System and method of classifying a depicted specimen

Abstract

A system and method of classifying a specimen depicted in an image may include receiving a sample image of a specimen, depicting a plurality of visual traits; inferring a machine- learning (ML)-based classifier model on the sample image, to obtain a baseline classification of the specimen; applying an ML-based generative model on the sample image, to obtain a reconstructed image of the specimen, via a latent feature vector, comprising a plurality of latent features; analyzing the reconstructed image, to identify at least one latent feature as correlative to the baseline classification of the specimen; and producing a refined classification of the specimen, comprising (i) the baseline classification and (ii) at least one visual trait, associated with the identified at least one latent feature.

Description

SYSTEM AND METHOD OF CLASSIFYING A DEPICTED SPECIMENCROSS-REFERENCE TO RELATED APPLICATIONS

[0001] This application claims the benefit of priority of U.S. Provisional Patent Application No. 63 / 538,083, filed September 13, 2023, titled “System and method of classifying a depicted specimen”, and U.S. Provisional Patent Application No. 63 / 547,552, filed November 7, 2023, titled “Method for interpreting image-based deep learning black box prediction models by using generative models and latent representation disentanglement”, the contents of which are all incorporated herein by reference in their entirety.FIELD OF THE INVENTION

[0002] The present invention relates generally to machine-learning based classification. More specifically, the present invention relates to refining classification of a depicted specimen.BACKGROUND OF THE INVENTION

[0003] With the rapid growing volume and complexity of modem biomedical visual data, it is no longer possible to rely on human capacity to identify visual patterns in biomedical images. Interpretable machine learning could enable computational-driven biological and clinical discovery, capitalizing on the machine’s unprecedented ability to automatically identify complex patterns within biomedical images, followed by human interpretation guided by visual explanation techniques. Combining human visual and cognitive capabilities with machine-learning (ML) based classification has the promise of leading to a generation of high-level insights, and testable hypotheses.

[0004] Currently available algorithms for visual explanation may provide heatmaps that highlight classification-driving image regions. However, such algorithms do not provide sufficient visual cues, to support specific hypotheses for classifier-relevant mechanistic understanding.SUMMARY OF THE INVENTION

[0005] Other algorithms for visual explanation, such as those commonly referred to as “counterfactual explanation” algorithms, aim to interpret classification models by visually altering an image sample to change its classification score. This is typically achieved by training a generative model which learns the distribution of the data and embeds this distribution in a lower dimensional space. This representation enables to alter the latent representation and correlate it to specific changes in the corresponding reconstructed image.

[0006] Embodiments of the invention provide a novel framework for visual explanation of a machine-learning based classifier output. Embodiments of the invention may incorporate output of the pre-trained classifier, with output of intermediate layers outputs (also referred to herein as latent features) to learn unique, semantic perceptual features (also referred to herein as visual traits) that drive the classifier’s decision making.

[0007] In other words, embodiments of the invention may disentangle the latent features in a novel manner, that may enable presentation of unique, specific visual traits to practitioners such as engineers, physicians and diagnosticians. The term “intuitive” may be understood in this context as facilitating high-level understanding of contribution of unique and specific visual traits of a studied specimen, to the classifier’s decision, without need for interpreting latent features in the deep learning domain.

[0008] Embodiments of the invention may thus provide an improvement over currently available counterfactual explanation algorithms, which cannot provide such high-level understanding.

[0009] As elaborated herein, embodiments of the invention may identify a subset of latent features as relevant or correlative to a classifier’s output in relation to a specimen of interest, and may facilitate high-resolution classification of the studied specimen based on the extracted, relevant latent features.

[0010] Additionally, or alternatively, embodiments of the invention may enable a practitioner to manually exaggerate the value of one or more relevant latent features, thereby affecting corresponding unique, and specific visual traits of the studied specimen. Embodiments of the invention may thus provide further improvement over currently available counterfactual explanation algorithms, by enabling a practitioner to gain additional insight regarding the classifier’s decision.

[0011] Embodiments of the invention have been experimentally applied to biomedical images, and most examples provided herein relate to that field. However, it should be appreciated that the applications discussed herein are provided as non-limiting examples: Embodiments of the invention may be applied to explain any type of ML classifier decision, including classification of images of non-biomedical specimens (e.g., for classifying a product based on its quality, such as operability of an integrated circuit or chip in a fabrication plant), classification of audio data (e.g., for voice recognition or speaker identification), and the like.

[0012] For example, embodiments of the invention have been experimentally demonstrated in the domain of In-Vitro Fertilization (IVF). In this application, a supervised classification model was pretrained directly on images depicting blastocyst specimens, to predict viability of blastocysts for transplant. Embodiments of the invention were applied to the outcome of the supervised classification model in order to provide further insight, and / or refine the classification model’s prediction of viability or non-viability.

[0013] Embodiments of the invention may include a method of classifying a specimen depicted in an image by at least one processor.

[0014] According to some embodiments, the at least one processor may receive a sample image of a specimen, depicting a plurality of visual traits, and may infer a machine-learning (ML) based classifier model on the sample image, to obtain a baseline classification of the specimen. The at least one processor may apply an ML-based generative model on the sample image, to obtain a reconstructed image of the specimen, via a latent feature vector, that may include a plurality of latent features, and disentangle the latent feature vector, to associate each latent feature with zero, one or more unique visual traits of the plurality of visual traits.

[0015] The at least one processor may analyze the reconstructed image, to identify at least one latent feature as correlative to the baseline classification of the specimen, and produce a refined classification of the specimen.

[0016] The refined classification of the specimen may include, for example (i) the baseline classification and (ii) at least one visual trait, associated with the identified at least one latent feature.

[0017] Additionally, or alternatively, the at least one processor may visualize the refined classification via a Graphical User Interface (GUI), by displaying the baseline classification via the GUI; and displaying the reconstructed image via the GUI, while highlighting the at least one visual trait, associated with the identified at least one latent feature.

[0018] According to some embodiments, visualizing the refined classification further may include perturbating the identified latent feature, to produce one or more perturbed latent feature vectors; applying a generative component of the generative model on the one or more perturbed latent feature vectors, to obtain one or more perturbed reconstructed images; and inferring the ML-based classifier model on the one or more perturbed reconstructed images, to obtain one or more respective perturbed classifications. The at least one processor maysubsequently select at least one perturbed reconstructed image based on the one or more perturbed classifications, which may be different from the baseline classification, and display, via the GUI, the selected at least one perturbed reconstructed image, while highlighting the at least one visual trait, associated with the identified at least one latent feature.

[0019] According to some embodiments, the method may further include selecting a target subset of the latent features of the latent feature vector obtained by applying the ML-based generative model on the sample image; inferring a secondary ML-based classifier model on the target subset to obtain a supplementary baseline classification of the specimen; based on the supplementary baseline classification, calculating, for each latent feature of the target subset, a Shapely value representing a degree of contribution of a respective latent feature to the supplementary baseline classification; and visualizing the refined classification via the Graphical User Interface (GUI) further by displaying, for said highlighted at least one visual trait, a visual representation of the calculated Shapely value of the respective latent feature associated with said at least one visual trait.

[0020] According to some embodiments, the specimen may be a biological specimen such as a blastocyst. In such embodiments, the at least one processor may receive a plurality of sample images, corresponding to a plurality of blastocysts; for each sample image, produce a refined classification of the corresponding blastocyst; and screen the plurality of blastocysts for In Vitro Fertilization (IVF), based on the refined classifications.

[0021] Additionally, or alternatively, the specimen may be a biological specimen such as biological cell. For example, receiving the sample image may include performing single cell flow cytometry. The at least one processor may thus receive a plurality of sample images, corresponding to a plurality of biological cells of the single cell flow cytometry process. For each sample image, the at least one processor may produce a refined classification of the corresponding biological cell, screen, or choose among the plurality of biological cells according to at least one criterion, based on the refined classifications. Such a criterion may be application-specific, and may include for example, a change in morphological characteristics of the cell, such as protrusions, or light reflectance properties that can indicate a change in cell stiffness or organelle composition.

[0022] Additionally, or alternatively, the specimen may be an Integrated Circuit (IC) chip. In such embodiments, the at least one processor may receive a plurality of sample images,corresponding to a plurality of chips (e.g., in a production line). For each sample image, the at least one processor may produce a refined classification of the corresponding chip, and screen, or select among the plurality of chips according to at least one criterion, based on the refined classifications.

[0023] According to some embodiments, the at least one processor may apply, or infer the generative model on a training image depicting a specimen, to produce a corresponding reconstructed training image. The at least one processor may calculate a reconstruction loss value, representing a discrepancy between the training image and reconstructed training image. The at least one processor may train the generative model to produce the reconstructed training image while minimizing the reconstruction loss value.

[0024] Additionally, or alternatively, the at least one processor may apply the generative model on a training image depicting a specimen, to produce a corresponding reconstructed training image, inferring the ML-based classifier model on the training image to obtain a first baseline classification of the specimen, and infer the ML-based classifier model on the reconstructed training image to obtain a second baseline classification of the specimen. The at least one processor may subsequently calculate a classification loss value, representing a discrepancy between the first baseline classification and second baseline classification, and train the generative model to produce the reconstructed training image while minimizing the classification loss value.

[0025] According to some embodiments, the at least one processor may disentangle the latent feature vector by: applying the generative model on a first image of a specimen, to produce a first latent feature vector; selecting an index value representing a specific latent feature in the latent feature vector; perturbating the latent feature of the selected index value, to produce a perturbed latent feature vector; obtaining, from an ML-based disentanglement model, a predicted index value, representing prediction of the selected index, based on the perturbed latent feature vector.

[0026] The at least one processor may subsequently calculate a disentanglement loss value, representing a discrepancy between the selected index value and the predicted index value, and train the generative model to produce the reconstructed image of the specimen while minimizing the disentanglement loss value . The at least one processor may thereby uniquely associate the latent feature of the selected index with one or more respective visual traits of the plurality of visual traits.

[0027] According to some embodiments, the at least one processor may obtain the predicted value of the selected index, based on the perturbed latent feature vector, by: applying a generative component of the generative model on the first latent feature vector, to obtain a first reconstructed image; applying the generative component of the generative model on the perturbed latent feature vector, to obtain a second reconstructed image; calculating a difference image, representing difference between the first reconstructed image and the second reconstructed image; and applying the ML-based disentanglement model on the difference image, to predict the value of the selected index.

[0028] According to some embodiments, the at least one processor may apply the generative model on a training image, to produce a respective non-perturbed, reconstructed training image. The at least one processor may select an index of a latent feature of the latent feature vector; perturbate the latent feature of the selected index, to produce a perturbed latent feature vector; and applying a generative component (e.g., a decoder) of the generative model on the perturbed latent feature vector, to obtain a perturbed, reconstructed training image. The at least one processor may calculate a difference image, representing difference between the perturbed reconstructed training image and the non-perturbed, reconstructed training image. Using a value of the selected index as supervisory data, the at least one processor may train the ML-based disentanglement model to predict the index of the perturbated latent feature.

[0029] According to some embodiments, the at least one processor may identify at least one latent feature as correlative to the baseline classification.

[0030] For example, the at least one processor may analyze the reconstructed image by: applying, for at least one latent feature of the latent feature vector, a series of perturbations, to obtain a respective series of perturbed reconstructed images; and inferring the ML based classifier model on one or more perturbed reconstructed images of the series of perturbed reconstructed images, to obtain one or more respective, perturbed classifications of the specimen. The at least one processor may compare the baseline classification of the specimen to the one or more perturbed classifications, and identify the at least one latent feature as correlative to the baseline classification of the specimen, based on said companson.

[0031] Additionally, or alternatively, the at least one processor may constrain a subset or group of latent features of the latent feature vector to be correlative, or affect the baseline classification.

[0032] For example, the at least one processor may apply the generative model on a training image depicting a specimen, to produce a corresponding reconstructed training image. The at least one processor may infer the ML-based classifier model on the training image to obtain a first baseline classification of the specimen. The at least one processor may select a subset of the latent features of the latent feature vector, and inferring a secondary ML-based classifier model on the subset of the latent features, to obtain a second baseline classification of the sample image. The secondary ML-based classifier model may, for example, be a single -neuron classifier, trained to produce a baseline classification of the depicted specimen, based on the subset of latent features.

[0033] The at least one processor may calculate a secondary classification loss value, representing a discrepancy between the first baseline classification and second baseline classification, and train the generative model to produce the reconstructed training image while minimizing the secondary classification loss value.

[0034] According to some embodiments, the at least one processor may apply the series of perturbations only on one or more latent features of the latent feature subset, e.g., on latent features that are identified as pertinent to the baseline classification.

[0035] According to some embodiments, the at least one processor may applying the generative model on a plurality of training images, to produce a corresponding plurality of latent feature vectors. Based on the plurality of latent feature vectors, the at least one processor may calculate a covariance matrix, representing covariance among the latent features. The at least one processor may calculate a covariance loss value, representing a discrepancy between the covariance matrix and a unity matrix of the same dimension, and subsequently train the generative model to produce the reconstructed training image while minimizing the covariance loss value.

[0036] Embodiments of the invention may include a system for classifying a specimen depicted in an image. Embodiments of the system may include a non-transitory memory device, wherein modules of instruction code may be stored, and at least one processor associated with the memory device, and configured to execute the modules of instruction code. Upon execution of said modules of instruction code, the at least one processor may beconfigured to: receive a sample image of a specimen, depicting a plurality of visual traits; infer an ML based classifier model on the sample image, to obtain a baseline classification of the specimen; apply an ML-based generative model on the sample image, to obtain a reconstructed image of the specimen, via a latent feature vector, may include a plurality of latent features; disentangle the latent feature vector, to associate each latent feature with zero, one or more unique visual traits of the plurality of visual traits; analyze the reconstructed image, to identify at least one latent feature as correlative to the baseline classification of the specimen; and produce a refined classification of the specimen, may include (i) the baseline classification and (ii) at least one visual trait, associated with the identified at least one latent feature.

[0037] Embodiments of the invention may include a method of screening biological specimens by at least one processor. Embodiments of the method may include: receiving a sample image depicting a biological specimen; providing an autoencoder model, trained to generate a reconstructed version of the sample image via a latent feature vector, wherein the reconstructed version depicts one or more visual phenotypes of the biological specimen; providing a predictor ML model configured to constrain the latent feature vector, such that at least one latent feature of latent feature vector may be linked by injective association to a respective visual phenotype of the one or more visual phenotypes, depicted in the reconstructed version; and screening the biological specimen based on the at least one latent feature.BRIEF DESCRIPTION OF THE DRAWINGS

[0038] The subject matter regarded as the invention is particularly pointed out and distinctly claimed in the concluding portion of the specification. The invention, however, both as to organization and method of operation, together with objects, features, and advantages thereof, may best be understood by reference to the following detailed description when read with the accompanying drawings in which:

[0039] Fig. 1 is a block diagram, depicting a computing device which may be included in a system for classifying depicted specimens according to some embodiments of the invention;

[0040] Fig. 2 is a block diagram, depicting an example for implementation of a system for classifying depicted specimens according to some embodiments of the invention;

[0041] Figs. 3A-3C are examples of sample images, in this case - depicting a blastocyte, having a plurality of visual traits and properties;

[0042] Fig. 4A is a block diagram, depicting an example of operation of a training module, which may be included in the system for classifying depicted specimens, according to some embodiments of the invention;

[0043] Fig. 4B is a block diagram, depicting another example of operation of the training module, according to some embodiments of the invention;

[0044] Fig. 4C is a block diagram, depicting another example of operation of the training module, according to some embodiments of the invention;

[0045] Fig. 4D is a block diagram, depicting another example of operation of the training module, according to some embodiments of the invention;

[0046] Fig. 4E is a block diagram, depicting yet another example of operation of the training module, according to some embodiments of the invention;

[0047] Fig. 5 is a flow diagram depicting a method of classifying depicted specimens by at least one processor, according to some embodiments of the invention; and

[0048] Fig. 6 is a flow diagram depicting a method of screening biological specimens by at least one processor, according to some embodiments of the invention.

[0049] It will be appreciated that for simplicity and clarity of illustration, elements shown in the figures have not necessarily been drawn to scale. For example, the dimensions of some of the elements may be exaggerated relative to other elements for clarity. Further, where considered appropriate, reference numerals may be repeated among the figures to indicate corresponding or analogous elements.DETAILED DESCRIPTION OF THE PRESENT INVENTION

[0050] One skilled in the art will realize the invention may be embodied in other specific forms without departing from the spirit or essential characteristics thereof. The foregoing embodiments are therefore to be considered in all respects illustrative rather than limiting of the invention described herein. Scope of the invention is thus indicated by the appended claims, rather than by the foregoing description, and all changes that come within the meaning and range of equivalency of the claims are therefore intended to be embraced therein.

[0051] In the following detailed description, numerous specific details are set forth in order to provide a thorough understanding of the invention. However, it will be understood by those skilled in the art that the present invention may be practiced without these specific details. In other instances, well-known methods, procedures, and components have not beendescribed in detail so as not to obscure the present invention. Some features or elements described with respect to one embodiment may be combined with features or elements described with respect to other embodiments. For the sake of clarity, discussion of same or similar features or elements may not be repeated.

[0052] Although embodiments of the invention are not limited in this regard, discussions utilizing terms such as, for example, “processing,” “computing,” “calculating,” “determining,” “establishing”, “analyzing”, “checking”, or the like, may refer to operation(s) and / or process(es) of a computer, a computing platform, a computing system, or other electronic computing device, that manipulates and / or transforms data represented as physical (e.g., electronic) quantities within the computer’s registers and / or memories into other data similarly represented as physical quantities within the computer’s registers and / or memories or other information non-transitory storage medium that may store instructions to perform operations and / or processes.

[0053] Although embodiments of the invention are not limited in this regard, the terms “plurality” and “a plurality” as used herein may include, for example, “multiple” or “two or more”. The terms “plurality” or “a plurality” may be used throughout the specification to describe two or more components, devices, elements, units, parameters, or the like. The term “set” when used herein may include one or more items.

[0054] Unless explicitly stated, the method embodiments described herein are not constrained to a particular order or sequence. Additionally, some of the described method embodiments or elements thereof can occur or be performed simultaneously, at the same point in time, or concurrently.

[0055] Reference is now made to Fig. 1, which is a block diagram depicting a computing device, which may be included within an embodiment of a system for classifying depicted specimens, according to some embodiments.

[0056] Computing device 1 may include a processor or controller 2 that may be, for example, a central processing unit (CPU) processor, a chip or any suitable computing or computational device, an operating system 3, a memory 4, executable code 5, a storage system 6, input devices 7 and output devices 8. Processor 2 (or one or more controllers or processors, possibly across multiple units or devices) may be configured to cany out methods described herein, and / or to execute or act as the various modules, units, etc. Morethan one computing device 1 may be included in, and one or more computing devices 1 may act as the components of, a system according to embodiments of the invention.

[0057] Operating system 3 may be or may include any code segment (e.g., one similar to executable code 5 described herein) designed and / or configured to perform tasks involving coordination, scheduling, arbitration, supervising, controlling or otherwise managing operation of computing device 1, for example, scheduling execution of software programs or tasks or enabling software programs or other modules or units to communicate. Operating system 3 may be a commercial operating system. It will be noted that an operating system 3 may be an optional component, e.g., in some embodiments, a system may include a computing device that does not require or include an operating system 3.

[0058] Memory 4 may be or may include, for example, a Random-Access Memory (RAM), a read only memory (ROM), a Dynamic RAM (DRAM), a Synchronous DRAM (SDRAM), a double data rate (DDR) memory chip, a Flash memory, a volatile memory, a nonvolatile memory, a cache memory, a buffer, a short term memory unit, a long term memory unit, or other suitable memory units or storage units. Memory 4 may be or may include a plurality of possibly different memory units. Memory 4 may be a computer or processor non-transitory readable medium, or a computer non-transitory storage medium, e.g., a RAM. In one embodiment, a non-transitory storage medium such as memory 4, a hard disk drive, another storage device, etc. may store instructions or code which when executed by a processor may cause the processor to carry out methods as described herein.

[0059] Executable code 5 may be any executable code, e.g., an application, a program, a process, task, or script. Executable code 5 may be executed by processor or controller 2 possibly under control of operating system 3. For example, executable code 5 may be an application that may classify depicted specimens as further described herein. Although, for the sake of clarity, a single item of executable code 5 is shown in Fig. 1, a system according to some embodiments of the invention may include a plurality of executable code segments similar to executable code 5 that may be loaded into memory 4 and cause processor 2 to carry out methods described herein.

[0060] Storage system 6 may be or may include, for example, a flash memory as known in the art, a memory that is internal to, or embedded in, a micro controller or chip as known in the art, a hard disk drive, a CD-Recordable (CD-R) drive, a Blu-ray disk (BD), a universal serial bus (USB) device or other suitable removable and / or fixed storage unit. Datapertaining to images of depicted specimens may be stored in storage system 6 and may be loaded from storage system 6 into memory 4 where it may be processed by processor or controller 2. In some embodiments, some of the components shown in Fig. 1 maybe omitted. For example, memory 4 may be a non-volatile memory having the storage capacity of storage system 6. Accordingly, although shown as a separate component, storage system 6 may be embedded or included in memory 4.

[0061] Input devices 7 may be or may include any suitable input devices, components, or systems, e.g., a detachable keyboard or keypad, a mouse and the like. Output devices 8 may include one or more (possibly detachable) displays or monitors, speakers and / or any other suitable output devices. Any applicable input / output (I / O) devices may be connected to Computing device 1 as shown by blocks 7 and 8. For example, a wired or wireless network interface card (NIC), a universal serial bus (USB) device or external hard drive may be included in input devices 7 and / or output devices 8. It will be recognized that any suitable number of input devices 7 and output device 8 may be operatively connected to Computing device 1 as shown by blocks 7 and 8.

[0062] A system according to some embodiments of the invention may include components such as, but not limited to, a plurality of central processing units (CPU) or any other suitable multi-purpose or specific processors or controllers (e.g., similar to element 2), a plurality of input units, a plurality of output units, a plurality of memory units, and a plurality of storage units.

[0063] The term neural network (NN) or artificial neural network (ANN), e.g., a neural network implementing a machine learning (MU) or artificial intelligence (Al) function, may be used herein to refer to an information processing paradigm that may include nodes, referred to as neurons, organized into layers, with links between the neurons. The links may transfer signals between neurons and may be associated with weights. A NN may be configured or trained for a specific task, e.g., pattern recognition or classification. Training a NN for the specific task may involve adjusting these weights based on examples. Each neuron of an intermediate or last layer may receive an input signal, e.g., a weighted sum of output signals from other neurons, and may process the input signal using a linear or nonlinear function (e.g., an activation function). The results of the input and intermediate layers may be transferred to other neurons and the results of the output layer may be provided as the output of the NN. Typically, the neurons and links within a NN are represented bymathematical constructs, such as activation functions and matrices of data elements and weights. At least one processor (e.g., processor 2 of Fig. 1) such as one or more CPUs or graphics processing units (GPUs), or a dedicated hardware device may perform the relevant calculations.

[0064] Reference is now made to Fig. 2, which is a block diagram depicting an example for implementation of a system 10 for classifying depicted specimens, according to some embodiments of the invention.

[0065] According to some embodiments of the invention, system 10 may be implemented as a software module, a hardware module, or any combination thereof. For example, system may be or may include a computing device such as element 1 of Fig. 1, and may be adapted to execute one or more modules of executable code (e.g., element 5 of Fig. 1) to classify depicted specimens, as further described herein.

[0066] As shown in Fig. 2, arrows may represent flow of one or more data elements to and from system 10 and / or among modules or elements of system 10. Some arrows have been omitted in Fig. 2 for the purpose of clarity.

[0067] According to some embodiments, system 10 may receive (e.g., via input device 7 of Fig. 1) at least one sample image 20 which depicts, or represents a specimen 20’.

[0068] Depicted specimen 20’ may be characterized by one or more (e.g., a plurality) of visual traits or properties 20T. In other words, image 20 may include depiction of one or more traits or properties 20T, that characterize depicted specimen 20’.

[0069] Reference is also made to Figs. 3A-3C, which are examples of a sample image 20. As shown in Figs. 3A-3C the depicted specimen is a blastocyte, having a plurality of visual traits or properties. In this non-limiting example, visual traits 20T include criteria of blastocy st quality, according to the Gardner three-part scoring scheme, as known in the art. Each of Figs. 3A-3C highlights a specific such visual trait 20T:

[0070] In Fig. 3A the highlighted visual trait 20T is a Blastocyst Expansion (BE) property (e.g., a degree of expansion of the blastocyst cavity)- In Fig. 3B the highlighted visual trait 20T is a Morphology of the Inner Cell Mass (ICM) (e.g., size and compaction of the mass of cells that eventually form a fetus). In Fig. 3C the highlighted visual trait 20T is a morphology of the trophectoderm (TE) (e.g., a number and cohesiveness of a single cell layer on the outer edge of the blastocyst that eventually forms a placenta).

[0071] As shown in Fig. 2, system 10 may include a machine-learning (ML) based classification model, or regression model, also referred to herein as classifier model 50. classifier model 50 may be pretrained to classify the image 20 (depicted specimen 20’) according to one or more predefined classes or groups.

[0072] Pertaining to the blastocyst example of Figs. 3A-3C, classifier model 50 may be pretrained, e.g., based on expert-provided annotations as known in the art, to classify the depicted specimen (blastocyst) 20' according to its quality or viability for hi Vitro Fertilization (TVF) transplant.

[0073] According to some embodiments, system 10 may infer classifier model 50 on sample image 2.0, to obtain a baseline classification 50C of the specimen.

[0074] Additionally, system 10 may include a refinement module 60, As elaborated herein, refinement module 60 may further analyze sample image 20, to obtain a refined classification 60C of the depicted specimen 20’, in relation to one or more depicted visual traits 20T. Classification 60C may be referred to herein as “refined” in a sense that it may provide additional information that may enhance, explain, or fine-tune baseline classification 50C.

[0075] Pertaining to the blastocyst example of Figs. 3A-3B, a baseline classification 50C may include, for example, classification or indication of a depicted cell as non-viable for IVF transplant. In this example, refined classification 60C may include classification or attribution of the blastocyst to a class that explains an underlying reason forthat non-viability baseline classification 50C, in relation to one or more specific traits 20T depicted in image 20.

[0076] For example, a refined classification 60C to the “non-viable” baseline classification 50C may be “non-viable because of an inappropriate value of a specific visual trait 20T (e.g., BE)”.

[0077] As elaborated herein system 10 may be configured to present refined classification 60C via a user interface (UI) (e.g., output device 8 of Fig. 1), as a textual message. Additionally, or alternatively, system 10 may present refined classification 60C via the UI 8 as a highlighted feature of specimen 20’, e.g., as seen in Figs. 3A-3C.

[0078] It may be appreciated that such refined classification 60C, and / or presentation thereof to a human expert, may provide additional insight and understanding regarding the underlying cause of baseline classification 50C.

[0079] For example, specimen 20’ may be a depicted biological cell, taken by biopsy from a patient, and baseline classification 50C may include classification of that biological cell as cancerous. Refined classification 60C may explain baseline classification 50C, so as to highlight a specific visual trait 20T (e.g., a specific morphology or phenotype) that has contributed to the classification of the depicted specimen (biological cell) 20’ as cancerous. Refined classification 60C may thus facilitate high-resolution understanding or segmentation of cancerous cells based on specific contributing visual features or traits 20T, and may allow high resolution diagnosis of specimen 20’ and / or treatment of the relevant patient.

[0080] As shown in Fig. 2, system 10 may include an ML-based generative model 100, such as an autoencoder model 100. As known in the art, a generative model 100 (e.g., autoencoder) may be trained to represent probabilistic distribution of features that describe objects within a given dataset. During an inference stage, specific values or weights of the generative model may be set to generate new objects of a similar distribution.

[0081] According to some embodiments, generative model 100 (e.g., autoencoder) may include a first portion referred to herein as an encoding portion, or encoder 120, and a second portion referred to herein as a generative portion, or decoder 140. As known in the art, encoder 120 may be configured to encode incident sample data (e.g., images 20) into a latent feature vector 130LV representation, in a reduced-dimension latent space. Generative portion (decoder) 140 may be configured to decode the latent feature vectors 130LV, so as to produce a reconstructed version (e.g., image 30REC) of the incident data (e.g., image 20).

[0082] According to some embodiments, generative model 100 (e.g., autoencoder) may be trained to receive an image 20 of a specimen 20’, and produce a reconstructed version 30REC of image 20, via a latent vector 130LV, having a plurality of latent features 130LF. Latent vector 130LV may be characterized by a reduced dimension (e.g., number of latent features 130LF), in relation to a dimension of image 20 (e.g., number of pixels in image 20).

[0083] Therefore, reconstructed version 30REC may be regarded as filtered by a latent space 130LS defined by the plurality of latent features 130LF. Moreover, by applying a change to one or more latent features 130LF, generative model 100 may produce one or more respective, different reconstructed versions 30REC of sample image 20.

[0084] As shown in Fig. 2, system 10 may include a training module 40, configured to train generative model 100, as elaborated herein. It may be appreciated that training module 40ofFig. 2 maybe the same as training module 40 of Figs. 4A-4E. Similarly, system 10 of Fig. 2 may be the same as system 10 of Figs. 4A-4E. Some modules and / or signals were omitted in these drawings for the purpose of clarity.

[0085] Reference is also made to Fig. 4A, which is a block diagram depicting an example of operation of a training module 40, which may be included in system 10 for classifying depicted specimens, according to some embodiments of the invention. Elements of system 10 in Fig. 4A may be the same as corresponding (e.g., equally enumerated) elements ofFig. 2.

[0086] As shown in the example ofFig. 4A, training module 40 may be configured to train generative model 100 while minimizing image reconstruction loss 405L, and maintaining classification consistency. In other words, generative model 100 may be trained to reconstruct (image 30REC) the original input data (image 20) with high perceptual quality. This implicitly means that the latent representation encodes the imperative features of the data distribution as well as the factors of variation which influence the decision 50C of classification model 50. By ensuring minimal discrepancy between scores or classifications 50C of classifier model 50 that pertain to original images 20 and reconstructed versions 30REC thereof, system 10 may produce reconstructed images 30REC that contain the visual information used by classifier 50.

[0087] According to some embodiments, during a training stage, training module 40 may apply the generative model on a training image 20 depicting a specimen 20’ (e.g., a biological cell), to produce a corresponding reconstructed training 30REC image of that specimen 20’. Training module 40 may include a reconstruction computation module 405, configured to calculate a loss term value, also referred to herein as a reconstruction loss value 405L. Reconstruction loss value 405L may represent discrepancy or difference between the training image 20 and the reconstructed version 30REC of training image 20. Training module 40 may provide reconstruction loss value 405L as feedback to generative model 100, so as to train the generative model 100 to produce reconstructed training image 30REC, while minimizing reconstruction loss value 405L. In other words, training module 40 may train generative model 100 to minimize discrepancy or difference between training image 20 and reconstructed image 30REC.

[0088] Additionally, or alternatively, generative model 100 may be trained to produce image 30REC such that it maintains information that is correlative to the classification of ML model 50.

[0089] For example, during a training stage, system 10 may infer the ML-based classifier model 50 on the training image 20 to obtain a first baseline classification 50C of the specimen 20’, and infer the ML-based classifier model 50 on the reconstructed training image 30REC, to obtain a second baseline classification 50C of the specimen.

[0090] Training module 40 may include a classification computation module 415, configured to calculate a loss term value, also referred to herein as a classification loss value 415L. Classification loss value 415L may represent discrepancy or difference between the first baseline classification (of image 20) and the second baseline classification (of reconstructed image 30REC).

[0091] Training module 40 may provide classification loss value 415L as feedback to generative model 100, so as to train the generative model 100 to produce reconstructed training image 30REC, while minimizing classification loss value 415L. In other words, training module 40 may train generative model 100 to minimize discrepancy or difference between ML model’s 50 classification of image 20, and ML model’s 50 classification of reconstructed image 30REC.

[0092] According to some embodiments, system 10 may (e.g., during an inference stage) apply ML-based generative model 100 on a sample image 20 to obtain a respective reconstructed image 30REC of the specimen 20’, via latent feature vector 130LV (e.g., via the plurality of latent features 130LF).

[0093] Reference is now made to Fig. 4B, which is a block diagram depicting another example of operation of the training module 40, according to some embodiments of the invention. Elements of system 10 in Fig. 4B may be the same as corresponding (e.g., equally enumerated) elements of Fig. 2.

[0094] As elaborated herein, training module 40 may be configured to disentangle the latent feature vector 130LV, thereby associating each latent feature 130LF with zero, one, or more unique visual traits of the plurality of visual traits.

[0095] As shown in Fig. 4B, training module 40 may include an ML-based disentanglement model 430, that may be trained to receive a perturbed latent feature vector 130LVP, and identify which latent feature 130LF of perturbed latent feature vector 130LVP was altered,by computing an input discrepancy (e.g., a difference image) between: a) the unaltered reconstructed image 30REC; and b) a reconstructed image 30RECP based on altered latent feature vector 130LVP.

[0096] Training module 40 may be configured to calculate a loss term value, referred to herein as a disentanglement loss value 430L. Disentanglement loss value 430L may represent an error of disentanglement model 430 in prediction or identification of the altered latent feature 130LF of perturbed latent feature vector 130LVP.

[0097] As elaborated herein, Training module 40 may utilize disentanglement loss value 430L to force disentanglement of the latent vector representation. By minimizing disentanglement loss value 430L (e.g., minimizing the error of disentanglement model 430), training module 40 may constrain each latent feature 130LF to encode visual image features or visual traits 20T that are visually distinct from other latent features 130LF, thus obtaining disentanglement of the latent representation 130LS.

[0098] It may be appreciated that such disentanglement may allow simpler traversal of the latent space 130LS, one latent feature 130LF at a time, under the assumption that each feature will encode independent, classification 50C-driving image features or visual traits 20T. It has been experimentally observed that such feature disentanglement may also push the latent representation 130LS such that each latent feature 130LF will tend to encode a single visual image feature or visual trait 20T.

[0099] According to some embodiments, during a training stage, training module 40 may apply the generative model on a training image 20 depicting a specimen 20’ (e.g., a biological cell), to produce a corresponding reconstructed training 30REC image of that specimen 20’, via a first latent feature vector 130LV. Reconstructed training image 30REC may be referred to herein as non-perturbed, in a sense that it may be derived directly from training image 20, without having any changes or perturbation applied to latent feature vector 130LV, as elaborated herein.

[0100] Training module 40 may select at least one index value 130LFS representing at least one respective, specific latent feature 130LF in the latent feature vector 130LV. For example, training module 40 may select the at least one latent feature 130LF (having index 130LFS) randomly, or in an ordered manner (e.g., according to their indices in latent feature vector 130LV).

[0101] Additionally, or alternatively, and as elaborated herein (e.g., in relation to Fig. 4E), training module 40 may select the at least one latent feature 130LF (having index value 130LFS) from a subset of latent features 130LF of latent feature vector 130LV that are correlative, or relevant to the classification 50C of classifier ML model 50.

[0102] According to some embodiments, training module 40 may include a perturbation module 410, configured to perturbate the at least one latent feature 130LF of the selected index values 130LFS, to produce a perturbed latent feature vector 130LVP.

[0103] For example, perturbation module 410 may perturbate the at least one (or exactly one) latent feature 130LVP by setting one or more values to the selected latent feature 130LF, to obtain one or more respective perturbed latent feature vectors 130LVP. The set values may be (i) different from the selected latent feature’s 130LF value (of index 130LFS) pertaining to the incident training image 20, and (ii) within a range of values as defined by the distribution of the latent feature’s 130LF values, corresponding to the plurality of training images 20 in the training dataset.

[0104] As shown in Fig. 4B, training module 40 may include a generative portion 140’ that may be the same instance as, or a copied instance of generative portion (e.g., decoder) 140 of ML-based generative model 100. Training module 40 may apply, or infer generative portion 140’ on the one or more perturbed latent feature vectors 130LVP to obtain one or more respective, perturbed, reconstructed training images 30RECP. In other words, generative portion 140’ may decode the one or more perturbed latent feature vectors 130LVP (e.g., in a similar manner to that of generative portion 140) to produce or generate one or more respective, perturbed versions 30RECP of reconstructed image 30REC.

[0105] As shown in Fig. 4B, training module 40 may include a differentiation module 420, configured to calculate a difference or discrepancy 420D between reconstructed image 30REC, and a perturbed image 30RECP (of the one or more perturbed versions of reconstructed image 30REC) based on a predefined metric.

[0106] For example, the predefined metric may include an absolute value of pixel-based difference between images 30REC and 30RECP. In other words, differentiation module 420 may calculate discrepancy 420D as a difference image, representing difference (e.g., pixelbased difference) between the perturbed reconstructed training image 30RECP and the nonperturbed, reconstructed training image 30REC. Alternatively, differentiation module 420may calculate discrepancy 420D by using both images, without directly calculating a difference image.

[0107] Additionally, or alternatively, differentiation module 420 may calculate discrepancy 420D based on another predefined metric, such as a difference in a value of one ormore features that may be extracted from images 30REC and 30RECP. Additional options for calculation of discrepancy 420D value may also be possible.

[0108] According to some embodiments, disentanglement model 430 may be trained to identify a value of an index of a perturbed latent feature 130LF in a perturbed latent feature vector, based on the perturbation’s effect on the outcome, perturbed image 30RECP.

[0109] For example, during a training stage, disentanglement model 430 may be trained to receive discrepancy 420D (e.g., a difference image), and predict, or produce a prediction 430P of an index of a perturbated latent feature 130LF in perturbated latent feature vector 130LVP, based on the received discrepancy 420D data. Disentanglement model 430 may use the value of the selected index 130LFS as supervisory data for this training.

[0110] According to some embodiments, the training of disentanglement model 430 and generative model 100 may be intermittent, or iterative (e.g., repetitive over time). For example, a first training iteration may include training of the generative model 100 to produce reconstructed image 30REC as elaborated herein (e.g., in relation to Fig. 4A), where generative portion 140’ may be updated. A subsequent training iteration may include training of the disentanglement model 430, where prediction 430P of a value of an index of a perturbed latent feature 130LF may be learnt or refined.

[0111] The term “disentangle” may be used herein to refer to a separation and / or prioritization among latent features 130LF of latent feature vector 130LV, in relation to visual traits of the depicted specimen 20’.

[0112] For example, and as elaborated herein (e.g., in relation to Fig. 4E) a disentangled latent feature vector 130LV may include separation of member latent features 130LF, based on effect that these latent features 130LF may have on classification 50C. Pertaining to the example of Figs. 3A-3C, a subset of latent features 130LF that affect general aspects of image 20 (e.g., brightness, contrast, etc.) may not have any effect on classification 50C (e.g., viability of blastocyst) and may be indicated as having low relevance for determining refined classification 60C. In contrast, a subset of latent features 130LF that affect traits 20T (e.g.,BE, ICM, TE) of specimen 20’ may be indicated as having low relevance for determining refined classification 60C.

[0113] In another example, a disentangled latent feature vector 130LV may include separation of member latent features 130LF, such that each latent feature 130LF may be associated with zero, one, or more unique visual traits of the plurality of visual traits.

[0114] The term “unique” may be used herein as demonstrated by the example of Figs. 3A-3C: A value of a first latent feature 130LF may uniquely affect appearance of a first visual trait 20T (e.g., BE), and a value of a second latent feature 130LF may uniquely affect appearance of one or more second visual traits 20T (e.g., ICM, TE). However, a single visual trait 20T (e.g., ICM) may not be affected by more than one latent feature 130LF.

[0115] The term “affect” may be used in this context in a sense that a change in latent feature 130LF will affect appearance of a corresponding unique visual trait 20T. Pertaining to the example of Figs. 3A, a change in a single latent feature 130LF will uniquely affect visual trait 20T by changing (e.g., extending or shortening) the depicted line representing BE.

[0116] According to some embodiments, training module 40 may use prediction 430P of disentanglement model 430 to further train generative model 100 to disentangle between latent features 130LF of latent feature vector 130LV.

[0117] During a training iteration of generative model 100, training model 40 may select 130LFS, and perturb a latent feature 130LF of the latent feature vector 130LV, to obtain a respective perturbed latent feature vector 130LVP. Based on the perturbed latent feature vector 130LVP, training model 40 may obtain (i) a respective perturbed reconstructed image 30RECP and (ii) a subsequent discrepancy data element 420D as elaborated herein. For example, system 10 may apply a generative component 140 of the generative model 100 on the first, unperturbed latent feature vector 130LV to obtain a first reconstructed image 30REC, and apply a generative component 140 of the generative model, or a copied instance 140’ thereof on the perturbed latent feature vector 130LVP, to obtain a second reconstructed image 30RECP, and calculate a discrepancy data element 420D (e.g., difference image), representing difference 420D between the first reconstructed image 30REC and second reconstructed image 30RECP.

[0118] Training model 40 may subsequently infer or apply (now trained) ML-based disentanglement model 430 on discrepancy data element 420D (e.g., the difference image) to predicted a value of the selected index 130LFS.

[0119] According to some embodiments, training model 40 may calculate a loss term value, also referred to herein as a disentanglement loss value 430L, representing discrepancy between the selected index value 130LFS and the predicted index value 430P (e.g., prediction of selected index 130LFS).

[0120] Training module 40 may feed disentanglement loss value 430L back to generative model 10. Training module 40 may thus train generative model 10 to produce the reconstructed image 30REC of specimen 20’ while minimizing the disentanglement loss value 430L.

[0121] It may be appreciated that by minimizing the disentanglement loss value 430, training module 40 may uniquely associate the latent feature 130LF of the selected index 130LFS with zero, one or more unique, respective visual traits 20T of the plurality of visual traits 20T of specimen 20’.

[0122] In other words, by training generative model 100 to minimize disentanglement loss value 430L, training model 40 may force disentanglement among latent features 130LF of latent feature vector 130LV, in a sense that each latent feature 130LF may encode, or be associated with a respective, identifiable visual pattern or visual trait 20T within image 20.

[0123] According to some embodiments, system 10 may analyze reconstructed image 30REC and / or latent feature vector 130LV, to identify at least one latent feature 130LF as correlative to the baseline classification 50C of the specimen 20’. Additionally, or alternatively, system 10 may facilitate analysis of reconstructed image 30REC and / or latent feature vector 130LV (e.g., by a human expert via a user interface (UI) such as input / output devices 7 and 8 of Fig. 1), to identify correlation, and associate between at least one latent feature 130LF and baseline classification 50C of the specimen 20’.

[0124] Such identification of correlation may be performed by perturbing latent feature vector 130LV and identifying specific latent feature 130LF values that affect baseline classification 50C.

[0125] In some embodiments, system 10 may gradually increase magnitude of these perturbations to obtain a respective series of perturbed reconstructed images 30RECP that exaggerate subtle image properties or visual traits 20T that drive classification 50C (e.g.,apply monotonic change in the classifier inference according to the magnitude of the latent feature 130LF perturbation) while minimizing other image variabilities. Such series of perturbed “in silico” images may then be used to manually, or automatically identify visual traits 20T (e.g., clinical or physical visual properties) that were correlated with the classification as underlying “cause” of classification 50C.

[0126] For example, perturbation module 410 may, for at least one latent feature 130LF of latent feature vector 130LV, apply a series of perturbations, to obtain a respective series of perturbed reconstructed images 30RECP. System 10 may infer ML-based classifier model 50 on one or more (e.g., each) perturbed reconstructed images 30RECP of the series of perturbed reconstructed images 30REC, thereby obtaining one or more respective, perturbed classifications 50CP (Fig. 2) of the specimen. The term “perturbed” may be used in this context to indicate a classification, or classification score 50C that may be derived from a perturbed, or changed latent feature vector 130LVP, and may therefore be different from a classification, or classification score 50C that originated from an original, or un-perturbed latent feature vector 130LV.

[0127] According to some embodiments, refinement model 60 of Fig. 2 may be configured to compare baseline classification 50C of the specimen to the one or more perturbed classifications 50CP, to identify at least one perturbed latent feature 130LF as correlative to the baseline classification 50C / 50CP of the specimen. This identified latent feature 130LF is denoted in Fig. 2 as element 60CLF.

[0128] Additionally, or alternatively, baseline classification 50C may include a score, or confidence value. As known in the art, such score, or confidence value may represent an extent to which a specific specimen may be regarded as pertaining to a specific class of classification 50C. In such embodiments, refinement model 60 may be configured to identify at least one perturbed latent feature 130LF as correlative to the baseline classification 50C / 50CP of the specimen (e.g., 60CLF) by calculating correlation of latent feature values with corresponding classifier scores or confidence values of baseline classification 50C, and determining the most correlative latent features 60CLF as those corresponding to extremum (e.g., maximal or minimal) correlation with corresponding classifier scores.

[0129] For example, refinement model 60 of Fig. 2 may be configured to find at least one image in a series of perturbed images 30RECP, in which the classification 50CP is changed in relation to classification 50C due to a change in a specific latent feature 130LFvalue, and may subsequently label that latent feature 130LF (now 60CLF), and / or that latent feature 130LF value as correlative, or pertinent to the decision of classification model 50.

[0130] Additionally, or alternatively, and as elaborated herein, refinement model 60 of Fig. 2 may allow a user to apply a change (e.g., via input 7 of Fig. 1) in value of one or more specific latent feature 130LF, present the subsequent perturbed image 30RECP (e.g., via output 8 of Fig. 1) adjoint with, or without classification 50C, and allow an expert user to mark, or attribute the relevant changed latent feature 130LF as pertinent to classification 50C.

[0131] According to some embodiments, refinement model 60 of Fig. 2 may subsequently produce a refined classification 60C of specimen 20’, which may include amalgamation of information of baseline classification 50C and at least one visual trait 20T, associated with the identified at least one latent feature 130LF (now 60CLF).

[0132] Pertaining to the example of Fig. 3B, refinement model 60 may identify a specific latent feature 130LF (now 60CLF) as correlative to, pertaining to, or affecting classification 50C, and that specific latent feature may, in turn, be associated with a specific visual trait 20T (e.g., ICM). Therefore, baseline classification 50C may categorize a specific specimen (blastocyst) of interest as “non-viable”, and a refined classification 60C to the “non-viable” baseline classification 50C may be “'non-viable because of the ICM value 20T surpassing a specific value” (e g., a value in which classification 50C of perturbed image 30RECP has changed, e.g., from “viable” to “non-viable”).

[0133] Reference is now made to Fig. 4C, which is a block diagram depicting yet another example of operation of the training module 40, according to some embodiments of the invention. Elements of system 10 in Fig. 4C may be the same as corresponding (e.g., equally enumerated) elements of Fig. 2.

[0134] As shown in Fig. 4C, training module 40 may include a random noise generator module 440, configured to create a random noise vector 440V of the same dimension as latent feature vector 130LV. Additionally, or alternatively, training module 40 may include an ML-based discriminator model 450, trained to receive a data vector, and classify received data vector as either a random noise vector 440V (e.g., originating from random noise generator module 440) or a latent feature vector 130LV (e.g., originating from an image 20 depicting a specimen 20’ of interest). This classification is denoted as element 450L.

[0135] During training of generative model 100, discriminator model 450 of training module 40 may produce classification 450L as a loss term value, also referred to herein as an adversarial loss value 450L. Training module 40 may provide adversarial loss value 450L as feedback to generative model 100, and train generative model 100 while minimizing a value of adversarial loss value 450L. It may be appreciated by a person skilled in the art that by doing so, training module 40 may regularize the latent space of latent feature vectors 130LV as having a probabilistic distribution. In other words, the latent space of generative model 100 may now be a stochastic space, having minimal sparsity of the latent space to sample from.

[0136] As elaborated herein, disentanglement model 430 may be configured to associate between a single latent feature 130LF and a unique set of one or more visual traits 20T (e.g., where a single visual trait 20T (e.g., ICM) of image 20 may not be affected by more than one latent feature 130LF).

[0137] Additionally, or alternatively, training model 40 may apply a similar constraint in the opposite direction, e.g., where a single latent feature 130LF affects exactly one, unique visual trait 20T of image 20. In other words, training model 40 may enforce an injective relation between latent features 130LF and respective unique visual traits 20T.

[0138] Reference is now made to Fig. 4D, which is a block diagram depicting another example of operation of the training module 40, according to some embodiments of the invention. Elements of system 10 in Fig. 4D may be the same as corresponding (e.g., equally enumerated) elements of Fig. 2.

[0139] As shown in Fig. 4D, training module 40 may include a covariance whitening module 460, configured to create a covariance matrix 460M of latent features 130LF. As known in the art, covariance matrix 460M may be a rectangular matrix having the same dimension as latent feature vector 130LV.

[0140] Covariance matrix 460M may represent covariance among latent features 130LF, and may be gradually calculated overtime, e.g., based on values of incoming latent features 130LF, originating from respective input images 20.

[0141] In other words, training module 40 may apply generative model 100 on a plurality of training images 20, to produce a corresponding plurality of latent feature vectors 130LV, and may calculate covariance matrix 460M based on the plurality of latent feature vectors 130LV, such that each entry (i, j) of covariance matrix 460M may represent a covariancevalue (e.g., in the region [-1, 1]) between latent feature 130LF (i) and latent feature 130LF 0).

[0142] During a training stage, training module 40 may calculate another loss term value, referred to herein as a covariance loss or whitening loss value 460L, representing discrepancy between the covariance matrix and a unitary matrix of the same dimension (e.g., where elements along the diagonal are ‘ 1 ’, and all others are ‘0’).

[0143] For example, training module 40 may (a) calculate an average of the values of covariance matrix 460M on the diagonal indices as a scalar value, denoted herein as "avg_diag"; (b) calculate an average of off-diagonal indices of covariance matrix 460M as a scalar value, denoted herein as "avg_offdiag" ; and (c) calculate covariance loss value 460L based on avg_diag and avg_offdiag, e.g., according to Eq. 1, below:Eq. 1Covariance loss value 460L = (avg_diag - 1)^2 + (avg_offdiag - 0)^2.

[0144] Training module 40 may then feed covariance loss value 460L back to generative model 100, and train generative model 100 to produce reconstructed training image 30REC while minimizing the covariance loss value 460L.

[0145] It may be appreciated that by doing so, training module 40 may decorrelate, or “whiten” latent features 130LF such that covariance matrix will be diagonalized, representing negligible correlation between dimensions of the latent feature vectors 130LV.

[0146] Reference is now made to Fig. 4E, which is a block diagram depicting yet another example of operation of the training module 40, according to some embodiments of the invention. Elements of system 10 in Fig. 4E may be the same as corresponding (e.g., equally enumerated) elements of Fig. 2.

[0147] As elaborated herein, training module 40 may constrain a subset, or a group of latent features 130LF to be correlative to, or affect classification 50C. In other words, training module may differentiate between latent features 130LF that affect, or are pertinent to classification 50C (denoted herein as subset 1308), and latent features 130LF that are merely required for the purpose of image reconstruction (e.g., for maintain lighting conditions between images 20 and 30REC).

[0148] System 10 may subsequently perform the perturbation and / or disentanglement of latent features 130LF as elaborated herein (e.g., in relation to Figs. 2, 4B) on these selected and constrained latent features 130LF.

[0149] According to some embodiments, during a training stage, system 10 may apply or infer generative model 100 on a training image 20 depicting a specimen 20’ (e.g., to produce a corresponding reconstructed training image 30REC). System 10 may infer ML- based classifier model 50 on training image 20 to obtain a first baseline classification 50C of specimen 20’.

[0150] System 10 may select, or receive (e.g., via input 7 of Fig. 1) a subset 130S (also referred herein as a “training subset”) or number (e.g., 10) of latent features 130LF of latent feature vector 130LV. It is the latent features 130LF of this subset 1308 that may be constrained to affect classification 50C.

[0151] As shown in the example of Fig. 4E, training module 40 may include a secondary ML-based classifier model 470. Secondary ML-based classifier model 470 may be, or include, for example, a single neuron classifier, trained to receive subset 1308 of latent features 130LF from latent feature vector 130LV (originating from an input image 20 of specimen 20’), and classify specimen 20’ based on the input subset 1308.

[0152] For example, secondary ML-based classifier model 470 may receive classification 50C from classifier model 50 as supervisory data, to be trained to classify specimen 20’ based on the input latent features 130LF of subset 1308, and thereby force input latent features 130LF of subset 1308 to be correlated to classification 50C. Secondary ML-based classifier 470 may thereby produce a secondary baseline classification 470C of the sample image 20 (e.g., of the specimen 20’ depicted in sample image 20).

[0153] According to some embodiments, training module 40 may include a classification loss module 475, configured to calculate another loss term value, referred to herein as a secondary classification loss value 475L. Secondary classification loss value 475L may represent discrepancy between the baseline classification 50C and secondary baseline classification 470C.

[0154] During a training stage, training module 40 may feed secondary classification loss value 475L back to generative model 100, and may train generative model 100 to produce the reconstructed training image 30REC while minimizing secondary classification loss value 475L. It may be appreciated that by doing so, training module 40 may provide at least partial correlation between representation of depicted specimens 20’ in the latent space 130LS, and classifier prediction 50C, and constrain or guide subset 1308 of the latent features 130LF to predict the classification model’s 50C score, or classification of the inputimage 20. This constraint may define which latent features 130LF are to be perturbated for interpretability at inference.

[0155] In other words, the series of perturbations elaborated herein (e.g., in relation to Fig. 2, 4B) may be applied or performed only on one or more latent features 130LF of the latent feature subset 130S.

[0156] Reference is now made back to Fig. 2, which shows an assistive diagnostics module 70 (or module 70, for short), as part of system 10.

[0157] According to some embodiments, module 70 may be, or may include a Graphical User Interface (GUI), configured to collaborate with input 7 and / or output 8 of Fig. 1, to visualize refined classification 60C to a user. Module 70 may display baseline classification 50C via the GUI, and / or display reconstructed image 30REC via the GUI. In such manner, module 70 may highlight at least one visual trait 20T (or a difference visual trait 20T between image 20 and image 30REC) that is associated with an identified at least one latent feature 130UF of latent feature vector 130UV.

[0158] For example, as elaborated herein, system 10 may be configured to identify at least one latent feature 130UF as a “cause” of baseline classification 50C. In this example, module 70 may display baseline classification 50C via the GUI, while highlighting the visual trait 20T’ (e.g., BE) corresponding to the identified latent feature 130EF.

[0159] Additionally, or alternatively, module 70 may collaborate with perturbation module 410 to perturbate the identified latent feature 130EF (e.g., identified as a “cause” of baseline classification 50C), to produce one or more perturbed latent feature vectors 130LVP. Module 70 may subsequently apply a generative component (e.g., a decoder) 140 ofthe generative model 100 on the one or more perturbed latent feature vectors 130LVP. to obtain one or more (e.g., a set of) perturbed reconstructed images 30RECP.

[0160] Module 70 may proceed to infer the ME-based classifier model 50 on the one or more perturbed reconstructed images 30RECP, to obtain one or more respective perturbed classifications 50CP, and select at least one perturbed reconstructed image based on the one or more perturbed classifications which is different from the baseline classification. It may be appreciated that module 70 may thus identify a tipping point at which a value of latent feature 130LF caused classifications 50C to change.

[0161] Module 70 may subsequently display, via the GUI, the selected at least one perturbed reconstructed image 30RECP, while highlighting the at least one visual trait 20T, associated with the identified at least one latent feature.

[0162] In some embodiments, secondary ML-based classifier model 470 (described with reference to Fig. 4E) may also be used at an inference stage.

[0163] In some embodiments, system 10 may select, or receive (e.g., via input 7 of Fig. 1) atarget subset of latent features (e.g., latent features 130LF) of latent feature vector (e.g., latent feature vector 130LV) which may be obtained by applying ML-based generative model 100 on sample image 20. Latent features 130LF of the target subset may be constrained to affect classification 50C. It shall be understood that the “target” subset as referred to herein may include same features as the subset 1308 described above, but different values of these features if they are derived from another image (e.g., sample image 20).

[0164] E.g., in some embodiments, system 10 may be further configured to infer secondary ML-based classifier model 470 on the target subset to obtain a supplementary baseline classification (not shown) of the specimen. System 10 may be further configured to calculate, based on the supplementary baseline classification, for each latent feature 130LF of the target subset, a Shapely value representing a degree of contribution of a respective latent feature 130LF to the supplementary baseline classification.

[0165] In some embodiments, Shapley values may be calculated by applying SHapley Additive exPlanations (SHAP) known in the art. SHAP is a method used in the art to interpret the output of machine learning models by assigning each feature an importance value for a particular prediction. SHAP values are derived from cooperative game theory and provide a way to fairly distribute the “payout” (in this case, the prediction) among the features. By calculating the contribution of each feature to the prediction, SHAP helps in understanding the impact of each feature on the model’s output. This method is known to be particularly useful for explaining complex models and making their predictions more transparent and interpretable.

[0166] In some embodiments, system 10 may be further configured to visualize the respective refined classification (e.g., refined classification 60C) via the GUI, further by highlighting visual traits 20T (as explained above) and displaying, for highlighted visualtraits 20T, a visual representation of the calculated Shapely values of the respective latent features associated with said visual trait 20T.

[0167] In some embodiments, said visual representation of the calculated Shapely values may be or may include, e.g., respective numeric values shown in association with respective visual traits 20T, and / or it may affect the way visual traits 20T are highlighted (e.g., the higher the respective Shapely value, the higher the brightness / saturation / transparency of the highlighted region). It shall be appreciated that other similar ways of providing visual representations of the Shapely values may be covered by the scope of the present invention.

[0168] The above-described application of Shapely values enables the interpretation of the connection between the latent features and the semantic properties, thereby enabling instance classification interpretability on a case-by-case basis. The “weight” attributed to each latent feature (e.g., latent feature 130LF) that Shapely value represents, along with the mapping from latent features to interpreted semantic properties (e.g. visual traits 20T), enables to identify and rank the semantic properties most influencing the classification of a specific instance (e.g., specimen 20’). Thereby, further improvement of the relevant technological field may be achieved.

[0169] It shall be appreciated that, for the matter of clarity, some aspects of the training stages described with reference to Figs. 4A-4E, are explained using same elements (in particular, sample images 20, specimen 20’, latent feature vector 130LV, latent features 130LF etc.), as for the inference stage. Furthermore, in some sections of the present description, similar elements (e.g., elements of the same type) are associated with the same reference numbers (e.g., sample image 20 and training image 20). Accordingly, it shall be understood that, in practice, the training stage may be performed based on different elements (e.g., training image(s), training subset of the latent features etc.) than the elements used for the inference stage.

[0170] It shall also be appreciated that, for the matter of clarity and conciseness, the present description refers to the terms such as “latent features” and “subset of latent features” as to both the features (e.g., “size”, “shape” etc.) themselves and the values thereof (e.g., “big”, “round” etc.) . Accordingly, it would be clear for the person skilled in the art that latent feature vector 130LV, calculated for different inputs (e.g., sample image 20 at the inference stage and training image 20 at the training stage) may have same set of latent features 130LF, but different values thereof.

[0171] Pertaining to the example of Fig. 3B, module 70 may present refined classification 60C to a “non-viable” baseline classification 50C as ‘‘non-viable because of the ICM value 20T, of a specific latent feature 130LF, having surpassed a tipping point value” (e.g., a value in which classification 50C of perturbed image 30RECP has changed, e.g., from “viable” to “non-viable”).

[0172] As elaborated herein, system 10 may be applied to use cases that include analysis of biomedical images. For example, as elaborated in the examples of Figs. 3A-3C, specimen 20’ may be a blastocyst. In this example, module 70 may receive a plurality of sample images 20 corresponding to a plurality of blastocysts 20’. For each sample image, module 70 may produce a refined classification 60C of the corresponding blastocyst 20’, and may be used for screening the plurality of blastocysts 20’ (of the plurality of images 20) for IVF transplant, based on the refined classifications 60C.

[0173] In another example, system 10 may be applied to screening cells in a process of single cell flow cytometry. For example, system 10 may receive sample images 20 that may include biological cell specimens 20’, depicted by single cell flow cytometry, as known in the art. In this example, system 10 may receive a plurality of sample images 20, corresponding to a plurality of biological cells 20’. For each sample image 20, system 10 may produce a refined classification 60C of the corresponding biological cell 20’. Module 70 may then be used for screening the plurality of biological cells 20’ according to at least one criterion, based on the refined classifications 60C. Such a criterion may be, for example, a change in morphological characteristics of the cell 20’ (e.g., protrusions), light reflectance properties of cell 20’ that can indicate a change in cell stiffness or organelle composition, and the like.

[0174] In yet another example, system 10 may be applied to screening an Integrated Circuit (IC) or chip. For example, system 10 may receive a plurality of sample images 20 corresponding to a plurality of specimens 20’ that are chips on a production line. For each sample image 20, system 10 may produce a refined classification 60 of the corresponding chip 20’, module 70 may screen the plurality of chips 20’ according to at least one criterion, based on the refined classifications 60C.

[0175] Other exemplary implementations of system 10 classification of pathological cell images, such as granulomas, by which refined classifications 60C may provide an expertinsight regarding sub-variants of the underlying pathology, and optionally insight regarding a possible treatment for the same.

[0176] It shall be appreciated that, additionally, or alternatively, system 10 of any of Figs . 2, 4A-4E may include the same modules as system 10 of another figure of Figs. 2, 4A-4E. It may be appreciated that some modules of system 10 as discussed with reference to any of Figs. 2, 4A-4E may have been omitted for the sake of clarity.

[0177] Reference is now made to Fig. 5, which is a flow diagram depicting a method of classifying at least one specimen depicted in an image, by at least one processor (e.g., processor 2 of Fig. 1), according to some embodiments of the invention.

[0178] As shown in step S 1005, the at least one processor 2 may receive a sample image 20 of a specimen 20’, depicting a plurality of visual traits 20T, also referred to herein as visual phenotypes 20T.

[0179] As shown in step S1010, the at least one processor 2 may infer an ML based classifier model 50 on the sample image 20, to obtain a baseline classification 50C of the specimen 20’.

[0180] As shown in step S 1015, the at least one processor 2 may apply an ML-based generative model 100 on the sample image 20, to obtain a reconstructed image 30REC of the specimen, via a latent feature vector 130LV. Latent feature vector 130LV may include a plurality of latent features 130LF, defining a latent feature space 130LS.

[0181] As shown in step S1020, and as elaborated herein (e.g., in relation to Fig. 4B) the at least one processor 2 may employ a disentanglement model 430 to disentangle the latent feature vector 130LV. As elaborated herein, such disentanglement may include association of each latent feature 130LF with zero, one or more unique visual traits 20T or visual phenotypes 20T of the plurality of visual traits. Examples for visual traits 20T or visual phenotypes 20T are provided herein, e.g., in relation to Figs. 3A-3C.

[0182] As shown in step SI 025, the at least one processor 2 may analyze reconstructed image 30REC, to identify at least one latent feature 130LF as correlative, or pertinent to baseline classification 50C of specimen 20’.

[0183] As shown in step SI 030, the at least one processor 2 may produce a refined classification 60C of specimen 20’. Refined classification 60C may include (i) the baseline classification 50C and (ii) at least one visual trait or visual phenotype 20T, associated with the identified at least one latent feature 130LF.

[0184] Reference is now made to Fig. 6, which is a flow diagram depicting a method of screening biological specimens 20’ by at least one processor 2, according to some embodiments of the invention.

[0185] As shown in step S2005, the at least one processor 2 may receive a sample image 20 depicting a biological specimen 20’.

[0186] As shown in step S2010, embodiments of the invention may provide a generative model, such as an autoencoder model 100. As elaborated herein, the at least one processor 2 may train autoencoder model 100 to generate a reconstructed version 30REC of the sample image 20 via a latent feature vector 130LV, such that reconstructed version 30REC may depict one or more visual phenotypes 20T of the biological specimen 20’.

[0187] As shown in step S2015, the at least one processor 2 may provide a predictor ML model 50 configured to constrain the latent feature vector 130LV, such that at least one latent feature 130LF of latent feature vector 130LV is linked by injective association to a respective visual phenotype, or visual trait 20T of the one or more visual phenotypes 20T, as depicted in the reconstructed version 30REC.

[0188] As elaborated herein, and as shown in step S2020, the at least one processor 2 may proceed to screen the biological specimen 20’ based on the at least one latent feature (e.g., based on refined classification 60C).

[0189] Experimental Data

[0190] The system and method of the present invention have been tested, demonstrating high efficiency and reliability, thereby contributing to improvements of the relevant technological field.

[0191] The experimental data used during tests were derived from a comprehensive dataset of in vitro fertilization (IVF) procedures. Specifically, the dataset comprised 11,211 embryo time-lapse videos collected from IVF cycles conducted at three clinic centers between March 2010 and December 2021. These videos were provided by AIVF LTD and included historical images of blastocyst-stage embryos along with associated metadata. The procedures and protocols for data collection were approved by an Institutional Review Board for secondary research use (IRB reference number HMO-006-20).

[0192] Fertilization was determined by the presence of two pronuclei (2PN) 16-18 hours after insemination. All zygotes were placed inside the EmbryoScope™ time-lapse incubator system (Vitrolife, Denmark) and incubated using a sequential media protocol until theblastocyst stage. The embryos were live-imaged with atemporal resolution of 15-20 minutes per frame. Each grayscale image (8-bit) was of size 500x500 pixels, with a physical pixel size of 294x294 μm2. Image stacks consisting of 7 slices, 15 μm apart, were acquired at each time point, with the middle slice used for analysis.

[0193] Analysis was performed on embryos at the blastocyst stage, with the typical onset of blastulation occurring approximately 103 hours post-insemination based on manual annotation of blastulation and hatching. A total of 6-10 frames from embryos at the blastocyst stage were collected with equal time intervals between them. High-saturation images and images with partially visible blastocysts were excluded, resulting in approximately 67,000 images (e.g., such as images 20, discussed with reference to Figs. 2, 4A-4E) used to train the model (e.g., generative model 100 and predictor ML model 50, discussed with reference to Figs. 2, 4A-4E).

[0194] Blastocysts were manually annotated by embryologists just before hatching or before the removal of the embryo from the microscope, according to the Gardner and Schoolcraft scoring criteria, one of the most common morphology-based blastocyst assessment criteria. The Gardner criteria are based on three morphology-based quality parameters: blastocyst expansion status (graded 1-6), Inner Cell Mass (ICM) morphology (graded A-C), and Trophectoderm (TE) morphology (graded A-C). In this dataset, blastocyst expansion status was not annotated. High-quality blastocysts were defined by corresponding ICM and TE labels of AA, AB, or BA, while low-quality blastocysts were defined by BB, BC, or CB.

[0195] The image pixel intensities were normalized to the range [0, 1] . To accommodate IVF-CLF training on a single GPU, the blastocyst images were preprocessed to reduce size, and their background was masked to eliminate irrelevant information. The preprocessing steps included semantic segmentation of the blastocyst from the raw image, centering the blastocyst in the image, and resizing the image to a lower resolution. Specifically, a Mask- RCNN object detection model was trained to detect 200x200 pixel bounding boxes around each blastocyst, using 800 raw images with manually annotated blastocyst bounding boxes. A Hough-transform detected the blastocyst's circular shape within the Mask-RCNN bounding box and was used to mask the non-blastocyst image regions and to center the blastocyst in the image. A U-NET was then trained to segment the blastocyst using 500 outof the 800 images successfully segmented by the Hough transform. The segmented image was resized to 64x64 pixels using nearest neighbors interpolation.

[0196] For the classification of high- versus low-quality blastocysts (as performed by predictor ML model 50, discussed with reference to Figs. 2, 4A-4E, also referred herein as “IVF-CLF”), an hnageNet pretrained VGG-19 network was selected. The model was finetuned by re-training it to discriminate between high- and low-quality blastocysts using a balanced training dataset of 977 high-quality and 977 low-quality blastocysts. The test dataset comprised 108 high-quality and 108 low-quality blastocysts. The IVF-CLF architecture included the VGG-19 feature extraction part, with the last fully connected layer replaced by a fully connected 16-node dense layer and an output node dense layer with a sigmoid activation, corresponding to a probability of high-quality blastocyst (0-1). The model was compiled with binary cross-entropy loss and Adam optimizer with a learning rate of 0.002. The IVF -CLF network was trained for 100 epochs with a batch size of 32, and data augmentation was performed by altering brightness, flipping, rotating, and adding Gaussian noise.

[0197] Overall, the experimental data provided a robust foundation for training and validating the system 10 (e.g., generative model 100, as discussed with reference to Figs. 2, 4A-4E), enabling the identification and interpretation of classification-driving morphological properties in blastocyst images.

[0198] A system (such as system 10, discussed with reference to Figs. 2, 4A-4E), designed for visual interpretability of image-based classification models, was structured as follows.

[0199] The system (also referred herein as “DISCOVER” model) was based on a deep learning generative framework that encodes image data into a disentangled latent representation. This framework comprises an adversarial autoencoder (AAE) with an encoder and a decoder (e.g., generative model 100, as discussed with reference to Figs. 2, 4A-4E). The encoder (such as encoder 120) is responsible for encoding incident sample data (e.g., images 20) into a latent feature vector representation (e.g., vector 130LV) in a reduced- dimension latent space (e.g., space 130LS). The decoder (such as decoder 140) is responsible for decoding the latent feature vectors to produce reconstructed versions of the incident data.

[0200] The encoder and decoder network backbones were based on residual blocks similar to those introduced in ResNet50. The outputs of the last convolutional downsamplingblock were flatened to a vector followed by a dense layer of 350 dimensions that defined the latent representation. The discriminator network was composed of six fully connected dense layers.

[0201] The reconstruction loss (such as reconstruction loss 405L) was a perceptual loss where the reconstruction minimized the Euclidean distance between the hidden layers of a VGG-19 pre-trained on hnageNet. This perceptual loss was preferred over minimizing LI or L2 pixel-wise differences because the later led to blurry and less realistic reconstructed images. The latent generative-adversarial loss enforced a probabilistic latent space such that samples were encoded into a continuous dense distribution.

[0202] To ensure that the visual semantic properties influencing the classification decision were maintained in the reconstructed image, a loss term minimized the discrepancy between the hidden representations of the of the real versus its corresponding reconstructed image. This perceptual loss was calculated by extracting the hidden representations of the IVF-CLF network (e.g., predictor ML model 50, discussed with reference to Figs. 2, 4A- 4E) from specific layers and minimizing the mean absolute error (MAE) between these layers.

[0203] The disentanglement module was designed to encode the image into a decorrelated latent space, where each latent feature is independent of the others. This was achieved by whitening the latent features’ covariance matrix (e.g., matrix 460M, as discussed with reference to Fig. 4D) and optimizing a new network to identify which latent feature was altered upon alteration of a single latent feature. The counterfactual disentanglement network was implemented as a single neuron trained to predict the IVF- CLF’ s classification score from the classification-driving subset of latent features (e.g., subset 1308, as discussed with reference to Fig. 4E).

[0204] The system (e.g., the generative model 100, as discussed with reference to Figs. 2, 4A-4E) was trained with the Adam optimizer, a learning rate of 0.0002, and a batch size of 64 for 30 epochs. Data augmentation was performed by altering brightness, flipping, rotating, and adding Gaussian noise. The overall loss of the system was defined as the addition of all six loss terms, with weights adjusted empirically to ensure convergence.

[0205] This structure enabled system 10 to provide high-quality and realistic reconstruction and traversal of the latent space, domain-specific classification-orientedencoding, and visual disentanglement of the latent representation, making it a powerful tool for interpreting and discovering classification-driving semantic image properties.

[0206] As known, classic explainable Al methods known in the art provided convoluted visual explanations that were unintuitive to embryologists. In contrast, the system according to the present invention provided clear and interpretable visual explanations by generating counterfactual images where classification-driving image properties were exaggerated. This allowed for the identification of subtle phenotypes and the establishment of new mechanistic hypotheses.

[0207] Specifically, the system identified two top classification-driving latent features: blastocyst size and trophectoderm quality. These features were visually disentangled and interpreted by domain experts. The system also discovered a latent feature encoding the blastocoel density, a classification-driving morphological property that was not explicitly annotated. This highlighted the potential of the system to define quantitative measures for morphologic properties that do not have explicit measurements and even identify novel visual classification-driving properties that were not known a priori.

[0208] Furthermore, the system enabled the determination of the cause of classification for specific blastocysts. By applying Shapley Additive exPlanations (SHAP) to the classification-driving subset of latent features, the system could weigh and rank the semantic properties most influencing the classification of a specific instance. This provided transparency and trust in the classification decisions, which is crucial in the biomedical domain.

[0209] Overall, the results established that the system has demonstrated a high efficiency of providing human-interpretable understanding of “black-box” classification models, proposing hypotheses to decipher underlying biomedical mechanisms, and providing transparency for the classification of individual predictions. This makes the present invention a powerful tool for computational-driven biological and clinical discovery in the domain of IVF and potentially other technical fields.

[0210] Unless explicitly stated, the method embodiments described herein are not constrained to a particular order or sequence. Furthermore, all formulas described herein are intended as examples only and other or different formulas may be used. Additionally, some of the described method embodiments or elements thereof may occur or be performed at the same point in time.

[0211] While certain features of the invention have been illustrated and described herein, many modifications, substitutions, changes, and equivalents may occur to those skilled in the art. It is, therefore, to be understood that the appended claims are intended to cover all such modifications and changes as fall within the true spirit of the invention.

[0212] Various embodiments have been presented. Each of these embodiments may of course include features from other embodiments presented, and embodiments not specifically described may include various features described herein.

Claims

CLAIMS1. A method of classifying a specimen depicted in an image by at least one processor, the method comprising: receiving a sample image of a specimen, depicting a plurality of visual traits; inferring a machine-learning (ML)-based classifier model on the sample image, to obtain a baseline classification of the specimen; applying an ML-based generative model on the sample image, to obtain a reconstructed image of the specimen, via a latent feature vector, comprising a plurality of latent features, said latent feature vector being disentangled to have each latent feature associated with zero, one or more visual traits of the plurality of visual traits; analyzing the reconstructed image, to identify at least one latent feature as correlative to the baseline classification of the specimen; and producing a refined classification of the specimen, comprising (i) the baseline classification and (ii) at least one visual trait, associated with the identified at least one latent feature.

2. The method of claim 1, further comprising disentangling the latent feature vector, to associate each latent feature with said zero, one or more unique visual traits of the plurality of said visual traits.

3. The method according to any one of claims 1-2, further comprising visualizing the refined classification via a Graphical User Interface (GUI), by: displaying the baseline classification via the GUI; and displaying the reconstructed image via the GUI, while highlighting the at least one visual trait, associated with the identified at least one latent feature.

4. The method of claim 3, wherein visualizing the refined classification further comprises: perturbating the identified latent feature, to produce one or more perturbed latent feature vectors;applying a generative component of the generative model on the one or more perturbed latent feature vectors, to obtain one or more perturbed reconstructed images; inferring the ML-based classifier model on the one or more perturbed reconstructed images, to obtain one or more respective perturbed classifications; selecting at least one perturbed reconstructed image based on the one or more perturbed classifications which is different from the baseline classification; and displaying, via the GUI, the selected at least one perturbed reconstructed image, while highlighting the at least one visual trait, associated with the identified at least one latent feature.

5. The method according to any one of claims 3-4, further comprising: selecting a target subset of the latent features of the latent feature vector obtained by applying the ML-based generative model on the sample image; inferring a secondary ML-based classifier model on the target subset to obtain a supplementary baseline classification of the specimen; based on the supplementary baseline classification, calculating, for each latent feature of the target subset, a Shapely value representing a degree of contribution of a respective latent feature to the supplementary baseline classification; and visualizing the refined classification via the GUI further by displaying, for said highlighted at least one visual trait, a visual representation of the calculated Shapely value of the respective latent feature associated with said at least one visual trait.

6. The method according to any one of claims 1-5, wherein the specimen is a blastocyst, and wherein the method further comprises: receiving a plurality of sample images, corresponding to a plurality of blastocysts; for each sample image, producing a refined classification of the corresponding blastocyst; and screening the plurality of blastocysts for In Vitro Fertilization (IVF), based on the refined classifications.

7. The method according to any one of claims 1-6, wherein receiving the sample image comprises performing Single cell flow cytometry, and wherein the specimen is a biological cell, and wherein the method further comprises: receiving a plurality of sample images, corresponding to a plurality of biological cells; for each sample image, producing a refined classification of the corresponding biological cell; and screening the plurality of biological cells according to at least one criterion, based on the refined classifications.

8. The method according to any one of claims 1 -7, wherein the specimen is an Integrated Circuit (IC) chip, and wherein the method further comprises: receiving a plurality of sample images, corresponding to a plurality of chips in a production line; for each sample image, producing a refined classification of the corresponding chip; and screening the plurality of chips according to at least one criterion, based on the refined classifications.

9. The method according to any one of claims 1-8, further comprising: applying the generative model on a training image depicting a specimen, to produce a corresponding reconstructed training image; calculating a reconstruction loss value, representing a discrepancy between the training image and reconstructed training image; and training the generative model to produce the reconstructed training image while minimizing the reconstruction loss value.

10. The method according to any one of claims 1-9, further comprising: applying the generative model on a training image depicting a specimen, to produce a corresponding reconstructed training image; inferring the ML-based classifier model on the training image to obtain a first baseline classification of the specimen;inferring the ML-based classifier model on the reconstructed training image to obtain a second baseline classification of the specimen; calculating a classification loss value, representing a discrepancy between the first baseline classification and second baseline classification; and training the generative model to produce the reconstructed training image while minimizing the classification loss value.

11. The method according to any one of claims 1-10, wherein disentangling the latent feature vector comprises: applying the generative model on a first image of a specimen, to produce a first latent feature vector; selecting an index value representing a specific latent feature in the latent feature vector; perturbating the latent feature of the selected index value, to produce a perturbed latent feature vector; obtaining, from an ML-based disentanglement model, a predicted index value, representing prediction of the selected index, based on the perturbed latent feature vector; calculating a disentanglement loss value, representing a discrepancy between the selected index value and the predicted index value; and training the generative model to produce the reconstructed image of the specimen while minimizing the disentanglement loss value, thereby uniquely associating the latent feature of the selected index with one or more respective visual traits of the plurality of visual traits.

12. The method of claim 11, wherein obtaining the predicted value of the selected index based on the perturbed latent feature vector comprises: applying a generative component of the generative model on the first latent feature vector, to obtain a first reconstructed image; applying the generative component of the generative model on the perturbed latent feature vector, to obtain a second reconstructed image; calculating a difference image, representing difference between the first reconstructed image and the second reconstructed image; andapplying the ML-based disentanglement model on the difference image, to predict the value of the selected index.

13. The method according to any one of claims 11-12, further comprising: applying the generative model on a training image, to produce a respective nonperturbed, reconstructed training image; selecting an index of a latent feature of the latent feature vector; perturbating the latent feature of the selected index, to produce a perturbed latent feature vector; applying a generative component of the generative model on the perturbed latent feature vector, to obtain a perturbed, reconstructed training image; calculating a difference image, representing difference between the perturbed reconstructed training image and the non-perturbed, reconstructed training image; and using a value of the selected index as supervisory data, to train the ML-based disentanglement model to predict the index of the perturbated latent feature.

14. The method according to any one of claims 1-13, wherein analyzing the reconstructed image comprises: for at least one latent feature of the latent feature vector, applying a series of perturbations, to obtain a respective series of perturbed reconstructed images; inferring the ML based classifier model on one or more perturbed reconstructed images of the series of perturbed reconstructed images, to obtain one or more respective, perturbed classifications of the specimen; comparing the baseline classification of the specimen to the one or more perturbed classifications; and identifying the at least one latent feature as correlative to the baseline classification of the specimen, based on said comparison.

15. The method of claim 14, further comprising: applying the generative model on a training image depicting a specimen, to produce a corresponding reconstructed training image;inferring the ML-based classifier model on the training image to obtain a first baseline classification of the specimen; selecting a training subset of the latent features of the latent feature vector obtained by applying the generative model on the training image; inferring a secondary ML-based classifier model on the training subset, to obtain a second baseline classification of the training image; calculating a secondary classification loss value, representing a discrepancy between the first baseline classification and second baseline classification; and training the generative model to produce the reconstructed training image while minimizing the secondary classification loss value.

16. The method of claim 15, further comprising applying the series of perturbations only on one or more latent features of the training subset.

17. The method according to any one of claims 1-16, further comprising: applying the generative model on a plurality of training images, to produce a corresponding plurality of latent feature vectors; based on the plurality of latent feature vectors, calculating a covariance matrix, representing covariance among the latent features; calculating a covariance loss value, representing a discrepancy between the covariance matrix and a unity matrix of the same dimension; and training the generative model to produce the reconstructed training image while minimizing the covariance loss value.

18. A system for classifying a specimen depicted in an image, the system comprising: a non-transitory memory device, wherein modules of instruction code are stored, and at least one processor associated with the memory device, and configured to execute the modules of instruction code, whereupon execution of said modules of instruction code, the at least one processor is configured to: receive a sample image of a specimen, depicting a plurality of visual traits; infer a machine-learning (ML)-based classifier model on the sample image, to obtain a baseline classification of the specimen;apply an ML-based generative model on the sample image, to obtain a reconstructed image of the specimen, via a latent feature vector, comprising a plurality of latent features; disentangle the latent feature vector, to associate each latent feature with zero, one or more unique visual traits of the plurality of visual traits; analyze the reconstructed image, to identify at least one latent feature as correlative to the baseline classification of the specimen; and produce a refined classification of the specimen, comprising (i) the baseline classification and (ii) at least one visual trait, associated with the identified at least one latent feature.

19. The system of claim 18, wherein the at least one processor is configured to disentangle the latent feature vector by: applying the generative model on a first image of a specimen, to produce a first latent feature vector; selecting an index value representing a specific latent feature in the latent feature vector; perturbating the latent feature of the selected index value, to produce a perturbed latent feature vector; obtaining, from an ML-based disentanglement model, a predicted index value, representing prediction of the selected index, based on the perturbed latent feature vector; calculating a disentanglement loss value, representing a discrepancy between the selected index value and the predicted index value; and training the generative model to produce the reconstructed image of the specimen while minimizing the disentanglement loss value, thereby uniquely associating the latent feature of the selected index with one or more respective visual traits of the plurality of visual traits.

20. The system of claim 19, wherein the at least one processor is configured to obtain the predicted value of the selected index based on the perturbed latent feature vector by: applying a generative component of the generative model on the first latent feature vector, to obtain a first reconstructed image;applying the generative component of the generative model on the perturbed latent feature vector, to obtain a second reconstructed image; calculating a difference image, representing difference between the first reconstructed image and the second reconstructed image; and applying the ML-based disentanglement model on the difference image, to predict the value of the selected index.

21. The system according to any one of claims 18-20, wherein the at least one processor is configured to analyze the reconstructed image by: for at least one latent feature of the latent feature vector, applying a series of perturbations, to obtain a respective series of perturbed reconstructed images; inferring the ML based classifier model on one or more perturbed reconstructed images of the series of perturbed reconstructed images, to obtain one or more respective, perturbed classifications of the specimen; comparing the baseline classification of the specimen to the one or more perturbed classifications; and identifying the at least one latent feature as correlative to the baseline classification of the specimen, based on said comparison.

22. A method of screening biological specimens by at least one processor, the method comprising: receiving a sample image depicting a biological specimen; providing an autoencoder model, trained to generate a reconstructed version of the sample image via a latent feature vector, wherein the reconstructed version depicts one or more visual phenotypes of the biological specimen; providing a predictor ML model configured to constrain the latent feature vector, such that at least one latent feature of latent feature vector is linked by injective association to a respective visual phenotype of the one or more visual phenotypes, depicted in the reconstructed version; and screening the biological specimen based on the at least one latent feature.

Images