A computer-implemented method for cell-level detection, prediction and visualization of cancer biomarker associated features

Abstract

The present invention provides a computer-implemented training method for cell-level detection and prediction of cancer biomarker associated features. The method comprises obtaining, for each of a plurality of cancer biopsy samples, a multiplex immunohistochemistry (mIHC) image and corresponding hematoxylin and eosin (H&E) stained tissue image, aligning the corresponding mIHC and H&E stained tissue images, obtaining a labelled cell tile dataset by labelling the H&E images based on the cell labels from the corresponding mIHC images, training a deep learning classifier model, using the labelled cell tile dataset, for classifying specified cell specificity states and training a conditional generative adversarial network model to map cell image features onto a latent space and generate synthetic H&E images, corresponding to specified cell specificity states, from the latent space.

Description

[0001]

[0002] The present invention relates, in general terms, to a computer-implemented method for cell-level detection, prediction and visualization of cancer biomarker- associated features and the system thereof.

[0003] Background

[0004] This background is provided for generally presenting the context of the disclosure. Contents of this background section are neither expressly nor implied admitted as prior art against the present disclosure.

[0005] Deep learning models, while transformative in various applications, face critical challenges in medical settings due to their intrinsic 'black-box' nature. This lack of transparency in how these models make decisions poses a significant hurdle, especially in medicine, where precise and explainable decision-making is paramount. Additionally, these models are prone to 'shortcut learning'. This phenomenon leads them to rely on superficial or spurious features for decisionmaking, rather than on robust, generalizable patterns, thus compromising their reliability in critical applications such as medical diagnostics.

[0006] The challenges are more pronounced in cell-level prediction tasks. Here, deep learning models are typically trained on cellular images that capture only a fraction of the myriad variations resulting from different operator techniques, machine calibrations, and preprocessing methods. Such a limited training scope can lead to models that are not robust or reliable enough for the diverse and complex scenarios encountered in real-world medical environments. Moreover, the complexity of associating cellular protein expressions with histological features (such as those seen in H&E stained images) often surpasses human expert analysis capabilities. As a result, not only is there uncertainty regarding the reliability of these deep learning models at the cellular level, but there is also a significant challenge in extracting clinically actionable or useful histomorphological features from them. Overcoming these limitations is crucial for advancing the application of artificial intelligence (Al) in medical diagnostics and ensuring its adoption in clinical practices.

[0007] It would be desirable to overcome or alleviate at least one of the abovedescribed problems, or at least to provide a useful alternative.

[0008] Summary

[0009] The present invention provides a computer-implemented training method for cell-level detection and prediction of cancer biomarker associated features. The method comprises obtaining, for each of a plurality of cancer biopsy samples, a multiplex immunohistochemistry (mIHC) image and corresponding hematoxylin and eosin (H&E) stained tissue image. The mIHC and corresponding H&E stained tissue images are aligned. A labelled cell tile dataset can then be obtained by labelling the H&E images based on the cell labels from the corresponding mIHC images. This enables training of a deep learning classifier model, using the labelled cell tile dataset, for classifying specified cell specificity states and training a conditional generative adversarial network (conditional GAN) model to map cell image features onto a latent space. The generator of the conditional GAN can be used to generate synthetic H&E images, corresponding to specified cell specificity states, from the latent space.

[0010] Additionally, the method further comprises performing cell-level detection and prediction of cancer biomarkers by process a stained image from a biopsy tissue slice using the deep learning classifier model to categorizing the stained image into one or more said cell specificity states and generating confidence probability of prediction of specificity state.

[0011] Preferably, the method comprises processing the stained image using the conditional-style generative adversarial network model to generate a synthetic H&E image, corresponding to the specified cell specificity states.

[0012] Preferably, the method comprises segmenting individual cells within the H&E images into image patches, wherein obtaining the labelled cell tile dataset comprises labelling the image patches based on the cell labels from corresponding mIHC image tiles.

[0013] More preferably, the deep learning classifier model is tested, using the labelled synthetic H&E images.

[0014] Preferably, alignment involves digital alignment using the Scale-Invariant Feature Transform (SIFT) and B-spline transformation techniques.

[0015] Preferably, the image patches are 64 x 64, 96 x 96, 100 x 100, or 112 x 112 pixel image tiles.

[0016] Further, the cell specificity states are T-cell tumor specificity states.

[0017] The present invention further provides a computer system for configurable for cell-level detection and prediction of cancer biomarkers associated features. The comprises memory, at least one processor, an image alignment module, a labeller module, a classifier module and a generative adversarial network module. The memory storing instructions that, when executed by the at least one processor, cause the system to obtain, for each of a plurality of cancer biopsy samples, a multiplex immunohistochemistry (mIHC) image and corresponding hematoxylin and eosin (H&E) stained tissue image, align, at the alignment module, the corresponding mIHC and H&E stained tissue images, use the labeller module to obtain a labelled cell tile dataset by labelling the H&E images based on the cell labels from the corresponding mIHC images, train a deep learning classifier model of the classifier module, using the labelled cell tile dataset, for classifying cell specificity states and train a conditional-style generative adversarial network model of the conditional generative adversarial network module, to map cell image features onto a latent space and generate synthetic H&E images, corresponding to specified cell specificity states, from the latent space.

[0018] Brief description of the drawings

[0019] Embodiments of the present invention will now be described, by way of nonlimiting example, with reference to the drawings in which :

[0020] Figure 1 illustrates an overview of an integrated system that combines a deep learning classifier with a conditional-style StyleGAN model.

[0021] Figure 2 illustrates a flow diagram of a method according to an embodiment of the invention.

[0022] Figure 3 illustrates StyleGAN3 model architecture and training method according to an embodiment of the invention.

[0023] Figure 4 illustrates a workflow of image generation for a continuum of specified cell specificity states according to an embodiment of the invention.

[0024] Figure 5 illustrates H&E Images generated by StyleGAN according to an embodiment of the invention.

[0025] Figure 6 illustrates PCA visualization of the latent vectors W, conditioned on the three defined cell classes and randomly retrieved from the trained StyleGAN according to an embodiment of the invention.

[0026] Figure 7 illustrates PCA visualization of the latent vectors W, conditioned on two disease stages, randomly retrieved from the trained StyleGAN according to an embodiment of the invention.

[0027] Figure 8 is a visualization of the dot product between latent vectors W and the normal vector wjtoiv of the hyperplane that separates the latent vectors into the corresponding disease stage according to an embodiment of the invention.

[0028] Figure 9 illustrates series of images showing the transition of a CR.C tumour cell as the disease advances from Stage I to Stage IV (left to right) according to an embodiment of the invention.

[0029] Detailed description

[0030] The present invention provides methods for cell-level detection and prediction of cancer biomarker-associated features and the system thereof.

[0031] Hereinafter, the system and method according to the present disclosure will be described in detail with reference to Figure 1 to Figure 9 according to the preferred embodiments. It is to be understood that limiting the description to the preferred embodiments of the invention is merely to facilitate discussion of the present invention. Various modifications may be made, to the preferred embodiments, without departing from the scope of the appended claims.

[0032] An embodiment of the present invention involves use of biopsy samples obtained from five patients with colorectal cancer. In practice, biopsy samples are collected, whether directly from patients, from frozen stores or other methods and are first placed on tissue slides and stained for CD8 and CD39 markers using the multiplex immunohistochemistry (mIHC) technique. This interpreted data was used in identifying three cell types, namely tumor-specific (CD39+) and non-tumor-specific (CD39“) CD8+T-cells, and non-CD8 cells

[0033] (CD39“CD8“). After mIHC staining and imaging, the same tissue slides were then washed and re-stained with Hematoxylin and Eosin (H&E) and imaged, to produce H&E images. Imaging can involve capturing images of H&E stained tissue using a camera, such as a colour camera, or other imaging device that can distinguish between the limited range of colours of tissue resulting from the staining process.

[0034] The present invention is designed to enhance histomorphological analysis of tumorous tissue, such as CD8+ T-cell tumor, specificity and cancer stage by generating and refining relevant image data. While mIHC images are generally associated with labels such as specificity (muscle cells, nerve cells, immune cells and others), cancer stage labels may need to be generated using an alternative labelling model or manual labelling.

[0035] Figure 1 outlines the three components of the present invention: preparation of training data 102, training of a cell classifier 104, and training of the StyleGAN model 106. In preparation of training data 102, the H&E-stained images were digitally registered to their corresponding mIHC-stained images by employing Scale-Invariant Feature Transform (SIFT) for feature detection, followed by B- spline transformation for precise spatial alignment, as illustrated in Figure 1. Subsequently, individual cells within the H&E images were segmented using the StarDist H&E segmentation algorithm. Each segmented cell was then extracted as an image tile of fixed dimensions, e.g. 64 x 64, 96 x 96, 100 x 100, or 112 x 112 pixels, with the cell positioned centrally within the tile. Corresponding image tiles from the mIHC-stained pair were used to assign cell labels to each H&E cell tile, as depicted in Figure 1.

[0036] These image tiles serve a dual purpose: first, to train the cell classifier 104, a DenseNet-121 deep learning model for classifying T-cell tumor specificity states, and second, to train a conditional StyleGAN3 model 106 to map cell image features onto a latent space, represented by the latent vector HZ, shown in Figures 3 and 4. The tiles were divided into specific training and held-out subsets

[0037] (61876 and 6875 tiles, respectively).

[0038] For the classification of three predefined cell categories— tumor-specific CD8+ T-cells, non-tumor-specific CD8+ T-cells, and non-CD8 cells— a two-stage binary classification system based on the DenseNet-121 deep learning architecture was developed. Both classification models were trained using a cross-entropy loss function, with a batch size of 50, and optimized via the Adam optimizer configured with default epsilon parameters as implemented in TensorFlow version 2.8. The initial learning rate was set to 0.001. Classification decision thresholds were established by optimizing the balance between sensitivity and specificity as determined from the Receiver Operating Characteristic (ROC) curve. In some embodiments, the training dataset is augmented using artificial cell images (e.g. synthetic labelled H&E stained images) generated via a conditional generative adversarial network, such as StyleGAN3. The artificial images are generated according to specified cell specificity states or cancer stages.

[0039] In a preferred embodiment, a conditional StyleGAN3 model is trained separately, consisting of a generator (including a mapping network) and a discriminator, conditioned on the three defined cell classes, namely tumorspecific and non-tumor-specific CD8+T-cells, and non-CD8 cells. The model was trained using a single GPU with mixed precision and default parameters: gamma 0.5, batch size 32, capacity multiplier 32768, and a maximum of 512 feature maps. For parameter optimization, we used the Adam optimizer with p_l=0, P_2=0.99, and e = 10 (-8). The learning rates were set to 2e-3 for the discriminator and 2.5e-3 for the generator network. Hyperparameters and architectural parameters were preset as defined in the 'stylegan3-t' configuration.

[0040] Figure 2 illustrates a flow diagram of a method according to an embodiment of the invention. The computer-implemented training method 200 for cell-level detection and prediction of cancer biomarker associated features comprises steps of obtaining images of mIHC and H&E stained tissue slides from biopsy cancer samples (step 202). As described above, biopsy samples are collected from five colorectal cancer patients and are first placed on tissue slides. These samples were then stained for CD8 and CD39 markers using mIHC. This staining helped identify three cell types: tumor-specific CD8+ T-cells (CD39+), non- tumor-specific CD8+ T-cells (CD39-), and non-CD8 cells (CD39- CD8-). After mIHC staining and imaging, the same tissue slides were then washed and restained with H&E. Furthermore, the corresponding mIHC and H&E stained images are digitally registered to achieve precise spatial alignment (step 204). Following alignment, individual cells within the H&E images are segmented (step 206), and each segmented cell is extracted as a discrete image patch or image tile of fixed dimensions, e.g. 100 by 100 pixels. Subsequently, labelled cell tile dataset is obtained by labelling the image tile based on the cell labels from its corresponding mIHC image tile (step 208). As shown in step 210, training on a deep learning classifier model is performed for classifying cell specificity states (namely tumor-specific CD8+ T-cells, non-tumor-specific CD8+ T-cells, and non-CD8 cells). Step 212 discloses performing training on a conditional generative adversarial network model or StyleGAN3 as a mapping model for mapping cell image features onto a latent space.

[0041] Step 214 discloses subjecting a stained image from a biopsy tissue slice through both classifier and mapping models in order to categorize the stained image into specificity states (step 216). Finally, confidence probability of prediction of specificity state is generated, as per step 218.

[0042] The method 200 further comprises processing a stained image from a biopsy tissue slice using the deep learning classifier model to categorize the stained image into one or more said cell specificity states, using the methods discussed above. Further, the method may include generating confidence probability of prediction of specificity state.

[0043] The method 200 further comprises processing the stained image using the conditional-style generative adversarial network model to generate a synthetic H&E image, corresponding to the specified cell specificity states.

[0044] Figure 3 describes StyleGAN3 model architecture and training methods. The method involves the following steps:

[0045] 1. Initialize mapping network and synthesis network.

[0046] 2. Forward pass with conditional class labels a. Input z: latent code (random noise vector) b. Input c: label vector (via embedding network) c. The mapping network generates the conditional latent vector w d. The synthesis network generates images conditioned on both w and c

[0047] 3. Discriminator and loss function a. The discriminator is trained to distinguish between real cell images (with correct labels) and generated cell images b. Conditional loss functions (e.g., cross-entropy for labels) are used to ensure label consistency

[0048] The conditional StyleGAN is used to generate artificial images corresponding to specified cell specificity states or cancer stages. Figure 4 illustrates a workflow of image generation for a continuum of specified cell specificity states according to an embodiment of the invention. In Figure 4, the initial latent code, Z, randomly sampled from a normal distribution, is transformed into a structured latent vector, 1 / 1 / , via a fully connected mapping network. In this conditional StyleGAN, the latent code is accompanied by cell class labels. Upon training, the model generates a synthetic image from this latent code. This image, a novel creation, corresponds to the point in the latent space represented by the initial Z code. It is transformed through the mapping network into the W code and realized as an image by the synthesis network, with the class label information guiding the generation process towards specific types of outputs.

[0049] Figure 4 describes methods for generation of inter-class images by categorizing the stained image into specificity states and generating confidence probability of prediction of specificity state

[0050] 1. Load a trained StyleGAN generator model

[0051] 2. Define starting and ending class labels: tumor non-specific CD8+T-cell (ci) and tumor-specific CD8+T-cell (C2)

[0052] 3. Sample latent codes: zi and z2for ci and cz; pass through the mapping network to obtain the intermediate latent vectors wi and w2

[0053] 4. Generate inter-class images: a. Define the number of interpolation steps = 100 b. For each interpolation step, compute interpolated w and c:

[0054] To visualize and understand the histomorphological transitions from non-tumor- specific to tumor-specific T-cell states, W vectors conditioned on these cell states are randomly sampled. The latent space trajectory between two randomly paired W vectors - one representing non-tumor-specific and the other tumorspecific T-cell class - was linearly interpolated in 0.01 increments. These interpolated W vectors, along with the original pair, were fed into the generator with zero noise, creating a series of images as shown in Figure 4. Figure 5 displays H&E images synthesized using randomly selected latent codes corresponding to different cell states, represented in the 3 columns. The images are arranged in rows, from top to bottom, showcasing the cell states CD39 CD8- , CD39+CD8+, and CD39 CD8+respectively. In a preferred embodiment, the artificially synthesized H&E images are directed toward augmenting (e.g. adding to) the testing dataset for the deep learning classifier model. The augmented testing set is utilized to evaluate the performance of the classifier, thereby facilitating improved understanding of classifier behaviour and enhancing the interpretability of model predictions.

[0055] Furthermore, to explore the distribution and relationships among all three cell classes, an equal number of W vectors associated with each class are retrieved. These were visualized using principal component analysis (PCA). The PCA revealed that CD8+T-cells, regardless of their tumor-specificity state, exhibit histomorphological similarities but remain distinguishable shown in Figure 6. Intriguingly, non-CD8 (i.e., CD8 CD39 ) cells cluster more closely with tumorspecific CD8+T-cells than with non-tumor-specific CD8+CD39_counterparts in Figure 6. The PCA plot illustrates the distribution of different cell classes, highlighting four exemplary trajectories that depict interpolated transitions between non-tumor-specific and tumor-specific T-cell states. This finding, corroborated by meticulous histological examination, suggests a unique histomorphological relationship influenced by tumor specificity in T-cells. This insight could significantly impact our understanding of T-cell functionality in the Tumor Microenvironment.

[0056] In various embodiments, the present invention is able to identify the change in observable histomorphological features of a CRC tumor cell as the cancer progresses, in particularly between Stage I and Stage IV tumor cells. 578 H&E- stained whole slide images (WSI) of CRC with the AJCC pathologic tumor stage (Stage I, II, III or IV) of the patients were collected from The Cancer Genome Atlas (TCGA). Tumor cell annotation was carried out by a pathologist using machine learning in QuPath. Tumor cell images of size 64 by 64 pixels were extracted, and each tumor cell was labelled with the tumor stage of the patient.

[0057] The conditional StyleGAN3 model was trained using tumor cells of Stage I and Stage IV, to model the cell image features in the latent space, represented by the latent vector W. Figure 7 illustrates PCA visualization of the latent vectors W, conditioned on two disease stages, randomly retrived from the trained StyleGAN. The PCA plot illustrates the distribution of the latent vectors. To characterize the histomorphological transitions of tumor cells progressing from Stage I to Stage IV, a decision boundary was established to distinguish between different disease stages based on the W vectors. A Support Vector Machine (SVM) classifier was trained using these W vectors to derive a hyperplane that functions as the decision boundary, as illustrated in Figure 8.

[0058] The hyperplane is characterized by a normal vector, labeled as wjtoiv. wJtoIV vector also represents a transition direction from Stage I to Stage IV. To visualize the histomorphological transitions of a tumor cell progressing from Stage I to Stage IV, the 1 / 1 / vectors corresponding to Stage I tumor cells were extrapolated along a defined trajectory, denoted as wjtoiv. A series of images was generated along this extrapolation path, thereby illustrating the morphological progression of the tumor cell as the disease advances from Stage I to Stage IV, as depicted in Figure 9.

[0059] As observed in the transition images shown in Figure 9, tumor cells in Stage IV are more enlarged than tumor cells in Stage I, suggesting that tumor cell size may be a potential meaningful H&E features for understanding of tumor cell phenotypes. Besides, the direct manipulation of cell images from one state to another enables us to capture the more important feature in the transition, as compared to interpolating between two random cells of different states.

[0060] It will be appreciated that many further modifications and permutations of various aspects of the described embodiments are possible. Accordingly, the described aspects are intended to embrace all such alterations, modifications, and variations that fall within the spirit and scope of the appended claims.

[0061] Throughout this specification and the claims which follow, unless the context requires otherwise, the word "comprise", and variations such as "comprises" and "comprising", will be understood to imply the inclusion of a stated integer or step or group of integers or steps but not the exclusion of any other integer or step or group of integers or steps.

[0062] The reference in this specification to any prior publication (or information derived from it), or to any matter which is known, is not, and should not be taken as an acknowledgment or admission or any form of suggestion that that prior publication (or information derived from it) or known matter forms part of the common general knowledge in the field of endeavour to which this specification relates.

Claims

Claims1. A computer-implemented training method for cell-level detection and prediction of cancer biomarker associated features, comprising: obtaining, for each of a plurality of cancer biopsy samples, a multiplex immunohistochemistry (mIHC) image and corresponding hematoxylin and eosin (H&E) stained tissue image, the mIHC images having corresponding cell labels including one or both of cell specificity states and cancer stage; aligning the corresponding mIHC and H&E stained tissue images; obtaining a labelled dataset by labelling the H&E images based on cell labels; training a conditional style generative adversarial network model (conditional StyleGAN), using the labelled dataset, to map cell image features of H&E stained tissue images onto a latent space and generate labelled synthetic H&E images, corresponding to one or both of a said specified cell specificity state and cancer stage, from the latent space; and training a deep learning classifier model, using the labelled cell tile dataset, for classifying specified cell specificity states.

2. The method of claim 1, further comprising performing cell-level detection and prediction of cancer biomarkers by: processing a further H&E stained image from a biopsy tissue slice using the deep learning classifier model to categorize the further H&E stained image into one or more said cell specificity states or cancer stage; and generating confidence probability of prediction of specificity state or cancer stage.

3. The method of claim 2, further comprising receiving one or more specified cell specificity states and / or a said cancer stage and processing thefurther H&E stained image using the conditional StyleGAN to generate a synthetic H&E image corresponding to the biopsy tissue slice, shown as a H&E stained image, at the one or more specified cell specificity states and / or the said cancer stage.

4. The method of any one of claims 1 to 3, further comprising segmenting individual cells within the H&E stained images into image patches, wherein obtaining the labelled cell tile dataset comprises labelling the image patches based on the cell labels from mIHC image tiles, each patch corresponding to one or more tiles.

5. The method of claim 1, wherein the deep learning classifier model is tested, using the labelled synthetic H&E images.

6. The method of any one of claims 1 to 5, wherein alignment involves digital alignment using the Scale-Invariant Feature Transform (SIFT) and B- spline transformation techniques.

7. The method of claims 4, wherein the image patches are 64 x 64, 96 x 96, 100 x 100, or 112 x 112 pixel image tiles.

8. The method of any one of claims 1 to 3, wherein the cell specificity states are T-cell tumor specificity states.

9. The method of any one of claims 1 to 8, wherein each said cancer stage specifies one of a plurality of progressive stages of cancer, including Stage I, Stage II, Stage III, and Stage IV.

10. A computer system for configurable for cell-level detection and prediction of cancer biomarkers associated features, comprising : memory; at least one processor; an image alignment module;a labeller module; a classifier module; and a generative adversarial network module, the memory storing instructions that, when executed by the at least one processor, cause the system to: obtain, for each of a plurality of cancer biopsy samples, a multiplex immunohistochemistry (mIHC) image and corresponding hematoxylin and eosin (H&E) stained tissue image, the mIHC images having corresponding cell labels including one or both of cell specificity states and cancer stage; align, at the alignment module, the corresponding mIHC and H&E stained tissue images; use the labeller module to obtain a labelled dataset by labelling the H&E images based on the cell labels; train a conditional style generative adversarial network model (conditional StyleGAN), using the labelled dataset, to map cell image features of H&E stained tissue images onto a latent space and generate labelled synthetic H&E images, corresponding to one or both of a said specified cell specificity state and cancer stage, from the latent space; and train a deep learning classifier model of the classifier module, using the labelled cell tile dataset, for classifying cell specificity states.

11. The system of claim 10, being configured, for cell-level detection and prediction of cancer biomarkers, to: receive a further H&E stained image from a biopsy tissue slice; process the further H&E stained image using the deep learning classifier model to categorize the H&E stained image into one or more said cell specificity states or cancer stage; and generate confidence probability of prediction of specificity state or cancer stage.12.The system of claim 11, being further configured to receive one or more specified cell specificity states and / or a said cancer stage and process the further H&E stained image using the conditional StyleGAN to generate a synthetic H&E image corresponding to the biopsy tissue slice, shown as a H&E stained image, at the one or more specified cell specificity states and / or said cancer stage.13.The system of any one of claims 10 to 12, wherein the labeller module is further configured to segment individual cells within the H&E stained images into image patches, wherein obtaining the labelled cell tile dataset comprises labelling the image patches based on the cell labels from mIHC image tiles, each patch corresponding to one or more tiles.14.The system of claim 10, wherein the deep learning classifier model is tested, using the labelled synthetic H&E images.15.The system of any one of claims 10 to 14, wherein alignment involves digital alignment using the Scale-Invariant Feature Transform (SIFT) and B-spline transformation techniques.16.The system of claim 13, wherein the image patches are 64 x 64, 96 x 96, 100 x 100, or 112 x 112 pixel image tiles17.The system of any one of claims 10 to 16, wherein the cell specificity states are T-cell tumor specificity states.18.The method of any one of claims 10 to 17, wherein each cancer stage specifies one of a plurality of progressive stages of cancer, including Stage I, Stage II, Stage III, and Stage IV.

19. A computer-implemented training method for cell-level visualization of cancer biomarker associated features, comprising:obtaining, for each of a plurality of cancer biopsy samples a hematoxylin and eosin (H&E) stained tissue image, and cell labels including cancer stage, to produce a labelled dataset; training a conditional style generative adversarial network model (conditional StyleGAN), using the labelled dataset, to map cell image features of H&E stained tissue images onto a latent space and generate labelled synthetic H&E images, corresponding to a said cancer stage, from the latent space.

20. A computer-implemented training method for cell-level visualization of cancer biomarker associated features, comprising: obtaining, for each of a plurality of cancer biopsy samples, a multiplex immunohistochemistry (mIHC) image and corresponding hematoxylin and eosin (H&E) stained tissue image, the mIHC images having corresponding cell labels including one or both of cell specificity states and cancer stage; aligning the corresponding mIHC and H&E stained tissue images; obtaining a labelled dataset by labelling the H&E images based on cell labels; training a conditional style generative adversarial network model (conditional StyleGAN), using the labelled dataset, to map cell image features of H&E stained tissue images onto a latent space and generate labelled synthetic H&E images, corresponding to one or both of a said specified cell specificity state and cancer stage, from the latent space.

Images