slide 102 and / or sample holder 26 laterally using his or her fingers as illustrated in FIG. 6C. In still other embodiments, combinations of automated and manual operations are contemplated. For example, an automated scanning operation may be interrupted by the pathologist or user to manually scan one or more regions of the microscope slide 102. After manual intervention, the automated scanning operation may proceed further. In some embodiments, the scanning operations or imaging aspects of the system 10 may be controlled via voice control or activation or through an application on a Smartphone or other portable electronic device.

[0066] In one embodiment, a method of automatically classifying one or more samples with an imager 12 as described herein. The method includes providing a one or more samples 100 on a stage 24 with electromagnets 40 therein located along an optical path 20 (FIG. 1) of the imager 12 between one or more light sources 18 and an image sensor 22. The samples 100 may be provided on a sample holder 26 or a microscope slide 102 that may include permanent magnets 42 disposed thereon. Motion-blurred video is then acquired of the one or more samples 100 while continuously moving the sample holder 26 and / or the microscope slide 102 the in one or more lateral directions (x, y). This continuous movement of the sample holder 26 and / or the microscope slide 102 may follow a pre-programmed pattern as described herein in response to actuation of one or more of the plurality of electromagnets. Alternatively, continuous movement of the sample holder 26 or the microscope slide 102 may manual movement of the same (e.g., with the hands of the user). A combination of both automatic movement and manual movement may also be possible. For example, a user may interrupt a scanning pattern with manual movement of the sample holder 26 or the microscope slide 102. After manual intervention, the automatic scanning pattern may proceed again. The acquired motion-blurred video image frames obtained from the video then stitched into a larger FOV image 106 containing the one or more samples 100. Optionally, motion-blurred video image frames are subject to a deblurring operation to convert the motion-blurred video image frames into motion-deblurred video image frames. The larger FOV image 106 or portion(s) thereof are then input into one or more trained machine learning model(s) 34 configured to output a classification and / or an analysis of the one or more samples 100.

[0067] In another embodiment, and with reference to FIGS. 6A-6C, the automated microscope stage 24 can also integrate both magnetic and air-cushion-based mechanisms to enable frictionless, high-speed manipulation of the microscope slide 102 and / or sample holder 26 across the stage 24. To minimize resistance and vibration, a low-pressure air cushion is introduced beneath the sample holder 26 or slide 102, allowing it to float on a thin layer of air. This effect is achieved through a perforated stage 24, which contains a grid of micro-perforations or holes 60 ranging from 50 to 200 pm in diameter, through which compressed air is evenly released. A small, quiet air pump 62, such as a diaphragm or centrifugal blower, supplies this controlled airflow, with a pressure control valve ensuring the maintenance of a stable air film. Beneath the stage 24, a plenum chamber 64 equalizes the airflow, guaranteeing uniform distribution through the perforations or holes 60 andpreventing turbulence or uneven lifting forces. To further refine stability, a pressure feedback loop dynamically adjusts the airflow intensity based on the weight of the slide 102 and / or sample holder 26 and motion requirements, maintaining a consistent floating height, typically between 10 and 100 pm above the surface of the stage 24. FIG. 6C illustrates how a user may manually manipulate the slide 102 or sample holder 26 (not shown) as it floats over the surface of the stage 24. For example, as explained herein, the imager 12 may scan the slide 102 in an automated fashion and a user may interrupt this automated scan to look at a particular region or regions of the slide 102 by physically grabbing or touching the slide 102 (or sample holder 26).

[0068] While the air-cushion-based mechanism described above as being used in conjunction with magnetic levitation, it should be appreciated that in some implementations or embodiments, the imaging system 10 may just use the air-cushion-based mechanism without magnetic levitation. For example, pressured air may be used to levitate the slide 102 and / or sample holder 26 and further enable the same to float on or in gliding contact with the stage 24 without the need for the electromagnets 40 or the permanent magnets 42.

[0069] In another embodiment, magnetic forces are used to control movement of the microscope slide 102 or sample holder 26 (which may be a microscope slide holder) in three dimensions including motion in the vertical (z) direction. The sample holder 26 may include a tubule, cuvette, vial, or the like. The microscope slide 102 or sample holder 26 is equipped with magnetic features (e.g.. permanent magnets 42), which may be integrated as an adhesive or embedded material as discussed herein. Coordinated activation of surrounding electromagnets 40 in or below the stage 24 allows for both rotational and translational motion at speeds of 1-5000 mm / s, facilitating precise orientation and positioning in three-dimensional space. This may be used with or without air-cushion-based forces as described above. This system 10 enables full 3D imaging at any angle, partially addressing the missing cone problem and allowing adjustable-speed illumination from any desired direction. This 3D movement may be integrated with the levitating microscope slide 102 or sample holder 26 that is illustrated in FIG. 7.

[0070] In another embodiment, and with reference to FIG. 8, the imager 12 may utilize cartridges 70, 72 that are used to stage or otherwise hold microscope slides 102 or other sample holders 26. An input cartridge 70 is provided that stores microscope slides 102 or other sample holders 26 prior to scanning. An output cartridge 72 is provided that stores microscope slides 102 or other sample holders 26 after scanning. The input and outputcartridges 70, 72 are each capable of holding between 1 and 500 microscope slides 102 or other sample holders 26. These cartridges 70, 72 are lined with permanent magnets 74 and designed to maintain the respective microscope slides 102 or sample holders 26 in a free-floating state using magnets, magnetic adhesives or stickers arranged in alternating polarity as seen in FIG. 8. The input cartridge 70 and the output cartridge 72 are positioned on either side of the stage 24 and enable automated slide or specimen handling. Microscope slides 102 or other sample holders 26 are moved from the input cartridge 70 to the active scanning area of the stage 24 and then stored in the output cartridge 72 (as seen by arrows in FIG. 8). To initiate scanning, a magnetic force is applied to the bottom microscope slide 102 or sample holder 26 disposed in the input cartridge 70, guiding it to the imaging position on the stage 24 through electromagnetic manipulation as described previously. Upon completion of scanning, the microscope slide 102 or sample holder 26 is transferred to the output cartridge 72 via controlled electromagnetic forces. The remaining microscope slides 102 or sample holders 26 in the output cartridge 72 are incrementally repositioned vertically. This may be accomplished through a mechanical rod or pusher member 76 that raises the stack of microscope slides 102 or sample holders 26. A pressure-based lift system may also be employed that uses compressed air or gas to push the stack upwards. In another option, a coordinated magnetic lift system may be used that alleviates weight accumulation on lower slides or sample holders by alternating the interaction between magnets located on the microscope slide 102 or sample holder 26 and the magnets 74 lining the respective cartridges 70, 72.

[0071] In one embodiment or aspect of the invention and with reference to FIG. 9, an imaging system 10 for classifying and / or analyzing one or more samples 100 or objects contained within the samples 100 is disclosed that includes automated and / or manual operating modes. In the manual mode, a user such as a pathologist manually manipulates the microscope slide 102 other sample holder 26. For example, the user will physically manipulate the microscope slide 102 or sample holder 26 (with microscope slide 102) with his or her hands while viewing details of the image from the imager 12. A camera system 80 is provided as part of the imaging system 10 that identifies the positioning of the microscope slide 102 or sample holder 26 as well as the user’s hands. A display 82 such as that illustrated in FIG. 10A may be provided to see microscope slide 102 or slide holder as well as the user hands. Other information may also be provided in the image such as an overlay of the layout of the stage 24, the location of the electromagnetics 40, permanent magnets 42, andholes 60 in the stage 24. This camera system 80 has a large field of view and lower resolution continuously monitors the motion of the microscope slide 102 or and / or sample holder 26, as well as the user’s hand interactions with these objects, facilitating real-time tracking and adaptive control of the scanning process. The camera 80 generates a video or movie that, as explained below, is used for training a machine learning or neural network. For example, the imaging system 80 may be set to a manual or “loose” mode thereby allowing the user to manipulate the microscope slide 102 or sample holder 26 freely as if using a fully mechanical system while still benefiting from digital tracking. This hybrid functionality enables seamless transition between manual and automated control. For example, automated control may be interrupted for manual control which can then move back to automated control after the user intervention. Movements made in manual mode are continuously tracked and recorded for machine learning or artificial intelligence training, allowing the image system to learn user-based interactions and further enhance autonomous operation. More precisely, the movement of the microscope slide 102 and / or sample holder 26, along with focus adjustments and user interactions, can be recorded and tracked for machine learning or artificial intelligence training. This enables the system 10 to leam, replicate, and autonomously execute scanning procedures, effectively rendering the digital scanning microscope fully self-driving, if desired. These learned preferences may be tied to a specific user or pathologist because typically these individuals have a preferred manner of analyzing images of microscope slides. For example, a first pathologist may want a full scan of the microscope slide 102 prior to analyzing particular regions in more detail. A second pathologist may intervene during the automated scanning process to analyze regions during the scan. The system is able to leam the preferences of each pathologist and adjust operations accordingly.

[0072] The imaging system 10 may optionally include a display 82 as seen in FIG. 10A or touchscreen 84 as seen in FIG. 10B (e.g., liquid crystal display (LCD)) that can be integrated into or with the imager 12. The touchscreen 84 may be operated by circuitry or microcontroller (e.g., a Raspberry Pi module, microcontroller or a similar lightweight computing component) to display images in real time or near real time. The microscope slide 102 and / or specimen holder 26 can be manipulated directly by the user through touchscreen 84 gestures such as dragging, pinching, or other digital (finger) interactions. These inputs are translated into corresponding movements of the magnetic or motorized stage allowing realtime control. This allows the user to manipulate the microscope slide 102 and / or sampleholder 26 without physically touching the same. Additionally, all user interactions through the touchscreen 84 are tracked for Al-training, enabling the system to leam and replicate these actions, ultimately supporting fully autonomous operation, if desired, with the possibility of mimicking specific user activity footprint.

[0073] The LCD touchscreen 84 may be enhanced with three-dimensional (3D) renderings generated from 3D scanning operations, viewable through 3D glasses or advanced display technologies. These include a high-resolution LCD panel 85 combined with a specialized lenticular lens array and a computational rendering engine to create a stereoscopic or multiview hologram 86 as seen in FIG. 11. Additional configurations may incorporate a transparent glass surface with an integrated LED hologram fan, a Pepper's Ghost display, or a holographic pyramid. The holographic pyramid setup uses a transparent acrylic or glass reflector positioned above a screen that projects four perspectives of an object, creating the illusion of depth. Alternatively, an AR headset can be used for an immersive mixed-reality' experience. In all cases, this embodiment of the system 10 captures and processes the user's interaction with the microscope slide 102 and / or sample holder 26, enabling real-time feedback and interactive engagement with the displayed holographic content. An example of this embodiment is seen in FIG. 11.

[0074] In front and below the display 82, touchscreen 84, or panel 85, a haptic platform can be integrated by incorporating a system of actuators, sensors, and ultrasonic arrays to allow the user to "feel" the 3D morphology of the scanned sample 100. Capacitive, resistive, or piezoelectric sensors detect hand position and pressure, transmitting this data to a haptic engine that processes the interaction in real time. Voice-coil or piezoelectric actuators generate localized vibrations to simulate surface textures, while electrostatic or ultrasonic phased arrays create pressure differentials that replicate the contours and edges of the virtual object. For more advanced feedback, electrorheological or magnetorheological fluids can adjust surface stiffness dynamically, enabling the perception of varying material properties. By precisely modulating these forces, the system provides an immersive tactile experience, allowing users to explore the depth, shape, and texture of the 3D-scanned sample 100 through touch.

[0075] As explained herein and with reference to FIG. 1, the imaging system 10 may include computing device 14 that forms part of the system 10 includes image processing software 30 executed thereby that is configured to acquire motion-blurred video image frames of the one or more samples 100 while moving the stage 24 and / or sample 100 inlateral directions. The image processing software 30, which is executed by one or more processors 32, is further configured to stitch motion-blurred video image frames 186 or optionally motion-deblurred video image frames 188 of the acquired video into a larger FOV image 106 (in some cases a whole slide image) containing the one or more samples 100 as well as execution of machine learning models 34 or otherwise inputting the larger FOV image 106 or portion(s) thereof into machine learning models 34 for classification of the sample 100, with the option to halt specimen movement as needed. This can be performed in real time using models 34 trained on previously acquired high-fidelity data from relatively slower scans. These models 34 leverage learned patterns of motion blur and image misalignment to reconstruct sharp and seamlessly stitched images during high-speed scanning, ensuring real-time enhancement without compromising accuracy.

[0076] While a CMOS-based imager 12 may be employed it should be appreciated that other types of cameras or image sensors 22 may be used with the imaging system. This includes one or a plurality (or array) of a CCD, CMOS, or TDI or single line-scanning imagers / cameras, with the latter two capable of being synchronized with stage movements, including preprogrammed directional changes in zigzag scanning patterns, to enhance imaging speed while minimizing motion blur. A phase retarder element can also be incorporated to precisely delay the light path accordingly, ensuring alignment with the charge accumulation process in TDI imaging and optimizing signal integration. Further increases in speed can be achieved by allowing controlled motion blur, which is later reconstructed using deep learning models trained on data obtained from slower-moving samples 100, enabling high-fidelity image restoration.

[0077] The imaging system 10 described herein can be configured for multiple imaging modalities, including brightfield, darkfield, phase contrast, quantitative phase imaging, differential interference contrast (DIC), fluorescence, confocal, multi-photon, polarization, spectroscopic, localization microscopy, light sheet microscopy, and atomic force microscopy (using an airborne probe - also adding details to the hologram and haptic system), all allowing for the process of continuous (relatively fast and slow) and normal style acquisitions. It accommodates samples 100 that are label-free, histochemically stained, immunologically tagged, fluorescently tagged, or otherwise chemically treated to enhance contrast. The system 10 supports both coherent and incoherent light sources 18, ensuring adaptability across various contrast mechanisms and imaging applications.

[0078] In fluorescent microscopy mode, this system 10 can integrate a rectangular or circular tray of fluorescent emission filter discs 88 and excitation filter discs 90 that are moving continuously (in vertical or horizontal movements in their respective planes) while the stage 24 is also moving continuously as illustrated in FIG. 12. A dichroic mirror 92 is interposed in the optical path 20 between the illumination path and the emitted fluorescence which is captured by the image sensor 22. The blurring signal mixing is then untangled with corresponding stop-and-stare ground truths trained using a deep learning deblurring network 180 as described herein. When combined with a sample 100 with the same proteins tagged with different fluorophores, this can in some cases lead to a form of localization microscopy, as the different filter sets 88, 90 sweep in continuous motion, revealing blur widths with different signal segments, allowing for the identification of separate features at the 10-20 nm level. More than one camera or image sensor 22 or light source 18 can be included, with the same or different type of modality simultaneously imaging different parts of the same sample 100, all while the stage 24 is being measured in continuous acquisition, enabling faster measurement in general and in potentially different modalities at once.

[0079] In another embodiment, the system 10 can undergo near-instantaneous whole slide imaging by capturing multiple overlapping frames from different angles using an array of cameras or image sensors 22 and dichroic mirrors 92 that selectively transmit and reflect light as seen in FIG. 13. The levitating sample 100 (which may be through actuation of electromagnetics 40 or drones / propellers 122 as illustrated) is illuminated from multiple light sources 18, with light bouncing through the system 10 via controlled reflections and refractions, ensuring comprehensive optical coverage. Each camera or image sensor 22 records a slightly different perspective, creating an extensive dataset with significant frame overlap. A deep learning algorithm or machine learning model 94 reconstruct the final image 96 by aligning and fusing these frames, correcting for distortions, variations in intensity, and optical aberrations introduced by the reflective pathways. This computational approach eliminates the need for mechanical scanning, dramatically increasing imaging speed while maintaining high resolution and depth accuracy, effectively enabling real-time whole slide imaging instantly.

[0080] In another embodiment and with reference to FIGS. 14, 15, and 16, levitation of the microscope slide 102 or sample holder 26 for 1-500 slides is accomplished with a drone system 120. A propeller-based drone system 120 enables controlled 3D movement of a microscope slide 102 and / or sample holder 26 such as a microscope slide 102, cuvette ortubule containing solid or fluid specimens by employing a propeller-based drone mechanism. Each drone 124 includes propellers 122. which may be part of a small e.g., mm-sized drone 124 are driven via motors 126 located on the microscope slide 102 or sample holder 26. These may be powered by their own power source (e.g., battery) or through wireless inductive charging, resonant energy transfer, or capacitive coupling using an external power source (now shown). The drone 124 may include multiple propellers 122 as illustrated (e.g. two or four propellers 122). In some embodiments, multiple drones 124 are used to levitate and move the microscope slide 102 and / or the sample holder 26. For example, two such drones 124 are illustrated as moving a microscope slide 102 (FIG. 14) or sample holder 26 with microscope slide 102 (FIG. 15) but it should be appreciated that fewer drones 24 (e.g., as few as a single drone 124) or more than two drones 24 may be used. FIGS. 14 and 15 illustrate drones 124 being used to levitate the microscope slide 102 and / or sample holder 26 above the stage 24 and provide lateral relative movement of the microscope slide 102 for image scanning. Also illustrated are an input cartridge 70 and output cartridge 70 where microscope slides 102 can be stored and retrieved as needed.

[0081] In some embodiments such as illustrated in FIG. 16, the drone 124 includes a magnet 128 (e.g., small electromagnet or permanent magnet) that is used engage with a corresponding magnet 42 located on the microscope slide 102 and / or sample holder 26. This may include, for example, permanent magnets, magnetic stickers or the like that can be adhered to the periphery of the microscope slides 102 and / or on the sample holder 26. In this regard, the propeller 122 and / or drone 124 that contains the propeller 122 acts as a sort of ferry or helicopter that can move and / or transfer the microscope slide 102 and / or sample holder 26. This includes moving the microscope slide 102 relative to the imaging components of the imager 12 for scanning of the microscope slides 102 as well as loading and / or unloading of microscope slides 102 to storage areas as described herein. In other embodiments, the propeller 122 and / or drone 124 can be used to adjust the location or move other components of the imaging system 10. This includes, for example, the camera or image sensor(s) 22 or optical detectors, optical components (e.g., lenses 16), the light source 18, etc.). The propellers 122 and / or drones 124 may also be used to transfer or move custom manufactured or printed components that form part of the imaging system 10. These carrier drones 124 may move these components according to a script or program that may be implemented using machine learning and / or deep learning for system optimization.

[0082] The sample holder 26 is stabilized through a combination of aerodynamic thrust and fine-tuned air currents, ensuring precise altitude control and smooth translational and rotational motion. Adjustments to individual propeller speeds allow for controlled movement at velocities ranging from 1 to 5000 mm / s, providing dynamic positioning and orientation in three-dimensional space. This system facilitates full 3D imaging at any angle, mitigating the missing cone problem while enabling adjustable-speed illumination from any direction. The aerial suspension method eliminates the need for mechanical contact, further reducing vibration and interference during imaging.

[0083] In another embodiment or aspect, on the side of the imaging stage 24 there is a pressure or acoustic actuator 130 as seen in FIG. 15 to assist with finer or more extensive rotations of the microscope slide 102 and / or a sample holder 26. A pressure-based or acoustic actuator 130 enhances control of the levitating slide by applying specific forces for micro-rotations around pitch, roll, and yaw axes. Acoustic actuation relies on phased arrays of ultrasonic transducers generating standing wave patterns, creating localized pressure differentials that steer the slide without physical contact. This approach enables precise angular adjustments, making it suitable for delicate specimens in biological or nanotechnology applications. Alternatively, microfluidic air jets or compressed gas actuators provide mechanical forces through piezoelectric nozzles, MEMS-based valves, or directional micro-actuated jets, allowing real-time adjustments to slide orientation. A hybrid system combining pressure actuation with electromagnetic stabilization further refines angular positioning using small electromagnets or Eddy-current damping coils, ensuring precise and contactless control. This system allows sub-micron accuracy, compensates for vibrations in real-time, and operates through software-controlled feedback loops for dynamic slide stabilization in advanced microscopy and automated imaging.

[0084] The acoustic or pressure actuator 130 can furthermore be used, through deep learning networks, to deliver a customized ‘sheet’ of vibrations to a levitating microscopy slide 102 and / or sample holder 26, enabling real-time control over resonance effects, molecular interactions, and fluid dynamics in bioimaging applications. Al-driven adaptive acoustics can amplify weak signals by inducing resonance in biological molecules, improving detection sensitivity for biomarkers, DNA strands, and proteins. Dynamic surface modulation through controlled pressure waves can allow for enhanced fluid mixing, reagent delivery, and adaptive calibration in microfluidic applications. Al can also optimize vibration profiles to enable precise, label-free manipulation of cells, nanoparticles, and biomolecules, facilitatingsingle-cell analysis and high-sensitivity assays. The system’s ability to leam and adjust acoustic wave patterns in response to biosensor feedback ensures real-time optimization. This combination of Al-driven adaptive vibrations and non-contact sensing creates a highly programmable and multifunctional biosensing platform, capable of reconfiguring dynamically to enhance detection accuracy and experimental flexibility in microscopy-based diagnostics.

[0085] In another embodiment as illustrated in FIG. 17, the camera or image sensor 22, light source(s) 18 and optics (e.g., lens(es) 16) can move laterally independent of the microscope slide 102 and / or sample holder 26 or optical component holder 27 enabling motion of the microscope slide 102, sample holder 26, and / or optical component holder 27 in the opposite direction to enhance scanning efficiency. As seen in FIG. 17, the light source 18, lenses 16, and camera or image sensor 22 are moved in one direction while the microscope slide 102 or sample holder 26 moves in the opposite direction. As seen in FIG.17, the various optical components are held on an optical component holder 27. The optical component holder 27 may operate in a similar manner as the sample holder 26 described herein and interface with magnetic forces (e.g., using permanent magnets 42) to interface with drones 124. The difference is that the optical component holder 27 holds optical components rather than a microscope slide 102 with a sample 100. This movement can be achieved using a motorized system, magnetic levitation with electromagnetics 40 and permanent magnets 42 as described herein or drones 124 with propellers on each component (as illustrated). By counterbalancing the motion of the microscope slide 102, this approach effectively increases the relative scanning speed, potentially doubling the rate of image acquisition.

[0086] The scanning pattern utilized by the imaging system 10 with any of the scanning modalities described herein can follow a zigzag pattern, a spiral pattern, a raster pattern, a beveled pattern or dynamically optimized 3D trajectory, with Al-generated path planning based on initial specimen snapshots. By analyzing structural features in real time, the system 10 selects the most efficient scanning strategy to minimize redundant passes and maximize capture speed. Adaptive algorithms may adjust motion parameters to optimize resolution, contrast, and acquisition time, ensuring high-throughput imaging with minimal distortion or photobleaching.

[0087] In one embodiment, the components of the imager 12 including the camera or image sensor 22. optics such as lenses 16, microscope slide 102, stage 24, and light source(s)18, can achieve airborne stability using drones 124 or the previously described polar magnetic arrangement with surrounding electromagnets 40 and permanent magnets 42, or a hybrid approach that includes both propellers 122 and magnetic levitation and movement using electromagnets 40 and permanent magnets 42. Propeller-based stability enables controlled thrust, while polar magnetic configurations interact with dynamically controlled electromagnets 40 to provide levitation and guided motion. Remote power delivery is achieved through wireless inductive charging, resonant energy transfer, or capacitive coupling, ensuring continuous operation of electronic components without direct wiring. This system allows for precise, contactless positioning and movement, optimizing scanning speed and adaptability. They can also be packed neatly in a cube (sized -5-20 x 20 x 20 cm3) because of their untethered independent state.

[0088] The distance and positioning / coordinates of levitating components can be dynamically optimized in real time via deep learning algorithms, ensuring precise alignment and stability. The system continuously leams and refines the coordinates and Ay movements of each airborne element as seen in FIG. 17 through specific deep learning networks or other machine learning models 94 like in FIG. 13, enabling adaptive adjustments that enhance imaging accuracy and efficiency.

[0089] Computational processing is either integrated within the system 10 or performed remotely and consists of electronic and / or photonic components, including waveguides, optical resonators, nonlinear optical elements, and optical memory for data transmission and storage. Alternatively, processing is conducted using in vitro biological neural network cultures, which are stimulated either electronically through electrophysiological interfaces or optogenetically via controlled laser or LED signaling. A hybrid approach may also be employed, combining these modalities to optimize computational efficiency and adaptive learning capabilities.

[0090] In one aspect and with reference to FIG. 18, this imaging system 10 may be linked to one or more inventory facilities 190 containing between 10 and 10,000,000 optical components 192 used in the imaging system 10. including mirrors, filters, waveguides, polarization elements, prisms, diffusive structures, digital micromirror devices, spatial light modulators, and apertures. These components 192 can be integrated into the system 10 through robotic mechanisms, propeller-based transport as illustrated), or magnetically controlled carriers that attach magnets 42, magnetic stickers or propeller-based drones 124 to these items, enabling dynamic adjustments to the optical modality. This capability allows forreal-time reconfiguration via deep learning networks 194 for measurement optimization in terms of speed, resolution, total component cost, total component number, total size of the system, or signal enhancements like signal-to-noise ratio. This process would enable the automated creation of entirely new optical setups tailored to specific imaging requirements.

[0091] The sy stem 10, in some embodiments, can also be connected to a printer facility 196 in proximity with between 1-1000 3D printers 198 (FIG. 19) capable of fabricating deep learning-designed optical components 192 using plastics, polymers, composites, metals, glass, concrete, biomaterials, electronic materials, and conductive materials. These optical components 192 are Al-designed using a trained neural networks 200 to fine-tune, simplify, accelerate, or enhance imaging performance by integrating seamlessly into the optical pathway. The system 10 supports the fabrication of 2D optical sheets or 3D components ranging of a range of custom shapes from 1 nm to 10 cm in any dimension, incorporating optical properties such as reflection, refraction, dispersion, diffusion, absorption, scattering, diffraction, interference, polarization, spectral filtering, collimation, convergence, divergence, coherence, fluorescence, phosphorescence, luminescence, whispering gallery modes, evanescent wave propagation, and total internal reflection. The design and specifications of these components 192 are dictated by a deep learning-driven optimization process with trained networks 200 that iteratively minimizes a dynamically modulated generalized loss function, adapting based on user-defined objectives or a broad system prompt. A networked transport system 202 autonomously delivers fabricated optical components 192 stored in inventory into the optical setup using propeller-driven carriers such as drones 124. As illustrated, the drones 124 have an optical component holder 27 that holds the optical components 192. The drones 124 may interface with the optical component holder 27 with magnetic attachments such as permanent magnets 42 located on the optical component holder 27 that interface with magnets 128 on the drones 124 as illustrated in FIG. 16. Robotic automation may handle all mechanical assembly, including fastening, structural integration, and alignment, with access to auxiliary' components such as nuts, bolts, and a variety of fasteners and connectors. This self-optimizing manufacturing and deployment framework ensures continuous adaptation, precision, and efficiency in optical system performance, and works in coordination with the facility of premade components, as described previously.

[0092] With reference to FIG. 20, an additional facility 204 can be integrated into this self-optimizing optical system, incorporating 3D / laser printing technology to fabricate AI-designed sample holders 26 and / or optical component holders 27 using materials optimizedfor feasibility and performance, as outlined previously as seen in FIG. 19. These custom-designed sample holders 26 or optical component holders 27 may feature ferromagnetic structures (e.g., metasurfaces) for seamless integration with the electromagnetic platform, microfluidic channels for precise fluid manipulation, and resonant holders engineered to support whispering gallery modes for enhanced optical interactions. Additionally, they can incorporate mirrors, waveguides, and other advanced optical elements to refine signal quality and system adaptability. Furthermore, specialized deep learning-generated chemical and immunological patterns can be printed directly onto any surface of these holders, enabling highly targeted protein detection and expanding the system’s biosensing capabilities.

[0093] A data center facility' can also be connected wirelessly so that servers can store and transfer new data and new models can be trained on previously acquired high-fidelity data from relatively slower scans. These models would leverage learned patterns of motion blur or other image defects and image misalignment to reconstruct sharp and seamlessly stitched images during high-speed scanning, ensuring real-time enhancement without compromising accuracy. Models can also be re-trained periodically and updated as needed at the user side.

[0094] As another alternative for the manipulation of a sample 100 and / or reagents, fluid samples can be detected and manipulated with enhanced precision using acoustic levitation, which employs standing sound waves to suspend liquids without physical contact. An array of ultrasonic transducers generates controlled acoustic pressure nodes, allowing stable levitation and precise positioning of droplets. By modulating the acoustic field, droplets can be dynamically merged, split, or transported, enabling sophisticated sample handling.Additional reagents can be introduced mid-air using microdroplet dispensers or nebulizers, ensuring controlled mixing or selective chemical reactions. Colloidal particles, such as functionalized nanoparticles, can be suspended within the fluid and guided through frequency -tuned acoustic forces to facilitate localized reactions, enhance analyte detection, or modify optical properties for imaging. Integrating this system with fluorescence microscopy, Raman spectroscopy, or hyperspectral imaging can allow real-time biochemical analysis while eliminating contamination risks and the need for physical substrates like a slide or other sample holder. Deep learning-driven optimization can further refine acoustic field parameters, maximizing signal clarity and detection accuracy, ultimately enabling automated, high-throughput fluid analysis with exceptional control.

[0095] Specially engineered magnetic plates with custom-designed force profiles, optimized through deep learning-driven designs, can be assembled to operate in coordinationwith levitating magnetically susceptible particles - between 10 and 10,000,000. These particles which may include droplets that have a sample and / or reagents contained therein are initially launched into precise three-dimensional configurations through a controlled magnetic "kick," with fine-tuned adjustments dynamically regulated by surrounding electromagnets. This system enables passive optical signal manipulation, allowing for highly precise wavefront control with exceptional flexibility. By leveraging this approach, the platform can facilitate advanced optical computational tasks, offering innovative solutions for adaptive optics and computational imaging.

[0096] In another embodiment, a detachable microscopy eyepiece 206 (FIG. 21) can be connected to the optical system, allowing the user to observe the specimen directly. The eyepiece’s position in space is continuously tracked, enabling real-time synchronization with adjustable mirrors 208 that dynamically redirects the image plane to the eyepiece 206. This setup allows the user to view the sample 100 as if the eyepiece 206 were physically attached to the system 10, mimicking a conventional microscope configuration. The eyepiece 206 can be positioned anywhere within a range of 1 mm to 10 m, with the source intensity automatically adjusting in proportion to distance to maintain a consistent and stable brightness throughout movement. In this regard, the eyepiece 206 is not permanently physically attached to the microscope or other imager 12 and can be manipulated by the user for aid in user comfort or ergonomics.

[0097] A deep learning training photonic system that physically trains models using light, without any circuits or electronics, can be integrated into this setup with a secondary mirror directing the image to its input. This system would incorporate custom-designed photonic computational components, utilizing materials that undergo iterative modifications through repeated light exposure and reflections, such as silver halide emulsions, photoresists, and photochromic materials. These optically responsive elements would be dynamically employed and decommissioned using drone carriers, enabling continuous self-optimization of the system. The deployment and retrieval of these components would be further refined through laser-based modulation of carrier movement, ensuring precise spatial positioning and adaptive control of the optical learning process. A series of such photosensitive parts that be constructed iteratively in coordination with the a nearby biobank, to train on many specimen types.

[0098] Furthermore, the eyepiece can be integrated with such a dedicated photonic deep learning training unit 210 (FIG. 22) that utilizes optical processing components. Separateadjustable mirrors 208 directs the image plane to the input of this system, where optical transformations such as detection, classification, segmentation, and translation are performed entirely within the photonic domain. The processed optical output is then overlaid onto the real image and transmitted to the eyepiece 206, allowing the user to inspect the enhanced view in real time. This system operates purely with light signals, without any digital processing or projection, ensuring that the user observes a direct optical overlay rather than a digitally rendered augmentation.

[0099] This imaging system 10 can be dynamically reconfigured to support a broad spectrum of biosensing methodologies, each incorporating deep learning networks trained on paired slow and fast measurements. This architecture enables either high-speed classification or deblurring, depending on the specific imaging objective. Additionally, the system’s adaptability is enhanced through fully airborne components utilizing magnetism and / or propellers or drones employing the same, providing unprecedented flexibility in positioning and control. In biomarker detection, spectrophotometry enables precise optical analysis of biological samples, while ELISA facilitates high-sensitivity protein and antibody quantification. Genomic applications benefit from accelerated sequencing and molecular analysis. Ultrasound imaging is augmented by propeller-equipped probes, allowing for autonomous positioning and optimized imaging angles. Optical coherence tomography provides high-resolution, non-invasive visualization of tissue microstructures, while advanced radiological modalities, including X-ray imaging, computed tomography (CT), and magnetic resonance imaging (MRI), expand the system’s multi-modal diagnostic capabilities. This modular framework creates a highly versatile biosensing platform, leveraging deep learning to enhance image fidelity, improve analytical precision, and significantly accelerate diagnostic workflows. Additionally features such as manual / digital hybridity, activity tracking, holographic displays and haptic sensing would also apply.

[0100] The aforementioned different biosensing or microscopy systems and methods can furthermore undergo Al-driven self-learning and self-optimizing to continuously refine their capabilities through synergetic targets. These technologies can thus dynamically interact with one another, as well as with the optical component inventory (in some embodiments), AI-customized 3D-printed component inventory, adaptive sample holder inventory, and the data center and biobank. Through advanced deep learning architectures designed to optimize and expand sensing capabilities, this interconnected system can computationally generate novel technologies, autonomously evolving new configurations and methodologies. The result canmake unprecedented, continuously advancing biotechnologies capable of producing innovations in optical and biomedical imaging.

[0101] In all embodiments discussed herein, a small compact incubation unit can be integrated in the system 10 with small replaceable CO2 cartridges and a heating rod or element for cell cultures to be measured. The incubation unit may be co-located with or adjacent to the imager 12 or other imaging system so that samples may be easily transferred from the incubation unit to microscope slides 102 or other sample holders for imaging.

[0102] Results

[0103] Design of BlurryScope Imaging System

[0104] The BlurryScope imaging system 10 is designed to be a fast, compact, and cost-effective imaging system. The optical architecture of the imager 12, in the exemplary’ working embodiment, includes components adapted from a dismantled Ml 50 Compound Monocular AmScope brightfield microscope with an RGB CMOS image sensor 22, integrated into a custom 3D-printed framework as the housing 28. The system 10 is driven by three stepper motors for precise movement of the stage 24, achieving a stable lateral scanning speed of -5,000 pm / s with a lOx (0.25NA) objective lens 16. The other main optical components include the condenser lens 16 and LED light source(s) 18. It should be appreciated that the speed may vary’ and may be higher or lower than 5,000 pm / s. For example, it could move slower (e.g.. 300 pm / s or more). The structural parts were printed using SUNLU PLA+ fdament. maintaining the total component cost under $500 for low-volume manufacturing (see Table 1 below' and Methods for details, ‘BlurryScope design and assembly’).

[0105] The Blurry Scope imager 12 is a cost-effective alternative for routinely performing specialized inference tasks where machine learning models 34 such as trained neural networks can provide rapid, automated and accurate information regarding tissue specimens, such as the HER2 score classification that is the focus herein. The key performance trade-offs of BlurryScope imaging system 10 involve concessions in resolution, signal-to-noise ratio (SNR), and the detection of smaller objects, prioritizing speed, affordability’, and compact design. These limitations mean that BlurryScope imager 12 may, in some scenarios, lack the precision required for full diagnostic applications, but is still suitable as a complementary tool in clinical pathology’ environments.

[0106] To shed more light on the specifications of the Blurry Scope imaging system 10, several parameters are reported, including cost, speed, weight, and size in Table 1.Traditional digital pathology scanners can perform diffraction-limited imaging of tissue specimens at extreme throughputs and form the workhorse of digital pathology systems; however, their versatility and powerful features come with significantly higher costs, with prices ranging from $70,000 to $300,000, making them harder to scale up, especially in resource-limited environments. The compact design (35 x 35 x 35 cm) and lightw eight nature (2.26 kg) of the Blurry Scope imaging system 10 make it a practical solution for various medical settings. In a preferred embodiment the imager 12 weighs less than or equal to 10 kg and has volume < 50,000 cm3.Table 1Scanner Price Speed Weight Dimensions Digital pathology $70,000 - 52 x 52 x 62 cm - 1 -20 mm2 / s 34-55 kgscanners $300,000 120 x 85 x 100 cm BlurryScope <$500 3 mm2 / s 2.26 kg 35 x 35 x 35 cm

[0107] HER2 IHC tissue imaging

[0108] To demonstrate the efficacy of the BlurryScope imaging system 10, a total of 10 HER2-stained TMAs w ere used. The training and testing datasets consisted of 1144 and 284 unique patient specimens (tissue cores), respectively. Each patient sample was scanned three times (non-consecutively) to assess the repeatability of the approach, with a total duration of 5 minutes per scan (3 mm2 / s). This extensive dataset allowed for a comprehensive evaluation of the imaging system’s capabilities in automated HER2 scoring. The standard of comparison was the output of the same set of slides imaged with a state-of-the-art digital pathology scanner (AxioScan Zl. Zeiss).

[0109] FIGS. 23A-23B illustrates the differences in the acquired images using the AxioScan Z2 system and the BlurryScope imaging system 10. The stitch generated with a standard scanner (FIG. 23 A) has a clear and crisp delineation of all the cores since the scan undergoes a "stop-and-stare" operation. That is, the stage is physically halted for the duration of each camera acquisition operation. The left inset highlights the image clarity of a tissue core obtained with this method. In contrast, FIG. 23B demonstrates BlurryScope's continuous scanning output by capturing images at a running lateral stage speed of 5,000 pm / s. This rapid acquisition introduces bidirectional motion-blur artifacts, as depicted in FIG. 23B.Though there is a widening and smudging of features due to the effect of motion blur, theindividual cores are still fully separated in the final stitched mosaic, which allows for automated cropping and labeling of each patient tissue core for training of the machine learning model(s) 34.

[0110] The scanned tissue images corresponding to different HER2 scores (0, 1+, 2+, 3+) for individual patient cores are compared in FIGS. 24A-24B. FIG. 24 A shows the results from a traditional pathology scanner, yielding sharp, well-defined images for each HER2 score. In contrast, FIG. 24B presents the results from Blurry Scope image system 10, with stitched images exhibiting opposing directions of blur. Despite the smearing of various details, some correspondence between both image descriptions is still discernible. Lower-scored HER2 images exhibit fewer brown hues and less geometrical heterogeneity compared with higher-scored ones. This suggests that HER2 classification tasks may still be successful on such compromised data. The pertinent diagnostic information, though perhaps not apparent to an expert pathologist, is preserved in the BlurryScope images obtained during the continuous image scan despite extensive motion blur artifacts.

[0111] Automated classification of HER2 scores using BlurryScope images

[0112] With reference to FIG. 25, the data processing pipeline for the imaging system 10 begins by automatically organizing the images of each patient sample into multi-scale stacks as seen in operation 500. The process starts with scanning the slides containing the sample 100 (e.g., biopsy slides) and recording them in video format using the BlurryScope imager 12. These BlurryScope videos are then processed by image processing software 30 through automated stitching (operation 510) and labeling (operation 520) algorithms, which seamlessly integrate the frames into a whole-slide image. Subsequently, the individual cores were arranged into a concatenated stack of subsampled and randomly cropped patches (operation 530), ensuring that the image data are both precise and representative. The resulting data are then processed by a machine learning model 24 operating as a classification neural network (operation 540), configured for either 4-class (0,l+,2+,3+) or 2-class (0 / 1+ vs.2+Z3+) HER2 scoring. This approach allow s for the efficient handling of complex image data and ensures the repeatability of the classification process (see Methods for details on 'BlurryScope image scanning, stitching, and cropping and labeling’). Note that video image files from the BlurryScope videos may be blurry or de-blurred using deblurring network 180 illustrated in FIGS. 4A and 4B.

[0113] Upon finalizing both of the HER2-score machine learning models 34 or classification networks (see Methods for implementation details), the trained machinelearning models 34 were run on the blind test sets imaged by the Blurry Scope imaging system 10, covering N = 284 unique patient specimens / cores never seen before in the training phase. Since each slide 102 was scanned three times, this allowed the use of this extra data to improve final accuracy results; see FIGS. 26A-26F and 27A-27F. These multiple scans also enabled the assessment of the consistency of HER2 classification results across repeated measurements for the same tissue core. The degree of variability was quantified that might arise from factors such as slide insertion, alignment differences, and potential fluctuations in the scanning process itself. To achieve this, the prediction consistency was calculated for each core by comparing the classification results across the three scans. Specifically, for each core sample 100, the most frequently occurring prediction category was identified (i.e., the mode) among the three scans and then determined the proportion of predictions that matched this mode. The results revealed an overall consistency of 86.2% across all scanned core samples 100, demonstrating a high level of repeatability in the imaging system’s classification performance. As displayed in a bar graph of prediction consistency for each core sample 100 (see FIG. 28), the majority of the cores exhibit strong consistency, where at least two out of three results have the same score, though some variability is present. This suggests that, while the model 34 performs reliably for most samples, there are still certain cores where predictions are less stable, possibly due to factors like slide placement or operational conditions.

[0114] As detailed in the following analyses, three different distributions based on the triple measurements were evaluated for both HER2 classification neural networks 34: 1) total scans (3N), 2) maximum confidence interval (CI), and 3) average CI. Total scans include all the measured 3N images, while the highest CI method selects the result with the highest overall CI value from the three repeats, and the average CI method uses a Ci-weighted calculation. This weighted CI calculation involves multiplying each score by its corresponding CI, summing the results, and rounding the final value (see Methods section, Sample preparation and dataset creation).

[0115] One way to heighten the reliability of the Blurry Scope-based HER2 classification system is by leaving out results with low CI values and excluding them from the final assessment. To evaluate the balance between CI selection and accuracy vs. left out (indeterminate) percentages, and plotted their relationship for each data distribution and classification case. FIGS. 26A-26F shows that, as expected, the accuracy is proportional to the CI threshold score chosen, and the number of patients left out as indeterminate cases.FIG. 26A shows the testing accuracy and indeterminate percentages for the 4-class case with 3N samples, while FIGS. 26B-26C present the same relationship for the highest CI and average CI, respectively. These FIGS, illustrate how the chosen CI threshold value begins to exclude indeterminate patients starting around the 50% CI value mark. FIG. 26D displays the testing accuracy and indeterminate percentages for the 2-class network with 3N samples, while FIGS. 26E-26F present the same relationship for the highest CI and average CI.

[0116] In all these cases, there is a notable rise in the HER2 classification accuracy, along with indeterminate cases for CI selections above the 50% mark. A 5% improvement in HER2 classification accuracy in this range corresponds to -10% increase in the number of indeterminate cases. This suggests that once the CI value exceeds 50%, the user should be mindful of pursuing further improvements in accuracy, as they may result in substantial increases in dropout rates with indeterminate results. Overall, these analyses serve to illustrate that the Blurry Scope imaging system 10 can achieve a high testing accuracy with a manageable percentage of indeterminate results.

[0117] The classification accuracies for both networks 34 (4-class and 2-class HER2 inference) were also evaluated with confusion matrices, as shown in FIGS. 28A-28F.Threshold CI values were selected based on the plots in FIGS. 27A-27F corresponding to a 15% indeterminate rate - indicated by the grey dashed lines, which was empirically selected. The confusion matrix for the 4-class HER2 score inference of all the acquired BlurryScope images (3N) has a testing accuracy of 75.3% based on a 15% indeterminate CI threshold. Confusion matrices were also generated for highest and average CI scores (FIGS. 28B-28C), achieving HER2 score classification accuracies of 78.9% and 79.3%, respectively. Compared to automated HER2 classification results using microscopic images from a standard digital pathology scanner, these numbers prove competitive in performance, lagging only by a margin of -8-9%.

[0118] FIG. 27D represents the confusion matrix of all tissue scans for the 2-class HER2 classification machine learning model 34 (i.e., neural network in this embodiment), where 0 and 1+. and 2+ and 3+ groups are merged together, combining the two lowest and highest scores; these upper- and lower-bound categories are known to pathologists to have highly nuanced distinctions that are often difficult to differentiate. For this network 34, there is a markedly higher testing accuracy of 88.4% for a 15% indeterminate rate. When using the averaging CI method, the testing accuracy is slightly better, as shown in FIG. 27F, reaching an accuracy of 88.8%, and for the highest CI method, the accuracy increases even further to89.7%. For this model 34, the lower-right sections of the confusion matrices, which represent correctly identified negative cases, consistently show higher values compared to the upperleft sections, where true positive cases are recorded. This suggests the model 34 is better at correctly identifying negative cases, reflecting higher specificity. On the other hand, the relatively lower numbers for positive cases indicate that sensitivity is slightly lower, meaning the model 34 misses more true positives. This observation is important to note because while the model 34 effectively avoids false positives, it could potentially overlook some true positive cases, which would be critical to capture in medical diagnostics.

[0119] The receiver operative characteristic (ROC) curves were also plotted for these 2-class cases (FIG. 29) and demonstrate varying balances between sensitivity and specificity across different methods. The area under the curve (AUC) is a key metric used to evaluate the overall performance of an inference model 34, with higher AUC values indicating a better ability to distinguish between classes. The maximum CI method, with an AUC of 0.76, achieves the best performance, indicating a strong capability to maximize sensitivity while minimizing false positives. The absolute average CI distribution (not Ci-weighted), with an AUC of 0.74, performs similarly, slightly trailing the maximum CI approach but still maintaining a favorable balance. Overall, the maximum CI approach emerges as the most effective, achieving a decent balance between specificity and sensitivity, as reflected by its higher AUC. These analyses and results collectively indicate that the BlurryScope imaging system 10 is a promising digital imaging platform for quick inference of tissue biomarkers to potentially prioritize urgent cases or to streamline pathologists’ busy workflow.

[0120] Discussion

[0121] The utility of Blurry Scope imaging system 10 was demonstrated for automated HER2 scoring using a compact, cost-effective and rapid scanning microscope. The results for automated HER2 scoring on TMA slides using the BlurryScope imaging system 10 are concordant with those obtained from a high-end digital pathology scanner, although the latter shows improved performance. The framework of BlurryScope imaging system 10 offers promising possibilities in the sphere of digital pathology’, particularly in resource-limited settings or understaffed, small, suburban facilities. Traditional pathology scanners are substantially expensive, with costs ranging from $70,000 to $300,000. This prohibitive price is further compounded by the necessary procurement of multiple scanners to ensure continuous operation in clinical departments, posing a significant financial burden. In contrast, BlurryScope's total component cost is < $500, making it highly affordable forvarious resource-constrained settings. Furthermore, the compact dimensions and lightweight design of the imaging system 10 enhance its practicality for use in medical settings with limited space and resources.

[0122] It is, however, important to recognize the limitations of this technology. The primary performance trade-offs of Blurry Scope imaging system 10 include sacrifices to resolution, SNR, and smallest detectable object size in favor of speed, cost and form factor. These compromises mean that Blurry Scope imaging system 10 may not capture the finer details necessary for certain diagnostic applications, and thus could primarily be used as a supplementary7tool rather than a standalone digital pathology7solution in clinical settings. To mitigate some of these limitations, Al may be further integrated to potentially handle resolution loss, poor SNR levels, and sensitivity issues inherent in early versions of the BlurryScope imaging system 10. By leveraging advanced deep learning algorithms, one can enhance the quality7of images and possibly improve classification accuracy further. However, it is also important to acknowledge that Al systems can hallucinate, generating false or misleading information. This would raise the need for autonomous hallucination detection mechanisms to ensure the reliability of BlurryScope imaging system 10 in unsupervised settings. Clinically deployed BlurryScope imaging systems 10 would thus need to incorporate robust validation protocols and uncertainty7quantification metrics to address some of these concerns.

[0123] Moving forward, the prospects for enhancing and building on the imaging system’s capabilities are extensive. Because this is an inherently scalable technique based on exposure time and deep learning power, scanning speeds can be considerably increased using stronger motors (or different modes of relative continuous movement like magnets and the like) and more intricate Al networks. Applying this technology to analyze other sample types, such as blood smears, bacterial specimens, or defects, also holds significant promise. Additionally, the programmable z-axis feature could be leveraged for continuous three-dimensional (3D) sample imaging and sensing. Such a vertical scanning capability7could challenge existing deep learning-enabled autofocusing methods, providing faster and more accurate 3D imaging with improved depth-of-field.

[0124] In general terms, the concept of trading off data quality for hardware bargains can be integrated with various components for diverse applications. These include soft optics, AI-optimized filter cubes, diffractive deep neural networks (D2NNs). microfluidic setups, and spatial light modulators (SLMs). among others. The incorporation of such tools withBlurryScope’s scanning configuration could lead to the development of several valuable devices for both research and clinical uses. This initial version of Blurry Scope stands at the forefront of a series of forthcoming refinements and upgrades. Consulting with a diverse group of board-certified pathologists should provide crucial insights for fine-tuning the device to better meet their needs. BlurryScope's most immediate application would likely be in triaging and identifying questionable or tricky sample areas. In this capacity, the BlurryScope imaging system 10 is poised to excel in the short-term future, potentially providing critical support in clinical diagnostics.

[0125] Methods

[0126] BlurryScope design and assembly

[0127] The BlurryScope design process utilized Autodesk Fusion 360 for creating the detailed 3D models of the components of the imager 12. These models were then used to print the necessary parts on a Creality Ender 3 Pro 3D printer, ensuring precision and durability. The basic optics of BlurryScope imager 12 were adapted from aM150 Compound Monocular AmScope brightfield microscope with an RGB CMOS camera used for the image sensor 22 (30fps at 640x480 resolution, pixel size of 5.6 pm, and a sensor size of 3.59 x 2.69 mm). Also included is a condenser lens 16, lOx objective lens 16 (0.25NA) and LED light source 18 (FIG. 1). To ensure accurate imaging, a microscope stage calibration slide was used to align the optics and test the imager 12. For outfitting mechanical components, micro switches were used for the end stops of the linear actuators and two lead screws and couplings were incorporated for precise movement control. The system 10 also included three stepper motors to drive the linear actuators for movement of the stage 24 in the x, y directions and movement of the objective lens 16 in the z direction (a stepper motor and linear actuator may also be used to move the stage 24 in the z direction). The structural parts were printed using SUNLU PLA+ filament. Altogether, the optical components and surrounding materials amount to a cost of less than $500 USD.

[0128] The integration of these components allowed for the creation of a custom-designed framework that maintained the integrity of the imaging process and enabled programmable stage 24 and objective lens 16 movement. The imager 12 is powered and controlled using Thonny (v4.14, Aivar Annamaa), a Python integrated development environment (IDE), simplifying the deployment of the control software for the stage 24 and imaging system 10. The stage 24 was programmed to move in a zigzag configuration at 5,000 pm / s with at most 20% frame overlap using a lOx (0.25NA) objective lens 16. Although a higher NA objectivelens 16 and faster speeds could have been implemented, conservative stability and processing measures were taken to avoid potential mechanical failures in the long run.

[0129] BlurryScope scanning

[0130] The scanning and stitching process in BlurryScope imager 12 is fully automated using Thonny and MATLAB (vR2022b, MathWorks, Inc) software, respectively. The Thonny program controls the motorized stage 24 to ensure precise movement and continuous image acquisition. The AmScope camera software was used simultaneously with this to record videos of the samples 100 during the scanning process. The software is not synced or coordinated mechanically with the stage 24. The scan follows a zigzag geometry7and generates rows of opposing blur width. A motion-blurred image is a function of the stage speed and the camera acquisition time, as follows,

[0131] / (x, y; t) = / (% + sxt, y) (1)

[0132] where / (x, y), at stage speed sx=0, is an image at rest. The x-translated blurry7image, / , has the following time dependence:

[0133] / (x,y) = / (^,y)(KknO (2)

[0134] where T is the camera acquisition time,represents the convolution operator Sxover the variable x'sx, which has dimensions of time, and n is the ID rectangular function with a width of the blur distance sxT. Eq. (2) encapsulates the physical effect of spatial smearing as the result of a convolution operation. Thus, the blurred image is, in essence, its counterpart crisp image convolved along the direction of the scan by a rectangular function, with a width dictated by to the acquisition time. For a scanning speed of sx= 5,000 ^and T = 7.8ms, one has sxT = 39pm. This means that the continuous overlap of consecutive frames covers a margin of 39pm, in opposite directions for each row of scanning. The total scanning duration for a whole TMA slide is under 5 minutes, with a throughput of ~3mm2 / s. In a preferred embodiment, the motion-blurred images or video image frames have a blur width of less than 40 pm.

[0135] BlurryScope image stitching

[0136] The stitching of the captured video image frames into whole-slide images is achieved using image processing software 30 that executes an automated algorithm that combines correlation and square wave-fitting methods. This approach ensures that the images are accurately aligned and stitched together, even when there are slight variations (jerks) inthe movement of the stage 24. The stitching algorithm operates by analyzing the correlation between consecutive frames of the video to classify each frame in the video file as either "moving" or "static." It identifies windows of frames where the average correlation is very high, indicating a stable scan line. In the optimized zigzag scan pattern, there is a 0.5-second pause before and after each scan line (~30 frames) and a vertical jump to the next scan line (~3 frames). These pauses are detectable through correlation analysis. However, the algorithm can encounter errors in regions of the slide 102 without distinct features, where motion cannot be inferred from the video alone. To address this, the algorithm incorporates a model of the zigzag motion pattern. This model labels each frame as "moving" or "not moving," fitting a square wave to the data as the motion alternates between scanning (moving) and pauses (not moving) at the start and end of each scan line. Using this refined data, the algorithm can accurately identify the start and end of each scan line, enabling the generation of a stitch from the video image frames. The BlurryScope stitch is carried out entirely in MATLAB and takes ~2 minutes for a single slide 102 (7.8 mm2 / s) on a GeForce 4090 RTX graphics processing unit.

[0137] Tissue image cropping and labeling

[0138] For training of the machine learning models 34, the automated algorithm to crop and label tissue cores utilizes a spreadsheet that contains HER2 scores for each tissue sample 100 on the TMAs that were independently verified by three certified pathologists. This is the source for ground truth labels for training of the machine learning models 34. The process begins with a color correction (white balance) of each stitched TMA image. This is achieved by sampling a region devoid of tissue and other artifacts and subtracting it from the image, ensuring a uniform white background across all data. A grayscale copy of this image is then created, thresholded and blurred. The border is also cleared to facilitate the accurate drawing of contour boxes over tissue samples 100. The outermost coordinates of these contour boxes are saved, yielding a refined rectangular region that neatly encapsulates the tissue samples 100. This region is then divided according to the rows and columns listed on the TissueArray.Com database, which provides detailed core information for all purchased slides. Each box in the resulting grid contains a tissue core that is matched with its corresponding label from the TissueArray.Com diagram. The isolated tissue core image within each box is then saved and assigned a label ranging from 0 to 3+, which is derived from the pathologist-verified spreadsheet. Upon completion of the labeling process, taking 10s per slide (94mm2 / s). this transitions into the dataset creation phase.

[0139] Sample preparation and dataset creation

[0140] Histological samples were acquired from TissueArray, specifically breast tissue sections, and were stained for HER2 using standard IHC staining protocols at the UCLA Translational Pathology Core Laboratory. These samples 100 were then scanned using the BlurryScope imager 12 in sets of three repeats with different orientations and setup orders to account for various sources of variability. The dataset creation involved capturing multiple video image frames of each sample, which were then processed and stitched together. This dataset was then used to train and validate the HER2 score classification machine learning models or networks 34. Each training instance was a combination (3D concatenation) of 1 fully downsampled 512x512 image and four randomly cropped sections of the same 512x512 dimensions from the original stitched result (a 5125x512x3x5). The full dataset consisted of 3-fold scans of 1144 unique patient specimens (cores) for training and 284 for testing - i.e., 3N = 4284 core images in total with all the repeats.

[0141] Three distinct distributions were evaluated for both networks 34: 1) total scans, 2) maximum CI, and 3) average CL The total scans distribution incorporates all measured images across the three repeats, providing a comprehensive overview of all available data. The maximum CI method selects the result from the repeat that has the highest overall CI value, ensuring that the most confident prediction is used for that image. In contrast, the average CI method takes a more nuanced approach by calculating a Ci-weighted average. This involves multiplying each score by its corresponding CI value across the three repeats, summing these products, and then rounding the final result to provide a balanced prediction that accounts for all measurements while weighing them according to their corresponding confidence. This approach ensures that the final score reflects not just the raw predictions but the reliability of each repeat.

[0142] HER2 Score 4-Class and 2-class classification network architectures and training scheme

[0143] The classification machine learning model or network 34 for 4-class HER2 scoring was based on an eFIN architecture, a Fourier-transform-based network that manipulates the spatial frequency domain information using dynamical linear maps. Alternative networks that may also be used include one or more of: a convolutional neural network, a recurrent neural network, a fully connected neural network, and an artificial neural network. The network's architecture was modified to output 2 classes or 4 classes of HER2 scores by setting the last layer channel number to 2 or 4 and appending a global average pooling layer to the tail. Theclassification network for 2-class HER2 scoring was trained with hyperparameters and variables that were similar to the 4-class case, following the original eFIN architecture. The classification networks’ 34 training was optimized using an AdamW optimizer with a weight decay factor of 10-4. The training commenced with an initial learning rate set at 10 ’. which was dynamically adjusted using a cosine annealing scheduler with warm restarts. The training and testing operations were conducted on a desktop computer equipped with a GeForce RTX 3090 graphics processing unit, 64GB of random-access memory, and 13lhGen Intel Core™ i7 processing unit. The classification networks were implemented using PyTorch, with a single testing core image stack taking about 0.85 seconds to classify.

[0144] Additional Embodiments & Aspects

[0145] While embodiments of the present invention have been shown and described, various modifications may be made without departing from the scope of the present invention. For example, while the stage 124 is described herein as moving the optically transparent substrate with the sample it should be appreciated that the imager 12 may move relative to the sample 100 or combinations of the same may be used. This may be effectuated by a moveable stage 24, servos, motors, magnets, propellors 122 and / or drones 124 or equivalent that moves the imager 12 relative to the sample 100 (which could also include moving the imager 12 and the sample 100 at the same time). What is required is relative movement between the sample 100 and the imager 12. The invention, therefore, should not be limited, except to the following claims, and their equivalents.

Claims

What is claimed is:

1. An imaging system for automatically classifying and / or analyzing one or more samples through motion-blurred images recorded with continuous relative movement between an imager and the one or more samples comprising:an imager comprising:one or more light sources configured to direct light along an optical path; an image sensor disposed along the optical path;a stage having a sample holder disposed along the optical path between the one or more light sources and the image sensor, wherein the stage and / or the imager is configured to move continuously in a lateral direction (x and / or y);image processing software executed by a computing device that is configured to acquire motion-blurred video image frames of the one or more samples while moving the stage and / or the imager in the lateral direction, wherein the image processing software is further configured stitch motion-blurred video image frames or optionally generated motion-deblurred video image frames into a larger field-of-view (FOV) image containing the one or more samples; andone or more trained machine learning model(s) executed by the computing device configured to receive the larger (FOV) image or portion(s) thereof and output a classification and / or an analysis of the one or more samples.

2. The imaging system of claim 1, wherein the larger FOV image is generated from motion-blurred video image frames and the image processing software is further configured to generate a de-blurred larger FOV image or de-blurred portions thereof.

3. The imaging system of claim 1, wherein the computing device is associated with the imager.

4. The imaging system of claim 1, wherein the computing device is separate or remote from the imager.

5. The imaging system of claim 1, further comprising one or more lenses or set of lenses disposed in the optical path.

6. The imaging system of claim 1, wherein the stage is continuously moveable in the lateral direction (x and / or y) at a speed of up to 10-100 mm / s.

7. The imaging system of claim 1, wherein the imager weighs < 10 kg.

8. The imaging system of claim 1, wherein the imager has a volume that is < 50,000 cm3.

9. The imaging system of claim 1, wherein continuous relative movement between an imager and the one or more samples comprises non-zero movement in an x-direction ory-direction.

10. The imaging system of claim 1, wherein continuous relative movement occurs at different speeds.

11. The imaging system of claim 1 , wherein the image sensor comprises a CCD image sensor, CMOS image sensor, or a single-line scanning image sensor or camera.

12. The imaging system of claim 11. wherein single-line scanning image sensor or camera is configured for time delay and integration (TDI).

13. A method of automatically classifying one or more samples with an imager comprising:providing a one or more samples on a stage located along an optical path of the imager between one or more light sources and an image sensor;acquiring motion-blurred video image frames of the one or more samples while continuously moving the stage and / or the imager in one or more lateral directions (x. y) to generate motion-blurred video image frames;optionally deblurring the motion -blurred video image frames to generate motion-deblurred video image frames;stitching either the motion-blurred video image frames or the optional motion-deblurred video image frames into a larger field-of-view (FOV) image containing the one or more samples;inputting the larger FOV image or portion(s) thereof into one or more trained machine learning model(s) configured to output a classification and / or an analysis of the one or more samples.

14. The method of claim 13, wherein the one or more samples comprise one or more of the following: a biological sample, a biomedical sample, a tissue sample, a glass sample, a polymer sample, a plastic sample, a liquid sample, a wafer sample, a fabric sample, or a cloth sample.

15. The method of claim 14, wherein the one or more samples comprises a tissue sample and wherein the classification and / or analysis includes the detection and / or quantification of one or more biomarkers in the tissue sample.

16. The method of claim 14, wherein the one or more machine learning model(s) is executed using a computing device associated with the imager.

17. The method of claim 14, wherein the one or more machine learning model(s) is executed using a computing device separate or remote from the imager.

18. The method of claim 14, wherein the stage and / or the imager is continuously moved in a zigzag pattern, a spiral pattern, a raster pattern, or a beveled pattern as motion-blurred video is acquired of the one or more samples.

19. The method of claim 18, wherein the zigzag pattern is adjustable with respect to one or more of width and length dimensions and row and / or column overlap.

20. The method of claim 14, wherein multiple motion-blurred videos are acquired of the one or more samples in different orientations.

21. The method of claim 14, wherein the one or more machine learning model(s) comprises one or more of the following: a convolutional neural network, a recurrent neural network, a Fourier-transform-based neural network, a fully connected neural network, an artificial neural network.

22. The method of claim 14, wherein the stage and / or the imager is continuously moved at a speed of 300 pm / sec or more.

23. The method of claim 14, wherein the stage or an objective disposed in the optical path moves in a (z) direction orthogonal to the lateral directions (x, y).

24. The method of claim 14, wherein the output of the one or more machine learning model(s) comprises a classification.

25. The method of claim 14, wherein the output of the one or more machine learning model(s) comprises a numerical analysis and quantification.

26. An imaging system for automatically classifying and / or analyzing one or more samples through motion-blurred images recorded with relative movement between an imager and the one or more samples, comprising:an imager comprising:one or more light sources configured to direct light along an optical path; an image sensor disposed along the optical path;a stage having plurality of individually actuatable electromagnets disposed in or adjacent to the stage;a sample holder or microscope slide holding the one or more samples disposed on the stage along the optical path between the one or more light sources and the image sensor, the sample holder or microscope slide containing a plurality of permanent magnets disposed thereon;image processing software executed by a computing device that is configured to acquire motion-blurred video image frames of the one or more samples while moving the sample holder or microscope slide relative to the stage, wherein the image processing software is further configured to stitch motion-deblurred video image frames or optionallygenerated motion-deblurred video image frames into a larger field-of-view (FOV) image containing the one or more samples; andone or more trained machine learning model(s) executed by the computing device configured to receive the larger FOV image or portion(s) thereof configured to output a classification and / or an analysis of the one or more samples.

27. The imaging system of claim 26. wherein the larger FOV image is generated from motion-blurred video image frames and the image processing software is further configured to generate a de-blurred larger FOV image or de-blurred portions thereof.

28. The imaging system of claim 16. the imager further comprising a camera configured to capture video of relative movement of the sample holder or microscope slide and manual manipulation of the same by a user.

29. The imaging system of claim 26, further comprising one or more machine learning model(s) trained on the relative movement of the sample holder or microscope slide and manual manipulation of the same by the user.

30. The imaging system of claim 26, wherein the sample holder comprises a plurality of permanent magnets disposed thereon, the sample holder configured to hold a microscope slide therein.

31. The imaging system of claim 26, wherein the one or more light sources comprise coherent or incoherent light sources.

32. The imaging system of claim 26, the imager further comprising an input cartridge disposed adjacent to the stage and comprising a plurality' of magnets configured to hold a plurality of sample holders or microscope slides therein.

33. The imaging system of claim 32, the imager further comprising an output cartridge disposed adjacent to the stage and comprising a plurality7of magnets configured to hold a plurality of sample holders or microscope slides therein.

34. The imaging system of claim 33, the input cartridge and / or the output cartridge further comprising a mechanical rod or pusher member that lowers or raises the plurality of sample holders or microscope slides contained therein.

35. The imaging system of claim 33, the output cartridge further comprising a pressure-based lift system coupled to air or gas.

36. A method of automatically classifying one or more samples with an imager comprising:providing an imager comprising one or more samples on a microscope slide or sample holder disposed on a stage located along an optical path of the imager between one or more light sources and an image sensor, wherein the stage comprises a plurality of electromagnets disposed in or adjacent to the stage and wherein the microscope slide or sample holder contains a plurality of permanent magnets disposed thereon;acquiring motion-blurred video image frames of the one or more samples while moving the microscope slide or sample holder in one or more lateral directions (x, y) in response to actuation of one or more of the plurality of electromagnets;optionally deblurring video image frames of the motion-blurred video image frames to generate motion-deblurred video image frames;stitching either the motion-blurred video image frames or the optional motion-deblurred video image frames into a larger field-of-view (FOV) image containing the one or more samples with image processing software; andinputting the larger FOV image or portion(s) thereof into one or more machine learning model(s) trained to output a classification and / or an analysis of the one or more samples.

37. The method of claim 36, wherein the operation of optionally deblurring video image frames is performed by the same one or more machine learning model(s) trained to output a classification and / or an analysis of the one or more samples.

38. The method of claim 36, wherein the imager further comprises a camera configured to capture video of relative movement of the sample holder or microscope slide and manual manipulation of the same by a user.

39. The method of claim 38, further comprising one or more machine learning model(s) trained on the relative movement of the sample holder or microscope slide and manual manipulation of the same by the user.

40. The method of claim 39, wherein moving the microscope slide or sample holder in one or more lateral directions (x. y) in response to actuation of one or more of the plurality of electromagnets is done automatically according to script or program.

41. The method of claim 39, wherein moving the microscope slide or sample holder in one or more lateral directions (x. y) is done manually by a user.

42. The method of claim 36, wherein the microscope slide or sample holder is retrieved from an input cartridge by actuation of one or more of the plurality of electromagnets.

43. The method of claim 42, wherein the microscope slide or sample holder is stored in an output cartridge by actuation of one or more of the plurality of electromagnets.

44. The method of claim 36, wherein the microscope slide or sample holder is moved in a vertical direction (z) in response to actuation of one or more of the plurality of electromagnets.

45. An imaging system for automatically classifying and / or analyzing one or more samples through motion-blurred images recorded with relative movement between an imager and the one or more samples, comprising:an imager comprising:one or more light sources configured to direct light along an optical path; an image sensor disposed along the optical path;a stage comprising a plurality of holes or perforations coupled to a pressurized air or gas source;a sample holder or microscope slide holding the one or more samples disposed on the stage along the optical path between the one or more light sources and the image sensor;image processing software executed by a computing device that is configured to acquire motion-blurred video image frames of the one or more samples while moving the sample holder or microscope slide relative to the stage, wherein the image processing software is further configured to stitch motion-blurred video image frames or optionally generated motion-deblurred video image frames into a larger field-of-view (FOV) image containing the one or more samples; andone or more trained machine learning model(s) executed by the computing device configured to receive the larger FOV image or portion(s) thereof configured to output a classification and / or an analysis of the one or more samples.

46. A method of using the imaging system of claim 45, comprising forcing pressurized air or gas through a plurality of a plurality of holes or perforations disposed in the stage to elevate the sample holder or microscope slide holding the one or more samples above the stage.

47. An imaging system for automatically classifying and / or analyzing one or more samples through motion-blurred images recorded with relative movement between an imager and the one or more samples, comprising:an imager comprising:one or more light sources configured to direct light along an optical path; an image sensor disposed along the optical path;a stage for holding a sample holder or microscope slide;a plurality of propellers and / or drones are configured to transport or move the sample holder or microscope slide relative to the stage;image processing software executed by a computing device that is configured to acquire motion-blurred video image frames of the one or more samples while moving the sample holder or microscope slide relative to the stage, wherein the image processing software is further configured to stitch motion-blurred video image frames or optionally generated motion-deblurred video image frames into a larger field-of-view (FOV) image containing the one or more samples; andone or more trained machine learning model(s) executed by the computing device configured to receive the larger FOV image or portion(s) thereof configured to output a classification and / or an analysis of the one or more samples.

48. The method of using the imaging system of claim 47, further comprising transporting or moving the microscope slide or sample holder relative to the stage using the plurality of plurality of propellers and / or drones.

Images