A computer implemented method for detecting a tertiary lymphoid structure (TLS)

Abstract

Conventionally, tertiary lymphoid structure identification is performed manually by pathologists. However, identifying TLS is a subjective procedure that results in issues of reproducibility caused by different pathologists making different decisions. The present invention relates to a computer-implemented method to provide decision support to clinicians and improve the accuracy and efficiency of TLS detection, preferably mature TLS detection in an image of a tumor tissue section previously stained with B-cell, T-cell, follicular B-cell and mature dendritic cell markers by applying a trained machine-learning model.

Description

[0001] A COMPUTER IMPLEMENTED METHOD FOR DETECTING A TERTIARY LYMPHOID STRUCTURE (TLS)

[0002] FIELD OF THE INVENTION

[0003] The present invention relates to a computer-implemented method for detecting tertiary lymphoid structure (TLS), preferably mature TLS in an image of a tumor tissue section previously stained with B-cell, T-cell, follicular dendritic cell (FDC) and mature dendritic cell (mDC) markers by applying a trained machine-learning model.

[0004] BACKGROUND OF THE INVENTION

[0005] Tertiary Lymphoid Structures (TLS) are ectopic and transient lymphoid aggregates associated with inflammation. High densities of TLS correlate with long-term survival for patients in most solid cancers (Schumacher TN and Thommen DS. Science (2022), 375(6576):eabf9419) such as in non-small-cell lung cancer (NSCLC) (Dieu-Nosjean et al (2008). J Clin Oncol 26:4410- 4417), suggesting that TLS facilitate the development of local anti-tumor immune responses. They are composed of a B-cell zone, organized as a germinal center and containing mainly naive B cells, follicular dendritic cells (FDC), and follicular-helper T cells (Tfh); surrounded by a T-cell rich area, where mature dendritic cells (mDC) interact with T cells (Germain C, et al (2014). Am J Respir Crit Care Med 189:832-844; Garaud S, Dieu-Nosjean M-C, and Willard- Gallo K (2022). Nat Commun 13:2259). Depending on the maturation stage of the TLS, high endothelial venules (HEV) can be found close to or crossing TLS, which can favor the local recruitment of peripheral blood immune cells expressing CD62L, the receptor of PNAd. Some of these actors are especially associated with improved outcomes for cancer patients such as the presence of TLS-DC-LAMP+mDC (Dieu-Nosjean M-C, Antoine M, Danel C, et al (2008). J Clin Oncol 26:4410-4417) and PNAd+HEV (Martinet L, et al (2011). Cancer Res 71:5678- 5687). More recently, the organization of the immune infiltrate into TLS and HEV presence have been associated with response to immunotherapy, particularly to immune checkpoint blockade (Cabrita R, Lauss M, Sanna A, et al (2020). Nature 577:561-565; Helmink BA, et al. (2020), Nature 577:549-555; Petitprez F, et al (2020). Nature 577:556-560; Patil N. et al. Cancer Cell. 2022 Mar 14;40(3):289-300; Asrir A, et al (2022). Cancer Cell SI 535610822000046; Liu Z, Meng X, Tang X, et al (2023). Cancer Immunol Immunother CII 72: 1505-1521; Wang Q, Sun K, Liu R, et al (2023). Clin Transl Med 13:el346; Kinker GS, Vitiello GAF, Diniz AB, et al (2023). Gut 72: 1927-1941). These findings highlight the role of TLS as a predictive marker of response to anti-cancer treatment.

[0006] Depending on the cellular composition and activation, TLS can progress through at least three sequential maturation stages (Kroeger DR, Milne K, and Nelson BH (2016). Clin Cancer Res Off J Am Assoc Cancer Res 22:3005-3015; Shu DH, Ho WJ, Kagohara LT, et al (2023) BioRxiv Prepr Serv Biol 2023.10.16.562104). Firstly, accumulating B cells and T lymphocytes in response to chemokine cues forms an immature TLS or immune aggregate. Secondly, follicular dendritic cells (FDCs) differentiate to generate TLS resembling primary follicles in secondary lymphoid organs (SLO). Finally, a germinal center reaction develops, resulting in TLS with all the characteristics of secondary follicles of SLO (Garaud S, Zayakin P, Buisseret L, et al (2018). Front Immunol 9:2660; Posch F, Silina K, Leibl S, et al (2018). Oncolmmunology 7:el378844). The anatomical segregation of B and T cells into distinct areas depends on chemokine secretion (mainly CXCL13 / CXCR5 and CCL19-CCL21 / CCR7 axis, respectively) (Teillaud, Houel, Panouillot et al. (2024). Nat Rev Cancer 24(9):629-646).

[0007] The detection of TLS predominantly relies on pathological diagnosis, including Hematoxylin and Eosin (HE) staining, Immunohistochemistry (IHC), and Immunofluorescence (IF). However, false negative results can arise from the failure to collect tumor tissue or when dealing with atypical samples. Indeed, conventionally, tertiary lymphoid structure identification is performed manually by pathologists. However, identifying TLS is a subjective procedure that results in issues of reproducibility caused by different pathologists making different decisions.

[0008] To date, applying deep learning-based medical image analysis to Computer- Assisted Diagnosis (CAD) could provide decision support to clinicians and improve the accuracy and efficiency of various processes for diagnosis and treatment. However, to date, the medical image analysis only applies to the hematoxylin-eosin staining section and does not allow analysis of specific tissue or cell markers to accurately determine the TLS density and its maturity stage. There remains a need to develop a method to accurately determine the TLS density and its maturity stage.

[0009] SUMMARY

[0010] In the present application, the inventors have developed a qualitative and quantitative approach allowing accurate analysis of the tertiary lymphoid structure (TLS) density and maturity. The technic is based on the localization of at least: B cell, follicular dendritic cell (FDC), T cell and mature dendritic cell (mDC) markers in an image of a tumor tissue section to realize precise quantitative analysis of tissues, TLS, and cell phenotypes using deep learning method.

[0011] The present disclosure relates to a computer-implemented method for detecting a tertiary lymphoid structure (TLS) in at least one image of a tumor tissue section previously stained with at least: B cell marker such as CD20, follicular dendritic cell (FDC) marker such as CD21, T cell marker such as CD3, and mature dendritic cell (mDC) marker such as DC-LAMP / CD208, wherein the method comprises the steps of: a) determining a region of interest, preferably comprising tumoral tissue, in said image, b) segmenting tissue in the region of interest via a trained model, c) locating B-cell zone and T-cell zone within the segmented tissue, FDC within B-cell zone and mDC within T cell zone via trained models, wherein the density of a TLS is determined by the densities of B-cell zone and T-cell zone, and mature TLS is determined by the presence of FDC within B-cell zone and mDC within T cell zone. In a preferred embodiment, said tumor tissue section is stained by immunohistochemistry or immunofluorescence with an anti-CD20, anti-CD21, anti-CD3, and anti-DC-LAMP antibodies.

[0012] In a specific embodiment, the TLS is detected in two images of adjacent tumor tissue sections, preferably in a first section stained with B cell and FDC markers and a second adjacent section to the first one stained with T cell and mDC markers, more preferably wherein in step a) the region of interest is determined in two images of adjacent tumor tissue sections, and in step b) B-cell zone and FDC are located in the image section via a trained model, and T-cell zone and mDC are located, in the adjacent section via a trained model.

[0013] In a more particular embodiment, said tumor tissue section is further stained with a nucleic marker, preferably DAPI and the method comprises, preferably before step c) of locating B- cells, T-cells, FDC and mDC within the segmented tissue, a step of segmenting each cell in said tissue via a trained model, preferably wherein the model has been trained by supervised learning on a training dataset formed by images of tumor tissue sections previously stained with nuclei markers and segmented into cell and nuclei.

[0014] In a preferred embodiment, the models in steps b) and c) have been trained by supervised learning on a training dataset formed by images of tumor tissue sections previously stained with at least: B cell, FDC, T cell and mDC markers, and manually segmented into tissue, zone of cells or cells classes of interest.

[0015] In a particular embodiment, said tumor tissue section is further stained with tumor cell marker such as pan-cytokeratins marker and / or high endothelial venule (HEV) cell marker such as Peripheral node addressin (PNAd) marker, preferably by immunohistochemistry or immunofluorescence with anti-pan-cytokeratins antibodies or an anti-PNAd antibody and preferably the method comprises in the step c) further locating tumoral cells and / or HEV cells via a trained model and wherein the maturity of the TLS is determined by the density of FDC within B-cell zones, mature DC within T-cell zones and the densities of high endothelial venules within the tumoral stroma. In a preferred embodiment, the model in step c) has been trained by supervised learning on a training dataset formed by images of tumor tissue sections previously stained with: B cell, FDC, T cell, mature DC markers, HEV and / or tumoral cell markers and manually segmented into the tissue, zone of cells or cells classes of interest.

[0016] In a preferred embodiment, the method as described above may comprise a step of exporting the count, surface, coordinates or mean intensity of each tissue, cell or zone of cells.

[0017] In a preferred embodiment, said tumor section is obtained from a patient suffering from cancer, preferably selected from the group consisting of lung cancer, colorectal cancer, cervical cancer and breast cancer, preferably a non-small cell lung cancer patient.

[0018] In another aspect, the present disclosure also relates to a method for prognosing the survival outcome of a patient suffering from a cancer or evaluating the response to a cancer therapy of a patient suffering from a cancer comprising detecting the tertiary lymphoid structure by a computer-implemented method as described above, wherein a higher TLS density, preferably a higher mature TLS density compared to a predetermined reference value is indicative of a higher survival outcome in said patient or higher response to cancer therapy in said patient.

[0019] In another aspect, the present disclosure relates to a computer program product comprising code instructions which, when executed by a computer, cause said computer to implement the methods disclosed above.

[0020] According to another aspect, the present disclosure relates to a non-transitory computer- readable storage medium having stored thereon code instructions which, when executed by a computer, cause said computer to implement the methods disclosed above.

[0021] According to another aspect, the present disclosure relates to a computing system, comprising at least one memory and one or more processors, configured to carry out the methods disclosed above.

[0022] DETAILED DESCRIPTION OF THE INVENTION

[0023] Computer-implemented method for detecting TLS The present disclosure relates to a computer-implemented method for detecting a tertiary lymphoid structure (TLS), preferably a mature TLS in at least one image of a tumor tissue section previously stained with at least: B cell marker such as CD20, follicular dendritic cell (FDC) marker such as CD21, T cell marker such as CD3 and mature dendritic cell (mDC) marker such as DC-LAMP / CD208.

[0024] By tertiary lymphoid structure (TLS) or tumor-induced lymphoid structure, it is meant the organization of tumor-infiltrating leukocytes in the stroma of the tumor mass, composed of T- cell cluster (T-cell areas or zone) and B-cell cluster also called B-cell follicle (B-cell areas or zone).

[0025] Tertiary lymphoid structures are organized aggregates of immune cells into lymph node-like structures within the tumor microenvironment, characterized by a central B-cell zone (B-cell follicle) comprising a cluster of B-cells forming cell-cell contact adjacent to a rich T-cell zone. The TLS can develop from an immature stage to a mature stage.

[0026] An immature TLS is characterized by lymph-node like structure within the tumor microenvironment without a network of FDC within B-cell follicle and without mDC within T- cell zone.

[0027] A mature TLS is characterized by the segregation of T and B cells and the presence of a germinal center containing B cells and follicular dendritic cells surrounded by a T zone comprising mature DC. Mature TLS can also comprise High Endothelial Venules (HEV) within the tumor stroma that corresponds to specialized vessels facilitating the transport of lymphocytes from / to the peripheral blood.

[0028] In a particular embodiment, mature TLS can be characterized by a high density of a B-cell zone with FDC, a high density of T-cell zone with mature dendritic cells (mDC) with or without HEV within the tumor stroma, whereas immature TLS can be characterized by high densities of B- cell zone and T-cell zones but no FDC within B-cell zones, and / or no mDC within T-cell zone.

[0029] The computer-implemented method, according to the present disclosure, is performed on an image of a tumor tissue section previously stained with at least a B cell marker, a follicular dendritic cell (FDC) marker, a T cell marker and mature dendritic cell (mDC) marker and preferably tumoral cell marker, nucleus marker and / or high endothelial venule (HEV) marker, preferably at least B-cell marker, FDC marker, T-cell marker and mDC marker, more preferably with at least B cell marker, FDC marker, T cell marker, mature dendritic cell (mDC) markers, and HEV marker, again more preferably at least B-cell marker, FDC marker, T-cell marker, mDC marker, HEV marker and tumoral cell marker. In a preferred embodiment, said tumor tissue section is further stained with a nucleus marker.

[0030] As used herein, the term “image” or “digital image” refers to an electronic image represented by a collection of pixels that can be viewed, processed and / or analyzed by a computer. The image can be acquired by a digital camera or other optical device capable of capturing digital images from a slide or portion thereof. The image may have been acquired prior to the implementation of the method described below, and stored in a memory where it can be recovered by a computing system for implementation of the method.

[0031] The method may be implemented by a computing system comprising at least one processor (for instance of the type CPU and / or GPU) and at least one memory (for instance magnetic hard disk, solid-state disk, optical disk, electronic memory such as read-only memory, for instance of the type EEPROM, PROM, etc. or any type of computer-readable storage medium) (e.g., Figure 3) in which a computer program is stored, in the form of a set of program-code instructions to be executed by the one or more processors in order to implement all or part of the steps of the method detailed below. The memory may also store parameters of one or more trained models used during the implementation of the method.

[0032] In a particular aspect, the present disclosure relates to a computer program product comprising code instructions which, when executed by a computer, cause said computer to implement a method according to the present disclosure or a non-transitory computer readable storage medium having stored thereon code instructions which, when executed by a computer, causes said computer to implement the method according to the present disclosure or a computing system, comprising at least one memory and one or more processors, configured to carry out the method of the present disclosure.

[0033] As used herein, the term "tumor tissue sample" means any tumor sample tissue derived from the patient. Said tissue sample is obtained for the purpose of the in vitro evaluation. In some embodiments, the tumor sample may result from the tumor resected from the patient. Typically, the tumor tissue sample is fixed in formalin and embedded in a rigid fixative, such as paraffin (wax) or epoxy, which is placed in a mold and later hardened to produce a block that is readily cut. Thin slices of material can then be prepared using a microtome, placed on a glass slide, and submitted e.g. to immunohistochemistry (IHC) (using an IHC automate such as BenchMark® XT, for obtaining stained slides). The terms "subject" and "patient" are used interchangeably herein and refer to both human and non-human animals. As used herein, the term “patient” denotes a mammal, such as a rodent, a feline, a canine, and a primate. Preferably, a patient according to the invention is a human, preferably a human cancer patient, preferably a human solid cancer patient.

[0034] The terms “cancer”, “tumor”, are used interchangeably herein to refer to cells that exhibit relatively abnormal, uncontrolled, and / or autonomous growth, so that they exhibit an aberrant growth phenotype characterized by a significant loss of control of cell proliferation. In a preferred embodiment, said cancer is selected from the group consisting of lung cancer, colorectal cancer, cervical cancer and breast cancer, preferably a non-small cell lung cancer patient.

[0035] According to the present disclosure, said specific cell markers can be cell surface protein (also called herein cell surface antigens), intracellular protein (also called herein intracellular antigens), or gene that is specifically expressed by said cell type. In a preferred embodiment, said specific marker is a cell surface or intracellular protein specifically expressed by said cell type.

[0036] Each cell type can be identified by using one or a combination of specific cell markers. Specific cell markers are well-known in art and some examples are provided below.

[0037] As used herein, the term "T cell" has its general meaning in the art and includes cells within the T cell lineage, including thymocytes, immature T cells, mature T cells and the like. According to the present disclosure, said T cell can be a CD3+T cell, or CD8+and / or CD4+cells, FoxP3+regulatory T cells, preferably CD3+T cells. In a preferred embodiment, CD3 surface antigen is detected by one of the several antibodies that specifically bind to CD3 that have been described in the prior art and are commercially available. FoxP3 is a transcriptional factor (intracellular antigen) that stains a subset of T cells, CD4+and CD8+cells, with regulatory immune function. T-cell zone refers to T-cell subsets that form cell-cell contact clustered in the T-cell zone. T-cells in T-cell zones can be identified by using the T-cell marker, preferably a pan surface protein marker expressed by T-cells such as CD3 or CD8, or a combination CD3 and CD4, CD3 and CD8, preferably CD3.

[0038] As used herein, the term “B cell” or “B lymphocyte” has its general meaning in the art and refers to immune cell that produce antibodies and generate immunological memory, function as antigen-presenting cell, and secrete cytokines. Surface antigen and / or intracellular markers for B cells that can be used for distinguishing B cells from other cell types are well-known in the art and can be selected as non-limiting examples in the group consisting of: CD 19 (pan-B cells), B220 (CD45R), CD20 (pan-B cells), IgD and Bcl2 (naive B cells), IgA, IgG and IgE (switched B cells), activation-induced cytidine deaminase (AID) and Bcl6 (germinal center B cells), CD27 (memory and marginal zone B cells), CD38 in combination with CD138, CD78, IgG, CD27 (CD38+plasma cells and CD38' memory B cells), CD24 (regulatory B cells), CD25 and CD30 (activated B cells) or a combination thereof. In a preferred embodiment, said B-cell marker is a pan-B-cell marker, such as CD 19 or CD20, preferably CD20. B-cell zone or follicular B cells refers to B cell subsets that form cell-cell contact clustered in a B-cell follicle. B-cells in B-cell zones can be identified by using the B-cell marker, preferably a pan surface protein marker expressed by B-cells such as CD 19 or CD20, preferably CD20.

[0039] As used herein, the term “dendritic cell” has its general meaning in the art and refers to antigen- presenting cell that present processed antigens on the cell surface to T cells. According to the present disclosure, dendritic cells can be distinguished from other cells by detecting surface antigen and / or intracellular markers, for example selected in the group consisting of: CD la, CDl lc (ITGAX), HLA-DR, CLEC9A, BDCA-3 / CD141, Langerin / CD207, BDCA-2 / CD303, BDCA-4 / Neuropilin-l / CD304, BDCA-1 / CDlc, or a combination thereof, preferably CDl lc. Said surface antigens and / or intracellular markers can be detected by one of the several antibodies that specifically bind to said surface antigens and / or intracellular markers that have been described in the prior art and are commercially available.

[0040] By mature dendritic cells (mDC), it means a population of dendritic cells that are professional for the presentation of processed antigens to T cells. The vast majority of mature dendritic cells infiltrating the tumor are selectively located in contact with T cells, in the T-cell rich areas of the tumor-induced lymphoid structure. Mature dendritic cells can be distinguished from other cells by detecting surface and / or intracellular antigen markers, for example DC-lysosome associated membrane glycoprotein (DC-LAMP, intracellular antigen) also named CD208. Only in lungs, DC-Lamp can be expressed by Type II pneumocytes but with a distinct morphology (staining of multiple intracellular vesicles for Type II pneumocytes in contrast to intracellular dot staining for mDC). Said antigens can be detected by one of the several antibodies that specifically bind to said antigens that have been described in the prior art and are commercially available.

[0041] As used herein, the term “follicular dendritic cell” or “FDC” has its general meaning in the art and refers to stromal cells of fibroblastic origin found in B cell follicles of lymphoid tissue including TLS. FDC can be distinguished from other cells by detecting surface antigen markers, for example selected in the group consisting of: the long isoform of CD21 (in contrast to B cells that express the short isoform of CD21 lacking the exon 10a), CD23, or CD35 or a combination thereof, preferably CD21. Said surface antigens can be detected by one of the several antibodies that specifically bind to said surface antigens that have been described in the prior art and are commercially available.

[0042] As used herein, “High endothelial venules” or “HEV” has its general meaning in the art and refers to specialized post-capillary venules characterized by plump endothelial cells as opposed to the usual flatter endothelial cells found in regular venules. HEVs enable immune cells circulating in the blood to directly enter a lymph node (by extravasation through the HEV). HEV can be characterized and distinguished from the other blood vessels (pan-blood vessel markers: CD34 class II, CD31) by detecting peripheral node addressin (PNAd) marker, in particular within the tumoral stroma. Said surface antigens can be detected by one of the several antibodies that specifically bind to said surface antigens that have been described in the prior art and are commercially available.

[0043] Tumor cells or cancer cells are cells that divide continually, forming solid tumors or flooding the blood or lymph with abnormal cells. For example, tumor cell markers can be any antigens expressed at the surface of the tumor cell. In a preferred embodiment, the tumor cells, in particular epithelial tumor tissues, can be identified by detecting cytokeratins. Pan-cytokeratin antibodies well known in the art can be used to stain all epithelial tumors and can be used to detect tumoral tissue. For example, pan-Cytokeratins (mix of clones AE1 and AE3), a cocktail of antibodies that targets a broad range of cytokeratins can be used to identify epithelial-derived tumors, cytokeratins 5 / 6 (CK5 / 6) can be used to identify squamous cell carcinomas and mesotheliomas, cytokeratin 7 (CK7), can be used to identify tumors originating from the lung, breast, ovary, and pancreas, cytokeratin 20 (CK20) can be used to identify colorectal and Merkel cell carcinomas, epithelial Membrane Antigen (EMA) can be used as a marker for various carcinomas, Carcinoembryonic Antigen (CEA) can be used in colorectal and pancreatic cancers, p63 and p40 can be used as markers for squamous cell carcinomas, mucins (e.g., MUC1, MUC2) can be used in the identification of adenocarcinomas, and claudins and cadherins can be used in differentiating various types of epithelial tumors. In a preferred embodiment, a cocktail of antibodies that targets a broad range of cytokeratins is used, preferably pan-Cytokeratins (both clones AE1 and AE3). In a particular embodiment, specific cell markers such as surface antigen and intracellular markers, as described above, can be detected with an antibody such as a polyclonal antibody, monoclonal antibody, antigen-binding fragment or antibody mimetic using any well-known methods in the art such as immunofluorescence, immunohistochemistry, as described in the examples 1 and 2 of the present application. The reactions generally include revealing labels such as fluorescent, chemiluminescent, radioactive, enzymatic labels or dye molecules, or other methods for detecting the formation of a complex between the antigen and the antibody or antibodies reacted therewith.

[0044] Several antibodies that specifically bind to said surface antigen and intracellular antigen as described above have been described in the prior art and are commercially available and for example are listed in antibody databank such as on-line databank Antibodypedia (Kiermer V. (2008) “Antibodypedia - A web portal to share antibody validation data” Nature Methods. 5: 860).

[0045] According to the present disclosure, the tumor tissue section can also be stained with a nucleus marker to segment all the cells in the image of the tumor tissue section. The nucleus of the cell can be stained by any dye known in the art, such as non-limiting examples, DAPI (4’,6- diamidino-2-phenylindole), Hoechst 33342, Propidium Iodide, DRAQ5™, DRAQ7™, CyTRAK orange, or 7-AAD (7- Aminoactinomycin D). Also, the nucleus of the cell can be counterstained by other dyes such as hematoxylin, eosin, and toluidine.

[0046] The present computer-implemented method is implemented by a digital image processing system configured to perform automated detection of TLS and preferably mature TLS.

[0047] According to the present disclosure, the computer-implemented method is implemented by a digital image processing system that may access an image of a tumor tissue section previously obtained from a subject. The computer-implemented method (e.g., digital image processing system) according to the present disclosure may detect TLS based on a trained model.

[0048] The sections can be presented to a digital camera on a microscope. The image can be captured at different magnifications using different objectives. In a preferred embodiment, the image is captured at a minimum of 20x magnification. The image that is processed by the digital image processing system may be recovered from the digital camera or from a memory where the digital image has been preliminarily stored. In a first step of the method of detecting TLS according to the present disclosure, a region of interest may be manually determined within the image by the user to exclude massive non- relevant areas such as blur, folding, healthy tissue outside the tumor mass. According to the present disclosure, the region of interest comprises tumoral tissue such as a global primary tumor (as a whole), a tissue sample from the center of the tumor, and / or a tissue directly surrounding the tumor which tissue may be more specifically named the "invasive margin" of the tumor that can comprise lymphoid organization (i.e., small lymphoid aggregates without any organization and segregation of the immune cells, immature and mature TLS) in close proximity with the tumor.

[0049] After determining the region of interest, a first segmentation step is performed by implementing, on the region of interest, a trained model to segment the tissue within the region of interest of the image, in particular to discriminate tissue from other part of the image such as areas without tissue, artifacts such as necrosis, red blood cells, dust, or tissue folds, and both macrophages and anthracosis. Then, in a second step, B-cell zone and T-cell zone are located within the segmented tissue, mature dendritic cells are located within T cell zone and FDC are located within B cell zone via a trained model on an image of tumor tissue section previously stained with B cell (e.g. CD20), T cell (e.g. CD3), FDC (e.g., CD21) and mDC (e.g., DC-LAMP) markers. By “located”, it is meant that the locations of the cells within the segmented tissue is determined.

[0050] In a specific embodiment, two images or more of adjacent tumor tissue sections stained with different markers as described previously are captured to localize the different classes of cells or cell-zones within the segmented tissue via the trained models. In particular, a first section is stained with B-cell and FDC markers and an adjacent section to the first section is stained with T-cell and mature DC markers. Then, the B-cell and T-cell zones are located within the segmented tissue of two adjacent section images to determine the densities of the B and T-cell zones, and FDC and mDC are located within the B-cell zone and T-cell zone, respectively of the two adjacent section images to determine the densities of FDC within B-cell zone and mature DC within T-cell zone. The densities of the B- and T-cell zones are indicative of the density of the TLS, and the densities of FDC within the B-cell zone or mature DC within the T- cell zone are indicative of the maturity of the TLS.

[0051] In another particular embodiment, the first segmentation step is performed by implementing, on the region of interest, a trained model to segment tumor tissue, TLS, and tumoral stroma without TLS (named “tumor stroma) within the region of interest of the image, in particular to discriminate tumor tissue, TLS and tumoral stroma from another part of the image such as areas without tissue, artifacts such as necrosis, red blood cells, dust, or tissue folds, and both macrophages and anthracosis. In this case, a priority order can be attributed to the different parts of the region of interest, i.e., the different classes of segmentation, for example from class 1 to 5, areas without tissue, stroma, artifacts, TLS and tumor.

[0052] In a particular embodiment, the cells can be segmented within the regions of interest via a trained model on an image of section previously stained with a nucleus marker (e.g., DAPI) to identify individually each nucleus and cells, preferably before locating (i.e., phenotyping) B cell, T cell, mature dendritic and FDC within segmented tissue via a trained model on an image of tumor tissue section previously stained with B cell (e.g. CD20), T cell (e.g. CD3), FDC (e.g., CD21) and mDC (e.g., DC-LAMP) markers. By “located”, it is meant that the locations of the cells within the segmented tissue is determined.

[0053] In a preferred embodiment, HEV are further located within the segmented tissue, particularly within tumoral stroma by applying a trained model on an image of a section previously stained with HEV marker such as PNAd antibody. The stroma consists of the part of the tissue surrounding the tumor tissue, as defined previously, that can be preferably marked with pan- cytokeratin markers.

[0054] In a preferred embodiment, the trained models are deep learning classifier models, such as neural networks, for instance convolutional neural networks, for instance having a U-Net, ResNet or MiniNet architecture.

[0055] The segmentation or localization step may be produced by external digital image analysis tools and passed along as a new digital image with the same image characteristics (size, resolution, pixel size) but where each pixel denotes tissue, classes of cells or cell zones instead of the color acquired by microscopy, or is labeled with an indication of tissue, class of cells or cell zones. Any finite number of pixel classes may be input to the system such as tissue, but also stroma tissue, necrosis tissue, artifacts, anthracosis or macrophages, and classes of cells or cell zones (e.g., B-cell zone, FDC, mDC, T-cell zone, tumor cells, HEV, cells).

[0056] In a particular embodiment, the models include at least one neural network classifier which may be trained to identify tissue, cell or cell-zone classes. For instance, the model includes a neural network classifier which may be trained to identify a tissue within the region of interest and for example discriminate the tissue from other areas of the region of interest such as areas without tissue (e.g., glass, holes in the tissue), artifacts such as necrosis, red blood cells, dust, blur, mucus, or tissue folds, or zones comprising macrophages and anthracosis, a neural network classifier to locate FDC, mDC, T cell zone, B-cell zone, tumor cells, HEV, and another neural network classifier to segment nuclei and cells within tissue. The other parts of the image not corresponding to tissue may be segmented within a unique class or may themselves be segmented according to the different categories recited above.

[0057] In embodiments, one or more models are trained as disclosed in Figure 1 with a training set of images of tumor tissue sections including manual annotations of tissue, cells or zone of cells of interest (e.g., tissue, FDC, mDC, T-cell zone, B-cell zone, tumor cells, HEV, nuclei, cells) from users such as pathologists. The different classes of tissue, zone of cells, cells are defined.

[0058] The models were trained and were validated only if, upon visual inspection, the matching accuracy between the segmented image and the native image was above 85%. In case of insufficient accuracy, the algorithm was further trained with additional examples, and the validation process was repeated. For instance, said trained model can be the Zen and HALO- AI® software.

[0059] The method according to the present disclosure may comprise a final step of exporting the count, surface, coordinates or mean intensity of each tissue, cells or zone of cells.

[0060] According to the present disclosure, in a particular embodiment, cell or cell zone density may be then defined as the number of cells per one unit of tissue of the region of interest. Cell density may be expressed as the number of these cells that are counted per cm2or mm2of tissue of the region of interest.

[0061] In another particular embodiment, the B- and T-cell zone densities may also be defined as the proportion of a cell-zone type among the tissue of the region of interest. For example, B-cell zone density or T-cell zone density can be measured as a total surface of B-cell zone or T-cell zone per one unit of surface area of the tissue in the region of interest, e.g., as the surface area of B-cell or T-cell zone in mm2per mm2of surface area of the tissue in the region of interest.

[0062] In another particular embodiment, the cell density may also be defined as the proportion of a cell type among the total cells, preferably a proportion of cell type such as FDC (e.g., CD21+) among the total B cells (e.g., CD20+) or B-cell zone; or mature DC (e.g., DC-LAMP) among the total T cells (e.g., CD3+) or T-cell zone, in the tissue of the region of interest. The densities of the B- and T-cell zones are indicative of the density of the TLS and the densities of FDC within the B cell-zone, mature DC within the T-cell zone, HEV within tumoral stroma are indicative of the maturity of the TLS.

[0063] In a particular embodiment, in B-cell zone, the germinal center is composed of FDC organized in dense mesh of cells with elongated fibrillar morphologies thus precise quantification of those individual cells cannot be obtained solely using CD21 intracellular marker as nuclei marker is required to delimit and segment FDC. Thus, when the cells are not segmented within the regions of interest via a trained model on an image of a section previously stained with a nucleus marker (e.g., DAPI) (e.g., IHC method), FDC were quantified in terms of surface. In contrast, when the cells are segmented within the regions of interest via a trained model on an image of section previously stained with a nucleus marker (e.g., DAPI) (e.g., immunofluorescence staining technique), the FDC could be detected as cells and thus be quantified in terms of density using their surface antigen marker CD21 and nuclei marker (DAPI).

[0064] A method for evaluating the response to cancer therapy of a patient suffering from a cancer

[0065] A higher value of the TLS density, preferably mature TLS density determined by the computer- implemented method, as described above in comparison to a control value is indicative that the cancer patient is responsive to cancer therapy.

[0066] According to the present disclosure, the terms "determining the response of a cancer therapy”, refers to an ability to assess whether the treatment (e.g., cancer therapy such as immunotherapy) is effective in (e.g., providing a measurable benefit or positive medical response to) the patient before and / or after some time of administration of the treatment. In another terms, according to the present disclosure, determining the response to cancer therapy refers to an ability to assess whether a tumor is responsive to cancer treatment and for example, whether following the cancer therapy the number of cancer cells or the size of a tumor is reduced, the progression of a cancer to a more aggressive form (i.e., maintaining the cancer in a form that is susceptible to a therapeutic agent) is reduced, the proliferation of cancer cells or of the speed of tumor growth are reduced, cancer cells are killed or the likelihood of recurrence of a cancer is reduced in a subject.

[0067] The present disclosure relates to a method for determining a response to a cancer therapy in a cancer patient. The cancer therapy may be chemotherapy or immunotherapy. Chemotherapy includes the use of cytotoxic anti-neoplastic agents, such as alkylating agents, platinum-based drugs, PARP inhibitors, ATM inhibitors or ATR inhibitors, anti-metabolites, anti -microtubule agents, Topoisomerase inhibitors, cytotoxic antibiotics and others. Examples of chemotherapeutic drugs include with no limitations: Capecitabine, 5-FU, docetaxel, SN-38, CPT11, cisplatin, carboplatin, etc.

[0068] Immunotherapy is a type of cancer treatment that activates the immune system to fight disease such as cancer. Immunotherapy that can be used to treat cancer includes as non-limiting examples: immune inhibitory checkpoint inhibitors which are drugs that block inhibitory immune checkpoint protein, T-cell transfer therapy, monoclonal antibodies or immune system activators.

[0069] As used herein the term "immune checkpoint protein" has its general meaning in the art and refers to a molecule that is expressed by T cells and NK cells and regulates the immune system. According to the present disclosure, immune checkpoint proteins are preferably inhibitory immune checkpoint proteins that dampen effector immune response. Inhibitory immune checkpoint molecules are recognized in the art to constitute immune checkpoint pathways similar to the CTLA-4 and PD-1 dependent pathways (see e.g., Pardoll, 2012. Nature Rev Cancer 12:252-264; Mellman et al., 2011. Nature 480:480- 489).

[0070] As used herein, the term "inhibitory immune checkpoint inhibitor" or “immune checkpoint inhibitor” has its general meaning in the art and refers to any compound inhibiting the function of an immune inhibitory checkpoint protein. Inhibition includes reduction of function and full blockade. In particular, the immune checkpoint inhibitor is particularly suitable for enhancing the proliferation, migration, persistence and / or cytotoxic activity of CD8+T cells in the patient and in particular the tumor-infiltrating of CD8+T cells of the patient.

[0071] In an embodiment, said immune checkpoint inhibitor of immune cells (T and B lymphocytes) is selected from the group consisting of anti-PD-Ll, anti-PD-1, anti-CTLA-4, anti-HVEM, anti- BTLA, anti-TIGIT, anti-TIM-1 / 3, anti-LAG-3, and anti-OX40 agonist, anti-CD40 agonist, CD40-L, TLR agonists, and B-cell receptor agonists, in particular selected from the group consisting of anti-PD-Ll, anti-PD-1 and anti-CTLA-4. In a particular embodiment of the disclosure, the immune checkpoint inhibitor is an anti-PD-Ll antibody.

[0072] In a particular embodiment, examples of immune checkpoint inhibitors are inhibitors that affect the PD-l / PDL-1 and CTLA-4 pathways and can be selected from the group consisting of: Ipilimumab, Nivolumab, Pembrolizumab, Atezolizumab, Avelumab and Durvalumab. According to the present disclosure, said immunotherapy can be T cell transfer therapy. As used herein, T-cell transfer therapy, also called adoptive cell therapy, adoptive immunotherapy, or immune cell therapy has its general meaning in the art and refers to a treatment that boosts the natural ability of T cells to fight cancer. In this treatment, immune cells are taken from the patient tumor. T-cell transfer therapy can be tumor-infiltrating lymphocytes or CAR T-cell therapy.

[0073] According to the present disclosure, immunotherapy can be monoclonal antibodies that bind to specific targets on cancer cells such as anti-CD20, anti-HER2, anti-EGFR, anti-VEGF, anti- CD52 or anti-CD33 antibody or immune system activators such as cytokines including as nonlimiting examples: interferon alpha (IFN-a), interleukin-2 (IL-2), interleukin- 11 (IL-11), interleukin-21 (IL-21), erythropoietin, granulocyte-macrophage colony-stimulating factor (GM-CSF) or granulocyte colony-stimulating factor (G-CSF).

[0074] Typically, the predetermined reference values for the cell density of follicular B cells and for the cell density of mature dendritic cells may be determined by applying statistical methods in large-scale studies on cancer patients.

[0075] A higher TLS density, preferably high mature TLS density in a tissue section obtained from a patient in comparison to a control value is indicative that the cancer patient is responsive to cancer therapy, preferably immunotherapy. A lower TLS density, preferably low mature TLS in a patient sample in comparison to a control value is indicative that the cancer patient is non- responsive to cancer treatment, preferably immunotherapy.

[0076] The therapeutic response can be evaluated according to the present method, before cancer treatment (e.g., immunotherapy) or throughout the course of cancer treatment (e.g., immunotherapy) for monitoring the therapeutic response over time.

[0077] According to the present method for evaluating the therapeutic response, the tumor tissue sample is previously collected from a patient having received a dose of a therapeutic agent. According to the present disclosure, a patient having received a therapeutic agent refers to a patient having received at least one dose of a therapeutic agent. The therapeutic response can be evaluated according to the present method, after each administration of a dose of a therapeutic agent throughout the course of treatment for monitoring the therapeutic response over time. In a preferred embodiment, the density of TLS, preferably the density of mature TLS is determined in a tumor tissue section collected from a patient at least 1, 2, 3, or 7, preferably 10, 14, 21, 28 days after therapeutic agent dose administration, more particularly at least 30, 35, 40, 45, 50, 55, 60, 70, 80, 90, 100, 120 days after therapeutic agent dose administration.

[0078] The density of TLS, preferably the density of mature TLS in a tumor tissue section of a patient having received at least one first dose of therapeutic agent during the treatment, preferably in comparison to the density of TLS, preferably the density of mature TLS in a tumor tissue section obtained from said patient at a prior time point of the treatment, preferably prior to the administration of at least one therapeutic agent, correlates with therapeutic response of said patient to therapeutic agent.

[0079] The method for evaluating therapeutic response according to the present disclosure can indicate success or failure of treatment to a patient.

[0080] A higher density of TLS, preferably mature TLS in a patient tumor tissue compared to a control value is indicative that the patient is responsive to said treatment. Typically, the density of TLS, preferably mature TLS is deemed to be higher than the control value if change in said patient to that of said control value is higher than at least 0.1, preferably 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, more preferably 1, 2, 3, 4 again more preferably 5.

[0081] A lower or similar density of TLS, preferably mature TLS in a patient sample compared to a control value is indicative that the patient is non-responsive to said treatment. If after treatment with the therapeutic agent, the density of TLS, preferably mature TLS in tumor tissue of a patient having received at least one dose of the therapeutic agent is not higher than a control value, the treatment should be interrupted or modified.

[0082] According to the present disclosure, the terms “threshold value”, “control value” or “cut-off value” can be used interchangeably and can be determined experimentally, empirically, or theoretically. The control value refers to the TLS density or mature TLS density in a tissue section obtained from a general population or from a selected population of subjects.

[0083] In a preferred embodiment, the control value refers to the TLS density or mature TLS density in a tissue section obtained from other source than the patient’s data, for example cancer patients who is not responsive to cancer treatment (e.g., immunotherapy) or with poor prognosis (e.g., a short disease-free survival time). The control value may be established based upon comparative measurements between cancer patients responsive to cancer therapy (e.g., immunotherapy) and cancer patients no responsive to cancer therapy (e.g., immunotherapy) and calculating the statistical significance between the TLS density or mature TLS density of a collection of patient tissue section and the corresponding responsiveness to cancer therapy (e.g., immunotherapy) of the cancer patients from which samples derive. Said predetermined reference value (cut-off value) can consists of the value for which the highest statistical significance is calculated. Typically, the control value can be determined using a Receiver Operating Characteristic (ROC) curve based on experimental data. For example, after determining the TLS density or mature TLS density in a group of reference, one can use algorithmic analysis for the statistic treatment of the measured values in samples to be tested, and thus obtain a classification standard having significance for sample classification. In a particular embodiment, Receiver operating characteristic (ROC) analysis was performed to calculate TLS density or mature TLS density using patient tumour tissue with known clinical status. The TLS density or mature TLS density offering the highest sensitivity and specificity was selected as cut-off point. This algorithmic method is preferably done with a computer. Existing software or systems in the art may be used for the drawing of the ROC curve, such as: MedCalc 9.2.0.1 medical statistical software, SPSS 9.0, ROCPOWER.SAS, DESIGNROC.FOR, MULTIREADER POWER.SAS, CREATE-ROC.SAS, GB STAT VIO.O (Dynamic Microsystems, Inc. Silver Spring, Md., USA), etc.

[0084] A method for prognosing the survival outcome of a patient suffering from a cancer

[0085] A higher value of the TLS density, preferably mature TLS density determined by the computer- implemented method as described above in comparison to a control value is indicative of a higher survival time of a cancer patient.

[0086] As used herein, the term “survival time” refers to the length of time from either the date of diagnosis or the start of treatment, that patients diagnosed with the disease are still alive.

[0087] The survival time can be the disease-specific survival (DSS) (i.e., the percentage of people in a study or treatment group who have not died from a disease in a defined period of time), the disease-free survival (DFS) (i.e., the number of people who have no cancer after treatment), or the overall survival (OS) (i.e., the time from diagnosis of the cancer or the surgery (removal of the tumor) until death from any cause).

[0088] In a particular embodiment, survival time denotes the percentage of people in a study or treatment group who are still alive for a certain period of time after they underwent a surgery to remove their tumor or a certain period of time after they were diagnosed with or started treatment for a disease, such as cancer. The survival time rate is often stated as a five-year survival rate, which is the percentage of people in a study or treatment group who are alive five years after their surgery or diagnosis or the start of treatment.

[0089] A higher TLS density, preferably high mature TLS density in a tissue section obtained from a patient in comparison to a control value is indicative that the cancer patient has a higher survival time. A lower TLS density, preferably low mature TLS in a patient sample in comparison to a control value is indicative that the cancer patient has a lower survival time.

[0090] A higher density of TLS, preferably mature TLS in a patient tumor tissue compared to a control value is indicative that the patient has a higher survival time. Typically, the density of TLS, preferably mature TLS is deemed to be higher than the control value if change in said patient to that of said control value is higher than at least 0.1, preferably 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, more preferably 1, 2, 3, 4 again more preferably 5.

[0091] A lower or similar density of TLS, preferably mature TLS in a patient sample compared to a control value is indicative that the patient has a lower survival time.

[0092] According to the present disclosure, the terms “threshold value”, “control value” or “cut-off value” can be used interchangeably and can be determined experimentally, empirically, or theoretically. The control value refers to the TLS density or mature TLS density in a tissue section obtained from a general population or from a selected population of subjects. For example, the general population may comprise apparently healthy subjects, such as individuals who have not previously had any sign or symptoms indicating the presence of cancer. The term "healthy subjects" as used herein refers to a population of subjects who do not suffer from any known condition, and in particular, who are not affected with any cancer.

[0093] In a preferred embodiment, the control value refers to the TLS density or mature TLS density in a tissue section obtained from other source than the patient’s data, for example cancer patients with poor prognosis (e.g., a short disease-free survival time).

[0094] Said predetermined reference value (cut-off value) can consists of the value for which the highest statistical significance is calculated. The control value may also be established based upon comparative measurements between cancer patients with poor or good prognosis (e.g., a long disease- free survival time). Typically, the control value can be determined using a Receiver Operating Characteristic (ROC) curve based on experimental data. For example, after determining the TLS density or mature TLS density in a group of reference, one can use algorithmic analysis for the statistic treatment of the measured values in samples to be tested, and thus obtain a classification standard having significance for sample classification. In a particular embodiment, Receiver operating characteristic (ROC) analysis was performed to calculate TLS density or mature TLS density using patient tumour tissue with known clinical status. The TLS density or mature TLS density offering the highest sensitivity and specificity was selected as cut-off point. This algorithmic method is preferably done with a computer. Existing software or systems in the art may be used for the drawing of the ROC curve, such as: MedCalc 9.2.0.1 medical statistical software, SPSS 9.0, ROCPOWER.SAS, DESIGNROC.FOR, MULTIREADER POWER.SAS, CREATE-ROC.SAS, GB STAT VIO.O (Dynamic Microsystems, Inc. Silver Spring, Md., USA), etc.

[0095] Method of treatment

[0096] In connection with the above methods for detecting TLS, the present disclosure relates to the treatment of a cancer of a patient previously determined as having a high density of TLS, preferably mature TLS using the method for detecting TLS as described above.

[0097] The present disclosure relates to a method of treating a cancer in a subject comprising detecting a TLS by the method as described above and administering to said subject a therapeutically effective amount of a cancer agent (e.g., immunotherapy), as described above.

[0098] As used herein, a "therapeutically effective amount" or an "effective amount" means the amount of a composition that, when administered to a subject for treating a state, disorder or condition is sufficient to affect a treatment. The therapeutically effective amount will vary depending on the compound, formulation or composition, the disease and its severity and the age, weight, physical condition and responsiveness of the subject to be treated.

[0099] As used herein, the term “treatment”, “treat” or “treating” refers to any act intended to ameliorate the health status of patients such as therapy, prevention, prophylaxis and retardation of the disease. In certain embodiments, such term refers to the amelioration or eradication of a disease or symptoms associated with a disease. In other embodiments, this term refers to minimizing the spread or worsening of the disease resulting from the administration of one or more therapeutic agents to a subject with such a disease.

[0100] The cancer agent described herein may be administered by any means known to those skilled in the art, including, without limitation, intravenously, orally, intra-tumoral, intra-lesional, intradermal, topical, intraperitoneal, intramuscular, parenteral, subcutaneous and topical administration. Thus, the compositions may be formulated as an injectable, topical, ingestible formulation. Administration of the cancer agent to a subject in accordance with the present invention may exhibit beneficial effects in a dose-dependent manner. Thus, within broad limits, administration of larger quantities of the compositions is expected to achieve increased beneficial biological effects than administration of a smaller amount. Moreover, efficacy is also contemplated at dosages below the level at which toxicity is seen.

[0101] It will be appreciated that the specific dosage of cancer agent administered in any given case will be adjusted in accordance with the composition or compositions being administered, the volume of the composition that can be effectively delivered to the site of administration, the disease to be treated or inhibited, the condition of the subject, and other relevant medical factors that may modify the activity of the compositions or the response of the subject, as is well known by those skilled in the art.

[0102] For example, the specific dose of a cancer agent for a particular subject depends on age, body weight, general state of health, diet, the timing and mode of administration, the rate of excretion, medicaments used in combination and the severity of the particular disorder to which the therapy is applied. Dosages for a given patient can be determined using conventional considerations, e.g., by customary comparison of the differential activities of the compositions described herein and of a known agent, such as by means of an appropriate conventional pharmacological protocol. The compositions can be given in a single dose schedule, or in a multiple dose schedule.

[0103] Suitable dosage ranges for a cancer agent may be of the order of several hundred micrograms of the agent with a range from about 0.001 to 10 mg / kg / day, preferably in the range from about 0.01 to 1 mg / kg / day.

[0104] The invention will now be exemplified with the following examples, which are not limitative.

[0105] FIGURE LEGENDS

[0106] Figure 1: Quantitative analysis workflow on HALO- AL Multi-immunofluorescence and their respective hematoxylin, eosin and saffron (HES) stained sections were imported into HALO-AI v3.6 or higher, region of analysis (ROA) was outlined on each immunofluorescence (IF) slide and several Al algorithms were built and optimized until validation (>85% accuracy) in order to detect and quantify tissue categories (Tumor, Stroma, TLS), as well as nuclei and cells, and cell types. Finally, quantitative data were generated and spatial analyses can be performed in order to study cell distribution, cell contacts and distances, cell clusters and cell infiltration.

[0107] Figure 2: Complete HighPlex IHC module settings used for COPOC (Proof of concept) TLS quant project to run the validated Nuclei seg algorithm in the “T zone” defined by the DenseNet Al algorithm.

[0108] Figure 3: Schematic representation of a computing device for implementing the method according to embodiments. Computing device 1 comprises at least one processor 10 and at least one memory 11.

[0109] Figure 4: Linear regression showing the correlation between TLS quant method (Immunohistochemistry (IHC) CD20 / CD21) and standardized immunofluorescence staining method.

[0110] EXAMPLES

[0111] Example 1: Multiplex immunofluorescence method

[0112] 1. Multiplex-Immunofluorescence

[0113] The tissue sections of Formalin-fixed paraffin-embedded (FFPE) of non-small cell lung cancer (NSCLC) (4 pm thick (cut for a maximum of 6 months)) are dried for a minimum of 30 min in a drying oven at 37°C.

[0114] The tissue sections are deparaffinized under a laboratory fume hood, and the slides are immersed in two successive baths of Clearene (Leica biosystem) (or 100% xylene) for 5 min each, then in one bath of absolute ethanol for 5 min, one bath of 90% ethanol for 5 min, one bath of 70% ethanol for 5 min, one bath of 50% ethanol for 5 min and two successive baths of distilled water for 5 min each (use glass containers for slides).

[0115] To retrieve the antigen, the slides are immersed in a bath of pre-warmed antigen retrieval solution Target retrieval solution (TRS) pH6 (10X, Dako; 1 volume + 9 volumes of distilled water) for 30 min at 97°C (in a water bath; use a plastic container for slides), then cool down on the bench for at least 30 min at room temperature. After completing the antigen retrieval steps, it is possible to leave the slides in a IX TBS bath at room temperature overnight before proceeding with the experiment. The slides are washed in TBS for 5 min under gentle agitation (use a glass container for slides and raise the slide support to place a magnet under the slides). The excess buffer is quickly removed with absorbent papers. When removing excess buffer using absorbent papers, it is important to take care not to let the slides dry for more than 2 minutes before adding the following reagent to prevent tissue alteration and the occurrence of staining background noise. The tissue is outlined with a PAP Pen (kisker) to allow buffers to strictly stay on the tissues. The slides are placed in a humidified chamber. The tissue is covered with 200 pL of 3% H2O2 for 30 min at room temperature (to inactivate endogenous peroxidases) and the slides are washed in lx TBS for 5 min. The excess buffer is removed with absorbent papers. The tissue is then covered with 200 pL of Protein Block (Dako) for 30 min at room temperature and the excess Protein Block is removed with absorbent papers. The protein Block contains casein, a hydrophilic protein. This formulation helps prevent diffuse background staining caused by hydrophobic or ionic interactions between the primary antibody, secondary reagents, and nontarget tissue components.

[0116] The tissue is covered with 200 pL of primary antibody solution (anti-human DC-LAMP (rat IgG2a, clone 1010E1.01, Eurobio Scientific), 6.25 pg / mL in antibody diluent (Dako REAL™, Dako)), for Ih at room temperature. The primary antibody solution is removed by gently spraying TBS-Tween above tissue slices and the slides are washed in TBS-Tween for 5 min, under gentle agitation. The excess buffer is removed with absorbent papers.

[0117] The tissue is then covered with 200 pL of secondary antibody solution (anti-rat IgG-HRP (ImmPRESS HRP Polymer Kit, Vector Laboratories), ready -to-use), for 30 min at room temperature and the secondary antibody solution is removed by gently spraying TBS-Tween (1 volume of lOx TBS + 9 volumes of distilled water + 0.04% Tween20) above tissue slices. The slides are washed in TBS-Tween for 5 min, under gentle agitation and the excess buffer is removed with absorbent papers.

[0118] The tissue is covered with 200pL of tertiary reagent (AF594-Tyramide (ThermoFisher Scientific), l / 125e in 0.03% H2O2 TBS (Tyramide conjugated to Fluorochrome is reconstituted before use in DMSO according to the manufacturer’s instructions), for 10 min at room temperature. The tertiary reagent is then removed by gently spraying TBS-Tween above tissue slices and the slides are washed in TBS-Tween for 5 min, under gentle agitation and the excess buffer is removed with absorbent papers. To retrieve the antigen, the slides are immersed in a bath of pre-warmed antigen retrieval solution (TRS, pH6) for 10 min at 97°C (in a water bath; use a plastic container for slides). This step allows the removal of the bound primary and secondary antibodies while saving the AF594-tyramide binding of DC-LAMP+cells.

[0119] The tissue section is subsequently labeled with anti-CD3 antibody. The slides are incubated with anti-human CD3 antibody (rabbit polyclonal, Dako) (8 pg / mL in antibody diluent), for Ih at room temperature, then with secondary antibody incubation (anti-rabbit IgG-HRP (ImmPRESS HRP Polymer Kit, Vector Laboratories), 1.6 pg / mL in TBS), for 30 min at room temperature; and with tertiary reagent (CF514-Tyramide (Interchim), l / 125e in 0.03% H2O2 TBS), for 10 min at room temperature with washing in TBS-Tween as described above between each step.

[0120] The slides are then incubated with anti-human pan-Cytokeratins antibody (mouse IgGl, clones AE1 / AE3, Dako) (3.33 pg / mL in antibody diluent), for Ih at room temperature; then with secondary antibody (anti-mouse IgGl-HRP (whole IgG, JIR), 1.6 pg / mL in TBS), for 30 min at room temperature; and with tertiary reagent (AF488-Tyramide (ThermoFisher Scientific), l / 125e in 0.03% H2O2 TBS), for 10 min at room temperature with washing in TBS-Tween as described above between each step.

[0121] The slides are then incubated with anti-human PNAd ((rat IgM, clone MECA-79, BD Pharmingen), anti-human CD21 (mouse IgGl, clone 1F8, Dako), and anti-human CD20 (mouse IgG2a, clone L26, Dako) antibodies (anti-human PNAd, 10 pg / mL + anti-human CD21, 6.6 pg / mL + anti-human CD20, 0.504 pg / mL in antibody diluent), for Ih at room temperature; then with secondary antibodies (anti-rat IgM-AF647, 15 pg / mL + anti-mouse IgGl-HRP, 1.6 pg / mL + anti-mouse IgG2a-BV480 (whole IgG, JIR), 4 pg / mL in TBS), for 30 min at room temperature; and with tertiary reagent (AF555-Tyramide (ThermoFisher Scientific), l / 125e in 0.03% H2O2 TBS), for 10 min at room temperature. The 1F8 antibody stains the FDC-specific CD21 long isoform and does not recognize the short CD21 isoform expressed by B cells

[0122] The tissue is then covered with 200pL of Nuclei staining solution (DAPI (ThermoFisher Scientific), 1 pg / mL in distilled water), for 5 min at room temperature.

[0123] The secondary antibody solution is removed by gently spraying TBS above tissue slices and the slides are washed in distilled water for 5 min, under gentle agitation. The excess buffer is removed with absorbent papers. The slides are finally mounted with glass coverslips using Fluorescent Mounting Medium and dried overnight at 4°C in the dark before storing. The optimal dilution of the primary, and secondary antibodies and tyramide reagents should be determined based on your experimental setup and following the instructions provided by the antibody manufacturer.

[0124] 2. Image Analysis and TLS Quantification

[0125] To obtain whole slide images, slides need to be scanned using a multispectral fluorescence scanner (PhenoImager / Vectra Polaris, Akoy a Bioscience) or otherwise a fluorescent microscope (Zeiss Axio Observer Z1 piloted by Zen software) with adapted filters to image at 385nm, 430nm, 475nm, 511nm, 555nm, 590nm, and 630nm. Depending on the tissues, significant autofluorescence from collagen fibers can be observed, particularly in the AF488 channel, and to a lesser extent in the CF514 channel. Additionally, these channels have closely related wavelengths, leading to overlapping signals. However, due to differences in morphology between T cells and tumor cells, the two cell types are easily distinguishable.

[0126] For a robust quantification at cell-level precision, scanning should be performed at a minimum of 20x magnification. It is important to favor the use of a fluorescent microscope or scanner with spectral unmixing capabilities to distinguish real signals from false positive signals induced by spectrally overlapping fluorophores. Thus, the use of a device like the PhenoImager (previous Vectra Polaris) solution (Akoya Bioscience) is advised. TLS quantification can be performed automatically, using image analysis software (HALO-AI® version 3.6 or higher). For the analysis using HALO-AI, several representative regions and samples need to be chosen to train and validate Al tissue segmentation, Al cell segmentation, as well as Al cell phenotyping algorithms.

[0127] HALO’s Al algorithms were validated only if, upon visual inspection, the matching accuracy between the segmented image and the native image was above 85%. Remaining untrained areas were used to validate tissue / cell segmentation and phenotyping algorithms by verifying the high matching accuracy between the segmented markup image and the corresponding fluorescent image in terms of structure identification, nuclei / cell detections, and phenotype attributions. In case of insufficient accuracy, the algorithm was further trained with additional examples, and the validation process was repeated.

[0128] The analysis workflow consists in:

[0129] 1. Annotate the Region of interest (ROI). 2. Perform the tissue segmentation to discriminate Tumor, Stroma, and TLS from Artifacts (Necrosis, red blood cells, dust, folding) and areas without tissue (“No Tissue”).

[0130] 3. Perform nuclei / cell segmentation to identify individually each round nucleus as well as large (pan-Cytokeratins) and elongated cells (i.e., CD21 and HEV positive cells).

[0131] 4. Perform phenotyping analyses to obtain quantitative information (counts, percentages, XY coordinates, mean intensities) about all cell populations detected: pan-Cytokeratins+tumor cells, CD3+T cells, CD20+B cells, DC-LAMP+mDC, CD21+FDC, and PNAd+HEV.

[0132] 5. Perform spatial analysis to identify proximity and relative spatial distribution of cells across the tissue section.

[0133] After validation, all algorithms and analysis modules can be launched in batch analysis by going to the “Study” list, selecting all images, and right-clicking on the “Analyze...” option. Furthermore, spatial analyses can be performed in batches as well on these data by rightclicking on “Results. . .”. (see Figure 1).

[0134] In more detail, image analysis and TLS quantification consist of the following steps:

[0135] / . Import the slide image generated after scanning into a HALO-AI study

[0136] 2. Annotation using HALO-AI: Outline the whole tissue of interest, corresponding to the region of analysis (ROA), also named region of interest on each image using the annotation tool in the toolbar. ROA is focused on the tissue of interest (tumor, stroma, tumor-associated TLS) but can also include some structures of non-interest such as holes without tissue, red blood cells, dust, folding, or necrosis present in the tissue. Afterward, those regions of non-interest will be excluded from the final analysis using the Al Tissue segmentation algorithm through the “No Tissue” and “Artifacts” classes. Therefore, only the “Tumor”, “Stroma”, and “TLS” regions will be analyzed.

[0137] 3. Classifier for tissue segmentation:

[0138] 1. Load a MiniNet classifier algorithm, on HALO-AI, to identify and segment tissues of a study containing less than 100 samples. MiniNet and DenseNet V2 algorithms both can be used to identify and segment tissues by adding representative examples of each tissue category on representative images of a project. MiniNet can be set up quickly with few training examples and has good robustness. However, it is difficult to handle training for very large cohorts (>100). DenseNet V2 algorithm optimization is slower to train than MiniNet and requires lots of training regions (around 1000) however it can create a highly robust classifier that can handle multiple staining and heterogeneity. It is recommended to start with MiniNet and change to DenseNet v2 if the performance of the classifier stops improving. Note that it is possible to create a robust DenseNet v2 algorithm from the combination of training regions created by several MiniNet of similar projects (same tissues, same panel). Add in the MiniNet all relevant tissue classes in priority order from less to more important. This order will be followed during the superposition of training annotations. For NSCLC tumors, create 5 classes and add tissue categories in priority order from class 1 to 5 like the following: “No tissue” (glass), “Stroma”, “Artifacts” (red blood cells, necrosis, folding, dust, ...), “TLS”, and “Tumor”. Set, if needed, MiniNet resolution and minimum object size depending on the precision necessary for the recognition of all structures present in the samples. Keep the default resolution (2 pm / pixel) and the default minimum object size value (200 pm2) if they match the tissue structures of your sample. In “Advanced options”, then in the “Fluorescence” tab, tick on the relevant fluorescent channels needed to distinguish all tissue types as well as glass, artifacts, and TLS. For NSCLC, all channels were selected for the MiniNet algorithm training to segment correctly all tissue types and distinguish organized TLS from dispersed immune infiltrate. Tick all channels for the algorithm training only if they are all relevant for tissue segmentation (caution: processing and training time will increase). For example, “DAPI” and “pan-Cytokeratins” can usually help to distinguish “Tumor” from other tissue categories. Moreover, a channel with an auto-fluorescence background can help to discriminate between loose stroma containing few dispersed cells and the glass. Add training examples on different representative areas and images by drawing a few regions specific to each tissue class. For that, use the brush or the pen tools and draw a few small to medium training regions in a way to covers all the intra and inter-sample heterogeneity. Check that the priority order of tissue classes is respected during the drawing of superposed training regions. For NSCLC slides, outline very precisely TLS to allow the algorithm to detect them based on their cell compositions, cell densities, and cell distributions. Start the MiniNet training via “Start Training”. 7. Wait for several iterations and for the cross-entropy to decrease and tend towards zero or to stabilize.

[0139] 8. Open the “Real-Time Tuning” window to live view the training results and verify the accuracy of tissue segmentation.

[0140] 9. Correct the algorithm until its validation (>85% accuracy) by drawing new training examples according to the results shown by the “Real-Time Tuning” window (follow step 5). In the “Results” tab, tissue segmentation analysis generated a “Summary Data” table with all classified areas (mm2) as well as an “Object Data” table including Classifier Label, Region perimeter (pm), Region Area (pm2) Id of each tissue regions, as wells as their XY coordinates (Xmin, Xmax, Ymin, Ymax).

[0141] 10. Save the validated algorithm.

[0142] 11. Launch the validated MiniNet algorithm in the ROA (see step 1) via the “Analyze. . .” option.

[0143] 12. If needed, annotations can be generated from tissue segmentation using “Classifier Actions”, “Advanced Options...”, and “Annotations”, on “Mask to Annotations:” and tick on “Tumor”, “Stroma” and “TLS”. Save the algorithm into a new name and launch the algorithm to generate the corresponding annotations (caution: this algorithm will create annotations every time it is launched).

[0144] 13. Results of tissue segmentation are displayed in the “Results” tab.

[0145] 4. Nuclei Seg (Halo Al) Network

[0146] 1. Open a new classifier algorithm and choose Nuclei Seg to segment nuclei exclusively based on DAPI nuclear staining. On fluorescent images, by default, the training of the Nuclei seg algorithm is exclusively based on the “DAPI” channel. Thus, the algorithm requires a channel named “DAPI”. If no DAPI channel is available, one channel needs to be renamed “DAPI”, for that right click on the channel name, choose “Change Channel Name” and write “DAPI”. If needed, it is also possible to use the Nuclei seg algorithm to segment cells instead of nuclei by adding, into training channels, DAPI, and one or several other channels required for the identification of the cells of interest. Cell segmentation is especially well adapted for elongated cells such as endothelial cells or FDC and cells with large morphologies such as tumor cells. 2. Keep HALO-AI default settings for the resolution (0.25 pm / px) and Minimum Object Size (0 pm).

[0147] 3. Open the “Real-Time Tuning” window and check the accuracy of the nuclei segmentation performed by the default Nuclei seg algorithm.

[0148] 4. If the default Nuclei seg is not accurate enough, create a custom Nuclei seg algorithm by adding relevant examples in the “Background” and “Nuclei” categories. For the “Nuclei” class, draw precisely a high number of nuclei (>2000-3000) with various shapes, intensities, and distribution on several representative images to cover all intra and inter-sample heterogeneity to train Al algorithms to detect and segment correctly nuclei. For best results, fully annotate the entire region instead of random areas across the tissue. For example, it is better to annotate completely several areas of approximately 0.5 or 1 mm2instead of annotating individual cells dispersed across the whole slide.

[0149] 5. If DAPI-based Nuclear segmentation does not match enough some cell types with specific morphologies, add the corresponding training channels to segment those cells. If needed, it is also possible to use the Nuclei seg algorithm to segment cells instead of nuclei by adding, into training channels, DAPI, and one or several other channels required for the identification of the cells of interest. Cell segmentation is especially well adapted for elongated cells such as endothelial cells or FDC and cells with large morphologies such as tumor cells. For NSCLC slides, dual Nuclei / Cell segmentation was performed. In addition to DAPI, three more channels were used for training to segment elongated and large cells (CD21+cells, HEV+cells, pan-Cytokeratins+cells) as well as other round nuclei. To add those training channels, click on “Classifier Actions”, “Advanced options”, then “Fluorescence” and tick on DAPI, pan- Cytokeratins, CD21, and HEV channels.

[0150] 6. Start the training of the Nuclei seg custom algorithm.

[0151] 7. Wait for several iterations and for the cross-entropy to decrease and tend towards zero or to stabilize.

[0152] 8. Verify the nuclei segmentation accuracy using the “Real-Time Tuning” window.

[0153] 9. Correct the algorithm until its validation (>85% accuracy) by drawing new training examples according to the results shown by the “Real-Time Tuning”. The remaining untrained areas were used to validate the nuclei / cell segmentation by verifying the matching accuracy of cell detection and cell segmentation aggressiveness between the segmented markup image and the corresponding fluorescent image (same method for tissue segmentation and phenotyping).

[0154] 10. Save the validated custom Nuclei seg classifier.

[0155] 11. If needed, launch the validated algorithm in the ROA via the “Analyze. ..” option.

[0156] 12. Results of dual Nuclei / Cell segmentation are displayed in the “Results” tab similarly to tissue segmentation (“Summary Data” table with area (mm2) and total nuclei count, “Object data” table).

[0157] 5. Phenotype Analysis by HALO-AI

[0158] Phenotype analysis can be performed in different ways on HALO-AI depending on the sample heterogeneity, the need to obtain mean intensities, and the need to combine data from Al tissue segmentation, Al cell segmentation, and Al phenotyping inside a single data table.

[0159] Here, three methods of phenotyping are presented. Regardless of the phenotyping method used (Object Phenotyper and / or HighPlex FL / FISH IF modules), high accuracy of nuclei / cell segmentation is always required for robust and accurate cell phenotyping. Phenotyping requires an accurate nuclear / cell classifier to detect correctly the cells depending on their morphologies (size, shape) and fluorescence pattern (intensity, pixel distribution). The first one uses the Al Object Phenotyper algorithm alone. The second one corresponds to the use of threshold-based supervised or unsupervised phenotyping method with HALO’s HighPlex FL module alone (see below Step b). The third method refers to the implementation of all Al algorithms, especially the Object Phenotyper algorithm, inside a single HighPlex FL module (see below Step c).

[0160] Object Phenotyper can perform robust and accurate phenotyping through a training-by-example process but does not give mean intensity information. Object Phenotyper classifier identifies cell types according to phenotype classes and training by example processes defined by the user.

[0161] In HALO-AI version 3.6, Al algorithms cannot be launched directly inside a tissue compartment (i.e., “Tumor” or “Stroma”). Corresponding annotations need to be created first via the “Mask to Annotation” option. However, it is important to highlight the fact that all positive annotations inside negative annotations are not considered as positives and thus the area can be underestimated when having concentric positive / negative layers. For example, in the case of “Stroma” areas containing successively positive, then negative (due to “Tumor” islet presence), then positive (“Stroma” inside a “Tumor” islet) concentric annotations, only cells in the first positive layer will be considered to be in the “Stroma” and will be phenotyped with Object Phenotyper. To avoid this problem of potential tissue area underestimation, the inventors advise using HALO-AI 4.0 or higher to directly run the Object Phenotyper (or Nuclei Seg) inside the categories of the Tissue segmentation algorithm. For HALO-AI version 3.6, the inventors advise using Object Phenotyper Al algorithm inside the tissue classes of interest using the HighPlex FL module.

[0162] On the other hand, HighPlex FL (or FISH-IF) module phenotypes cells based on parameters and thresholds inside cell compartments (nucleus, cytoplasm, membrane). This module can be used individually to phenotype cells or can combine all Al algorithms to obtain a single analysis.

[0163] User-trained Al algorithms, such as MiniNet, Nuclei Seg, and Object Phenotyper can be included in the HighPlex FL (or FISH-IF) module. This analysis module can perform a global single analysis combining all Al algorithms as well as threshold-based supervised phenotyping. The final data contains Summary Data and Object Data and provides XY coordinates, tissue compartments, nuclei area, cell ID, mean intensity in each cell compartment (nuclei / cytoplasm / membrane / entire cell), phenotype defined by Object Phenotyper, phenotypes defined by HighPlex-FL or FISH-IF modules).

[0164] Using either Al Object Phenotyper alone (see section a) or the HighPlex FL (or FISH-IF) module in combination with Al algorithms (tissue, cell, and phenotype segmentations) (see below Step c) is recommended to generate more accurate data. a. 1stPhenotyping method using Object Phenotyper Al algorithm alone

[0165] 1. Open a new classifier algorithm and choose the Object Phenotyper algorithm to phenotype segmented nuclei / cells.

[0166] 2. Click on “Load Segmentation” and choose the validated Nuclei seg algorithm.

[0167] 3. Click on “Start segmentation” to visualize nuclei / cell outlines.

[0168] 4. Set the resolution to 0,25pm / pixel to match Nuclei Seg algorithm precision. By default, Object Phenotyper resolution is at 2 pm / pixel to phenotype broadly cell types such as “tumor cells”, “immune cells” and “stromal cells”. In the present protocol, this classifier is used to detect precisely all cell types based on their marker (“CD21+”, “CD3+”, “pan- Cytokeratins+”, “CD20+”, ...), therefore resolution was set to 0.25 pm / pixel to match Nuclei Seg precision and to increase detection precision during the phenotyping process. 5. Add classes corresponding to negative cells and each positive cell type of each exclusive marker of the panel. To phenotype the cells, the Object Phenotyper algorithm must contain classes of cells that correspond only to negative cells and exclusive markers. Each co-expressing cell must be phenotype in a separate Object Phenotyper algorithm. To cover sample heterogeneity, examples of negative cells as well as positive cells of each phenotype were taken from different representative slides to fill the algorithm training. In the present example, 7 classes were created: “Other” (=Negative cells) in blue, “CD3+” in cyan, “CD20+” in red, “PanCk+” in white, “DC-LAMP+” in green, “CD21+” in yellow, and “HEV+” in magenta (see Figure 1). For co-expressed markers, another Object Phenotyper algorithm with the same resolution and the same Nuclei seg algorithm as the previous Object Phenotyper was created. Then the negative cells and the co-expressed markers of interest were added. For a panel with co-expressed markers, several Object Phenotyper algorithms, including only exclusive markers, must be created and the analysis must be performed separately for each co-expression. All Object Phenotyper algorithms must have the same resolution and require the same Nuclei seg algorithm. Those requirements are needed to keep the same precision, same cell coordinates, and same cell ID. Data can be merged afterward into a single data table on R using cell ID or XY coordinates to obtain the complete phenotype combinations for each cell.

[0169] 6. In “Classifier Actions”, “Advanced options” and in the “Fluorescence” tab, tick on the relevant fluorescent channels needed to distinguish all phenotype classes listed in Object Phenotyper. Object Phenotyper default settings use all fluorescent channels during training however only relevant channels corresponding to phenotype classes should be kept to decrease the processing time and improve detection accuracy. Tick all channels for the training of the algorithm only if all channels are relevant for the phenotyping (caution: processing and training time will increase). Due to differences in the working image zoom parameters, cell counts are given through Nuclei seg (or Object Phenotyper) and HighPlex FL (or FISH-IF) could not exactly match inducing small cell segmentation differences and therefore a bias of 1 to 5% depending on cell aggregations. As a result, the inventors advise testing, in the HighPlex FL module, several image zoom and choosing the image zoom value that matches the best with validated Nuclei seg and Object Phenotyper data performed inside the entire image or ROA. The inventors advise using the 1st method for Phenotyping with HALO-AI v4.0 (or higher) and if necessary, performing in addition the 2ndphenotyping method to obtain a in parallel marker mean intensity value for each cell. Or to directly use only the 3rdPhenotyping method to obtain in a single analysis all informative data (including cell counts, percentages, coordinates, Object Phenotyper-based cell phenotypes, HighPlex FL threshold-based cell phenotypes). In case of a huge number of markers and co-expressing markers or to perform broad phenotyping, it is possible to use the 2ndmethod directly and to determine phenotype in an unsupervised fashion.

[0170] For NSCLC analysis, all markers were analyzed with a single Object Phenotyper algorithm thus all fluorescent channels were used for the training. Add, using the brush tool, a high number (>100) of training examples of individual cells and cells in cell aggregates for each phenotype class. Those examples must cover all intra and inter-sample heterogeneity in terms of intensity, cell and tissue distribution, and morphology, and thus must be from different representative areas and images. Start the Obj ect Phenotyper training. Wait for several iterations and for the cross-entropy to decrease and tend towards zero or to stabilize. Open the “Real-Time Tuning” window to check the results of the training and to verify the accuracy of the phenotyping. Correct the algorithm until its validation (> 85% accuracy) by drawing new training examples according to the results shown by the “Real-Time Tuning” window (see section a step 7). Save all validated Object Phenotyper algorithms. Run the Object Phenotyping Al algorithms inside each tissue compartment of interest (“TLS”, “Tumor”, and “Stroma”) or in the entire ROA. Object Phenotyper classifier identifies cell types according to phenotype classes and training by example processes defined by the user. In HALO-AI version 3.6, Al algorithms cannot be launched directly inside a tissue compartment (i.e., “Tumor” or “Stroma”). Corresponding annotations need to be created first via the “Mask to Annotation” option. However, it is important to highlight the fact that, all positive annotations inside negative annotations are not considered as positives and thus the area can be underestimated when having concentric positive / negative layers. For example, in the case of “Stroma” areas containing successively positive, then negative (due to “Tumor” islet presence), then positive (“Stroma” inside a “Tumor” islet) concentric annotations, only cells in the first positive layer will be considered to be in the “Stroma” and will be phenotyped with Object Phenotyper. To avoid this problem of potential tissue area underestimation, the inventors advise using HALO-AI 4.0 or higher to directly run the Object Phenotyper (or Nuclei Seg) inside the categories of the Tissue segmentation algorithm. For HALO-AI version 3.6, the inventors advise using Object Phenotyper Al algorithm inside the tissue classes of interest using the HighPlex FL module.

[0171] For each analysis, a Summary Data table (counts, percentages per phenotypes) and a single-cell Object Data table (cell ID, XY coordinates, phenotype class, confidence frequency of phenotype attribution) are generated. b. 2ndPhenotyping method using HighPlex FL module alone

[0172] 1. Go to the “Analysis” tab and load a new HighPlex FL (or FISH-IF) module to perform supervised or unsupervised phenotyping.

[0173] 2. Click on “Settings Actions” and “Save ...” to save the analysis module into a new name.

[0174] 3. Set image zoom to match validated Nuclei seg segmentation precision. For 0,65pm / pixel resolution NSCLC images (“Image Details”), HALO-AI default lx “Image Zoom” matched the best the Nuclei seg precision. In “Nuclei Detection”, if necessary, “Default Al” Nuclei seg can be optimized by setting “nuclear contrast threshold”, “minimum nuclear intensity”, “nuclear segmentation aggressiveness” and the “minimum / maximum nuclear size” (comprised between 10 and 200 pm).

[0175] 4. In the “Dye Selection” section, click on “Autofill” to automatically fill dyes into the analysis module.

[0176] 5. Fill the parameters in each section (Dye Selection, Nuclear Detection, Membrane, Cytoplasm Detection, Dyes, Advanced) to detect nuclei, cytoplasm, and / or membranes based on the dye’s positivity threshold, segmentation aggressiveness, morphology, and size of cell compartments.

[0177] 6. For this 2ndphenotyping method, do not fill “HALO Al Phenotyper” with any Object Phenotyper algorithm. Only load an Object Phenotyper algorithm in “HALO Al Phenotyper” for the 3rdmethod (see Step c) to use Object Phenotyper inside the HighPlex FL (or FISH-IF) module. 7. Go to the “Advanced” section, choose “True” in “Store Object (cell) Data” and “Mask” in “Classifier Output Type” to obtain XY coordinates, ID, and data for each cell. Those are needed for spatial and proximity analysis. If needed, instead of “Mask”, “Regions as Objects” can be chosen in order to obtain the number of phenotyped cells per tissue regions (“Tumor”, “Stroma”, “TLS”).

[0178] 8. In “Advanced”, choose the validated tissue segmentation algorithm in “Classifier” and tick on the relevant tissue classes in “Class List”. For NSCLC analysis, a validated MiniNet algorithm was chosen, and “Stroma”, “Tumor” and “TLS” tissue classes were selected.

[0179] 9. In “Nuclear Detection”, choose the type of Al nuclear segmentation (Al Default or Al Custom) and import the corresponding Nuclei seg classifier. User-trained Al algorithms, such as MiniNet, Nuclei Seg, and Object Phenotyper can be included in the HighPlex FL (or FISH-IF) module. This analysis module can perform a global single analysis combining all Al algorithms as well as threshold-based supervised phenotyping. The final data contains Summary Data and Object Data and provides XY coordinates, tissue compartments, nuclei area, cell ID, mean intensity in each cell compartment (nuclei / cytoplasm / membrane / entire cell), phenotype defined by Object Phenotyper, phenotypes defined by HighPlex-FL or FISH-IF modules).

[0180] In NSCLC, for example, “Custom Al” was chosen and the validated Nuclei seg algorithm was loaded. To obtain a similar segmentation, all additional filters were removed by setting “Minimum Nuclear Intensity”, “Nuclear Segmentation Aggressiveness” and “Minimum Nuclear Roundness” to 0. Then “Maximum Image Brightness”, “Number of Nuclear Dyes” and “Nuclear Dye 1 Weight” (DAPI) to 1. Finally, by setting “Nuclear Size” between 0 and 10000.

[0181] 10. Identify the membrane and the cytoplasm compartments of the cells (Maximum Cytoplasm Radius around 2 pm). For NSCLC, in the “Membrane and Cytoplasm Detection” section, “Maximum Cytoplasm Radius” was set to 2, then “Membrane Segmentation Aggressiveness” and “Number of Membrane Dyes” to 0, and finally “Cell Size” between 0 and 30000.

[0182] 11. Identify the positivity threshold for each dye in its corresponding cellular compartment by setting the positivity threshold and percentage of completeness for each dye. 12. To perform supervised phenotyping, enter the number of phenotype combinations, fill each phenotype section with its name, and define phenotypes using Boolean filters (AND Positive / Negative, OR Positive / Negative). Note that there is no need to fill any phenotype to perform unsupervised analysis using exclusively mean intensity values.

[0183] 13. View the colocalized Markup Image by opening the “Real-Time Tuning” window in “Analyze”.

[0184] 14. Verify visually the matching accuracy.

[0185] 15. Correct the parameters if the accuracy is below 85%.

[0186] 16. Save the validated HighPlex FL analysis module.

[0187] 17. Run the analysis in the ROA to obtain quantitative analysis in TLS, Tumor, and Stroma regions.

[0188] 18. The HighPlex FL analysis gives quantitative data (count, percent, mean intensities, Stroma / Tumor / TLS compartments) for each marker of interest including Summary Data table (counts, percentages per phenotypes) and single-cell Object Data table (cell ID, XY coordinates, cell type defined by threshold-based phenotyping, confidence frequency of phenotype attribution) are generated. c. 3rdPhenotyping method combining Object Phenotyper Al algorithm inside HighPlex FL module

[0189] Object Phenotyper can be loaded inside the HighPlex FL analysis module to obtain both robust Al phenotyping and threshold-based phenotyping. HighPlex FL can process together Al tissue and cell segmentations as well as Al Object Phenotyper. It generates a complete Summary Data table as well as Single Cell Object data. For batch analysis, select images in the “Study”, rightclick on the selected images, and choose “Results...”, then select all analyzed images of one analysis and click on “Result Actions” to draw a plot or perform Spatial Analysis.

[0190] 1. Open a new HighPlex FL module.

[0191] 2. Load the validated Al Object Phenotyper algorithm in “HALO Al Phenotyper” and select all Al Phenotypes except “N / A”.

[0192] 3. View the Object Phenotyper Markup by choosing the validated Object Phenotyper algorithm in the “output image” and by opening the “Real-Time Tuning” window. 4. Add a single phenotype in “Number of Phenotype” (“Dye Selection” section) as DAPI positive.

[0193] 5. Set HighPlex FL and perform analysis as detailed in section b Steps 2 to 17.

[0194] 6. The HighPlex FL analysis gives quantitative data (count, percent, mean intensities, Stroma / Tumor / TLS compartments) for each marker of interest including Summary Data table (counts, percentages per phenotypes) and single-cell Object Data table (cell ID, XY coordinates, cell type based on thresholding (HighPlex FL settings), phenotypes based on Object Phenotyper (Al Algorithm), confidence frequency of phenotype attribution).

[0195] 6. Spatial analysis

[0196] The Object Data generated on HALO gives XY coordinates as well as ID and data for each cell needed for subsequent spatial analysis using the Spatial Analysis module.

[0197] 1 . Create a spatial plot by clicking on “Object Data” from the analysis results or “Result Actions” in the Study for batch analysis.

[0198] 2. Four spatial analyses are possible: a. Nearest neighbor analysis that calculates the distance (pm) from one phenotype cell type to its nearest neighboring cell of another phenotype. For example, it can measure the distance between the nearest CD3+T cells of each mDC. b. Proximity analysis that calculates for a cell phenotype of interest, the number and percentage of cell types present within its specified radius. For example, it can count the number of CD3+T cells within 10pm close to PNAd+HEV and their precise number every micron from 1 to 10 pm. c. Infiltration analysis that allows visualization of the T cell infiltration inside Tumor nest annotations. d. Density heatmaps that visualize the density of a specific cell phenotype in an area of interest, for example, the density B cells on the whole slide, and compare it inside and outside of TLS.

[0199] 7. Data export

[0200] After analysis, HALO-AI generates Single Cell Object Data as well as Summary Data. Summary data includes: a. For tissue segmentation (MiniNet): surface values (mm2) and surface percentages of each tissue category. b. For Nuclei / Cell segmentation (Nuclei Seg): Total Nuclei counts, Classified Area (mm2), Background Area (mm2), Nuclei Area (mm2). c. For Object Phenotyper: Counts and percentages of Total Cells and Total of each cell type. d. For HighPlex FL including Al algorithms: same as tissue category, counts and percentages of cell types detected by both phenotyping methods (Object Phenotyper and HighPlex FL thresholding), cells surface, cell mean intensity value of all pixels per fluorescent channel and per cell compartment. e. For nearest neighbor analysis: for each cell phenotype (reference cell) the mean distance (pm) to the nearest neighbor of each other cell phenotype in each tissue category. f. For proximity (for example 20 pm radius): for each phenotype and in each tissue category, the number and percentage of cells of the different phenotypes present within a specified radius of 20 pm and their number every micron from 1 to 20 pm.

[0201] All analyses generated data tables that can be exported into .csv or .fsc in several folders for each analysis (Tissue segmentation, Nuclei Phenotyper, HighPlex FL, nearest neighbor, proximity analysis). Plots as well as analysis markup images can be exported as .jpeg if needed.

[0202] Example 2: Immunohistochemistry (IHC) method

[0203] 1. IHC method

[0204] Formalin-fixed, paraffin-embedded (FFPE) human NSCLC or CRC tumor tissue, freshly obtained 4 pm thick sections (maximum 6 months) stained by immunohistochemistry for CD20 / CD21 or CD3 / DC-LAMP, were scanned at 20x or 40x magnification with a slide scanner (NanoZoomer, Hamamatsu) allowing obtention of detailed whole slide digital images with precise visualization of tissues and cells.

[0205] 2. Image Analysis and TLS Quantification

[0206] To build algorithms for TLS detection, the steps consisted in, briefly: - choosing a Classifier Neuronal Network (DenseNet v2 for tissue segmentation, Nuclei Seg for cell segmentation)

[0207] - filling associated parameters (resolution, minimum object size)

[0208] - adding classes to define the object of interest (Tissue types, cell type of interest)

[0209] - adding training examples and launching the training of the algorithm until stabilization of the cross-entropy value.

[0210] - optimization of the algorithm until validation with high detection accuracy (>85% upon visual slide inspection)

[0211] For that purpose, several representative regions and samples needed to be chosen to train and validate Al tissue segmentation and Al cell segmentation algorithms.

[0212] To assess the representativity nature of a slide, several factors were evaluated including tissue architecture, tissue and cell types, morphologies, sizes, shapes, distribution and organization, as well as arrangements of cells (sparse, in clusters). Slides were representative if the occurrence frequency of those evaluated factors was above or equal to 10%.

[0213] Contrary to cell segmentation, tissue segmentation required a higher number of slides, distributed over all NSCLC cohorts, for the training of DenseNet algorithms. This was imperative in order to cover all intra and inter-cohort heterogeneities of structure, organization and staining intensity.

[0214] For both staining, between 1 to 15 representative slides per cohort were chosen for the training and optimization processes and between 10 to 30 untrained slides were chosen to validate the Al DenseNet algorithms.

[0215] For very small cohorts (< 10 samples), all slides were integrated into the training set to cover cohort heterogeneities. On those cohorts, representative areas were chosen to train the algorithm and the remaining untrained areas were chosen to validate Al algorithms.

[0216] For custom Nuclei seg algorithm training, 16 samples distributed over 3 NSCLC cohorts were sufficient to validate the final algorithm by precisely manually outlining relevant and representative DCs and background areas.

[0217] Each HALO’s Al algorithm was validated only if, upon visual inspection, the matching accuracy between the segmented image and the native image was above or equal to 85%. In case of insufficient accuracy, the algorithm was further trained with additional examples, and the validation process was repeated until sufficient improvement.

[0218] The analysis workflow consists in:

[0219] 1. Annotating the Region of interest (ROI) based for example on hematoxylin-eosin-saffron (HES) stain or hematoxylin phloxine saffron (HPS) stain, or CK (cytokeratins) IHC stained slides.

[0220] 2. Building, for both staining, Al DenseNet algorithms for robust tissue segmentation to distinguish between the areas of interest including “Tissue”, “Anthracosis and Macrophages”, “T zones” (for CD3 / DC-LAMP staining) or “B Follicles” and “CD21+GCs” (for CD20 / CD21 staining) from “Artifacts” (mucus, red blood cells, dust, folding) and from “No Tissue” (areas without tissue).

[0221] HALO-AI DenseNet Neuronal Network can accurately identify and segment tissues while handling high inter and intra sample heterogeneity. For the tissue segmentation of both staining, 1 to 15 representative slides of all NSCLC cohorts were included into the training set.

[0222] 3. Building, for CD3 / DC-LAMP stained slides, a custom Nuclei Seg Al algorithm for robust and precise cell segmentation and specific identification of DC-LAMP+DC inside CD3 enriched T zones.

[0223] 4. Using HighPlex IHC module, for CD3 / DC-LAMP stained slides, to perform the cell segmentation of DC-LAMP+cells specifically (step 3) in “T zone” areas detected by the validated tissue segmentation algorithm (step 2). In Halo-AI v4.0 or higher, using HighPlex IHC is not necessary as classifier pipelines can be created to directly launch the validated Nuclei seg algorithm inside the DenseNet tissue class of interest. By using the DC Nuclei seg algorithm and T zones DenseNet algorithm, Halo-AI v4.0 Classifier pipeline option allows to directly run the analysis of DC detected by Nuclei seg algorithm into the T zone class defined by the DenseNet algorithm.

[0224] 5. After validation of all algorithms and analysis module, analysis was launched in batch on NSCLC cohorts and Colorectal cancer (CRC) tumors and summary data as well as single cell data tables were generated.

[0225] Data can be exported in csv format and if needed, spatial analyses can be performed using a single object data table.

[0226] Table 1: Training Annotation Number of CD3, DC-LAMP and B Follicles and GC for the DenseNet algorithm

[0227] 2.1. Import of images on HALO-AI

[0228] IHC slide images generated after scanning were imported into HALO-AI study (Halo- AI®v3.6.4134).

[0229] 2.2. Annotations of ROI

[0230] For the analysis, region of interest (ROI), validated by pathologists, were manually outlined on slides to focus the analysis on the relevant region for the analysis of TLS associated to the tumor. ROI included the tumor region and its surrounding peritumoral area. Massive non relevant areas (blur, folding, healthy tissue outside tumor mass, other problematic artifacts due to technical issues...) were excluded from the analysis to prevent problems of non-specific detection or segmentation. HALO’s pen, brush, magnet and flood annotation tools were used to manually draw and create positive or negative annotations. The annotation of the Region of interest (ROI) can be based for example on HES or HPS stained slides if necessary.

[0231] 2. 3. Al algorithm for tissue segmentation o f CD20 / CD21 stained IHC slides

[0232] 1. DenseNet Al algorithm was loaded into HALO-AI Classifier tab. The aim was to build a robust and broad Al algorithm for the detection of TLS’s B follicles (B-cell zone) and CD21+follicular dendritic cell (FDC) (germinal center (GC)) on NSCLC cohorts stained by IHC for CD20 / CD21. For that, five NSCLC cohorts were chosen for the training, optimization and validation steps. The algorithm was also successfully validated in the CRC cohort of the study.

[0233] 2. Six tissue classes were added into the algorithm in priority order from less to more important like the following: “Tissue”, “B Follicles”, “CD21+”, “No Tissue” (glass), “Artifacts” (red blood cells, necrosis, folding, dust, ...), “Anthracosis and Macrophages”. This priority order was followed in case of superposition of training annotations. 3. The resolution was set to 4 pm / px and minimum object size at 200 pm2. Those parameters gave the precision necessary for the accurate recognition of all structures present in the samples and especially B follicles and GC of TLS.

[0234] 4. For the training of B follicles-GCs DenseNet algorithm, 40 samples distributed over all 5 NSCLC cohorts were used to generate the final algorithm by manually outlining a high number of accurate, precise and representative training annotations on different areas. To that aim, zoom magnification was set around 60x by using HALO’s “zoom beyond image resolution” keyboard shortcuts. Then, a few regions specific to each tissue class were precisely drawn in a way to cover intra and inter-sample heterogeneities. In total, 392 “B Follicles” and 279 “CD21+GCs” annotations were manually drawn to train the algorithm to detect those structures based on the density and organization of CD20+B cell and CD21+FDC.

[0235] Anthracose and Total of training

[0236] Cohorts Artifacts CD21 GC B Follicles No Tissue Tissue

[0237] Macrophages annotations

[0238] 1 343 277 115 157 92 253 1237

[0239] 2 82 18 75 71 27 111 384

[0240] 3 6 12 32 67 6 21 144

[0241] 4 15 31 6 52

[0242] 5 82 114 42 66 588 129 1021

[0243] Total 513 421 279 392 713 520 2838

[0244] 5 cohorts 40 slides

[0245] Table 2: Training Annotation Number of B Follicles and CD21+cells for the 1stDenseNet algorithm

[0246] 5. DenseNet training was performed by clicking on “Start Training” and by waiting for several iterations and for the cross-entropy to decrease and tend towards zero or to stabilize.

[0247] 6. “Real-Time Tuning” window was used to live view the training results and verify the accuracy of the segmentation.

[0248] 7. The algorithm was corrected until its validation (>85% accuracy) by drawing new training examples according to the results shown by the “Real-Time Tuning” window. 8. The validated algorithm was saved as “SGB COPOC IHC DenseNet TLS B.Foll- GCs FINAL” and launched in the ROI via the “Analyze. ..” option.

[0249] 9. Results of tissue segmentation were displayed in the “Results” tab

[0250] 2.4. Al algorithm for tissue segmentation o f CD3 / DC-LAMP stained IHC slides

[0251] 1. DenseNet Al algorithm was loaded into HALO-AI Classifier tab in order to detect areas of DC-LAMP associated to the TLS-T zones on NSCLC cohorts stained by IHC for CD3 / DC- LAMP. Four NSCLC cohorts as well as one cervical cancer cohort were chosen for the training, optimization and validation steps.

[0252] 2. Five tissue classes were added into the DenseNet algorithm following a priority order from less to more important: “Tissue”, “T zone”, “No Tissue” (glass), “Artifacts” (red blood cells, necrosis, folding, dust, ...), “Anthracose and Macrophages”. This priority order was followed in case of superposition of training annotations.

[0253] 3. The resolution was set to 2 pm / px and minimum object size at 200 pm2in order have enough precision to distinguish between positive areas containing DC in T zones and negative areas containing either DC-LAMP+pneumocytes or Anthracose / Macrophages in or around T zones. This allowed to focus the detection of “T zone“ specifically on T zone associated DC-LAMP+expressing DC while excluding DC-LAMP+pneumocytes and Anthracose / Macrophages.

[0254] 4. For the algorithm training, a high number of accurate, precise and representative training annotations were outlined on different representative areas and images. 41 samples distributed over 4 NSCLC and 1 CRC cohorts were used to generate the final algorithm. Several regions specific for each tissue class were manually drawn in a way to cover all the intra and intersample heterogeneity and to allow detection of T zone associated DC-LAMP+expressing DCs. For T zone manual training annotations, magnification was set between 20x and 40x depending on the size of TLS-T zones. For those structures, each annotation contained both high density of CD3+T cells compared to their surrounding tissue and presence of DC-LAMP+DC.

[0255] Manual outlines were performed by drawing around the T zone DC in a way to keep, when possible, an outside margin between 20 and 40 pm. The objective was to include into the annotation T zone CD3+T cells and DCs in a way to train the algorithm to take into account both T zone density and intensity for the detection of DC-LAMP+DCs. In total, 671 positive “T zone” annotations were precisely manually outlined to train the algorithm to detect T zones based on their organization and CD3+T cell intensity and density as well as DC-LAMP+DC distribution, intensity and morphology.

[0256] . ,, . Total of

[0257] Anthracose and Artifacts No Tissue T zone Tissue training

[0258] . . Macrophages . ..

[0259] CohortsK® annotations

[0260] 1 636 187 139 285 686 1933

[0261] 2 844 69 122 209 429 1673

[0262] 3 416 31 25 133 339 944

[0263] 4 613 11 8 36 75 743

[0264] 6 1 8 15 24

[0265] Total 2509 299 294 671 1544 5317

[0266] 5 cohorts 41 slides

[0267] Table 3: Training Annotation Number of T zones for the 2ndDenseNet algorithm

[0268] 5. DenseNet training was performed by clicking on “Start Training” and by waiting for several iterations and for the cross-entropy to decrease and tend towards zero or to stabilize.

[0269] 6. “Real-Time Tuning” window was used to live view the training results and verify the accuracy of the segmentation.

[0270] 7. The algorithm was corrected until its validation (>85% accuracy) by drawing new training examples according to the results shown by the “Real-Time Tuning” window. 8. The validated algorithm was saved as “SGB COPOC IHC DenseNet TLS T zone FINAL” and launched in the ROI via the “Analyze. ..” option.

[0271] 9. Results of tissue segmentation were displayed in the “Results” tab.

[0272] 2.5. Nuclei seg Al algorithm for cell segmentation of DCs

[0273] 1. Anew classifier algorithm was created and Nuclei Seg was chosen. By default, on IHC slides, HALO-AI’s Nuclei seg detects nuclei on hematoxylin and eosin (HE) staining slides based on the blue nuclear staining pattern. Here, CD3 staining corresponded to blue staining and DC to red staining, thus no DC-LAMP+DC could directly be detected by this neuronal network by default. Customization of Nuclei seg was imperative to obtain accurate detection of T zone associated DC. The customization was required for the cell segmentation of DC in order to switch the detection from blue to red staining and to detect DC-LAMP+DC based on their entire shape instead of their nuclei.

[0274] 2. HALO-AI default resolution (0.25 pm / px) and Minimum Object Size (0 pm) settings were retained.

[0275] 3. “Real-Time Tuning” window was used to check the accuracy of the cell segmentation performed by the default Nuclei seg algorithm.

[0276] 4. Custom Nuclei seg algorithm was created by adding relevant training examples in “Background” and “Nuclei” classes. Important precision was needed to outline positive cells, thus zoom magnification was set around 200x by using HALO’s “zoom beyond image resolution” keyboard shortcuts. For the “Nuclei” class, a high number of DC with various shapes, intensities, as well as organization were precisely outlined at this magnification on several representative images to cover intra and inter-sample heterogeneities.

[0277] For custom Nuclei seg algorithm training, 16 samples distributed over 3 NSCLC cohorts were used to generate the final algorithm. In total, for cell segmentation, 1221 positive DC and 601 negative regions were precisely manually outlined to train the algorithm and obtain high accuracy detection of DC-LAMP+DC especially in T zones based on image cell compositions, cell morphologies, cell associations, cell densities, and cell distributions.

[0278] Cohorts Background Nuclei Total number of training annotations

[0279] 1 229 929 1158

[0280] 2 7 7

[0281] 3 365 292 657

[0282] Total 601 1221 1822

[0283] 3 cohorts 16 slides

[0284] Table 4: Training Annotations Numbers for DCs Nuclei Seg algorithm 5. Custom Nuclei seg algorithm training was performed by clicking on “Start Training” and by waiting for several iterations and for the cross-entropy to decrease and tend towards zero or to stabilize.

[0285] 6. “Real-Time Tuning” window was used to live view the training results and verify the accuracy of the segmentation of T zone-associated DC.

[0286] 7. The algorithm was corrected until its validation (>85% accuracy) by drawing new training examples according to the results shown by the “Real-Time Tuning” window.

[0287] 8. The validated algorithm was saved as “SGB COPOC IHC Nuclei seg DCs FINAL” and launched in the ROI via the “Analyze. ..” option.

[0288] 9. Results of tissue segmentation were displayed in the “Results” tab (“Summary Data” table with area (mm2) and total nuclei count, “Object data” table).

[0289] 2.6. Use of HighPlex IHC module to detect DC speci fically in T zones.

[0290] For CD3 / DC-LAMP IHC staining, HighPlex H4C module was used to perform the cell segmentation of DC-LAMP+DC specifically inside the classified “T zone” areas detected by 2ndDenseNet algorithm.

[0291] This step is not necessary when using HALO-AI v4.0 thanks to the Classifier pipeline option. HALO-AI v4.0’s Classifier pipeline enable users to build a daisy chain of classifiers for the analysis in order to, for example, directly launch a Nuclei seg algorithm inside a specific class of tissue defined by a tissue segmentation algorithm. Thus, in HALO-AI v4.0, Nuclei seg of T zone DC (3rdalgorithm) can directly be launched inside the “T zone” class of the associated DenseNet (2ndalgorithm) bypassing the need to use any HighPlex IHC modules.

[0292] In Halo- Al v3.6, HighPlex IHC was needed to allow simultaneously the use of the validated Nuclei seg of DCs with DenseNet algorithm in order to obtain detection of DC inside T zones. HighPlex IHC module is usually used to phenotype cells based on parameters and cell intensity thresholds. This module can be used individually to phenotype cells or can combine different types of Al algorithms to obtain a single analysis.

[0293] 1. A new HighPlex IHC module was opened in the “Analysis” tab.

[0294] 2. Custom Al algorithm option was chosen via “Al Custom” option in the “Nuclear segmentation type” section. 3. Then validated Al Nuclei seg algorithm was loaded in the “Nuclear Segmentation Classifier” section, as well as validated DenseNet algorithm in the “Classifier” section.

[0295] 4. Other parameters were filled in a way to remove all cell segmentation filters set by default.

[0296] 5. HighPlex IHC was set as detailed in the Figure 2.

[0297] 6. Results were visualized using “Output image” then “Colocalization (Interactive)” and by opening the “Real-Time Tuning” window.

[0298] 7. Finally HighPlex IHC settings were saved as “SGB COPOC IHC High Plex IHC DC-T zone FINAL” and complete analysis was performed in order to specifically detect and quantify DCs inside T zones.

[0299] 8. The HighPlex FL analysis gives quantitative data (count, percent, mean intensities, Stroma / Tumor / TLS compartments) for each marker of interest including Summary Data table (counts, percentages per phenotypes) and single-cell Object Data table (cell ID, XY coordinates, cell type based on thresholding (HighPlex FL settings), phenotypes based on Object Phenotyper (Al Algorithm), confidence frequency of phenotype attribution).

[0300] After validation of the three Al algorithms and settings of the analysis module, analysis was launched in batch on all NSCLC cohorts and CRC tumors and generated summary data as well as single cell data tables.

[0301] 2. 7. Validation on CD20 / CD21 stained IHC slides

[0302] The analysis was then performed and validated on CD20 / CD21 stained IHC slides from a new cohort of 125 NSCLC patients.

[0303] Based on visual inspection, the reliability of TLS detection is very high (approximately 85- 98%) for the majority of CD20 / CD21IHC slides in the cohort.

[0304] The surface area values obtained by quantifying B follicles and germinal centers (GCs) in TLS, as detected by Immunohistochemistry (IHC) CD20 / CD21 method as described above (TLS quant) was compared to the surface area obtained by another standardized method on slides stained by immunofluorescence (IF) as described in Example 1.

[0305] One IHC slides from a patient was excluded from the analysis due to an issue with the CD20 IHC staining (the staining too weak to be quantifiable). Figure 4 showed the correlation between the two analyses, validating the use of the present method for TLS quantification on this new cohort for CD20 / CD21 staining.

Claims

49CLAIMS1. A computer implemented method for detecting tertiary lymphoid structures (TLS) in at least one image of a tumor tissue section previously stained with at least: B cell marker such as CD20, follicular dendritic cell (FDC) marker such as CD21, T cell marker such as CD3 and mature dendritic cell (mDC) marker such as DC-LAMP, wherein the method comprises the steps of: a) determining a region of interest, preferably comprising tumoral tissue, in said image, b) segmenting tissue in the region of interest via a trained model, c) locating B-cell zone and T-cell zone within the segmented tissue, FDC within B-cell zone and mature DC within T-cell zone via trained models, wherein the density of a TLS is determined by the densities of B-cell zone and T-cell zone, and mature TLS is determined by the presence of FDC within B- cell zone and mature DC within T-cell zone.

2. The method according to claim 1 wherein said tumor tissue section is stained by immunohistochemistry or immunofluorescence with an anti-CD20, anti-CD21, anti- CD3 and anti-DC LAMP antibodies.

3. The method according to claim 1 or 2, wherein the TLS is detected in two images of adjacent tumor tissue sections, preferably in a first section stained with B cell and FDC markers and a second adjacent section to the first one stained with T cell and mDC markers.

4. The method according to claim 3 wherein in step a) the region of interest is determined in two images of adjacent tumor tissue sections, and in step b) B-cell zone and FDC are located in the image section via a trained model, and T-cell zone and mDC are located, in the adjacent section via a trained model.

5. The method according to any one of preceding claims, wherein the models in steps b) and c) have been trained by supervised learning on a training dataset formed by images50 of tumor tissue sections previously stained with at least: B cell, FDC, T cell and mDC markers and manually segmented into tissue, zone of cells or cells classes of interest.

6. The method according to any one of claims 1 to 5, wherein, said tumor tissue section is further stained with tumor cell marker such as pan-cytokeratins marker and / or high endothelial venule (HEV) cell marker such as Peripheral node addressin (PNAd) marker, preferably wherein said tumor tissue section is stained by immunohistochemistry or immunofluorescence with anti-pan-cytokeratins antibodies or an anti-PNAd antibody.

7. The method according to claim 6 wherein the method comprises in the step c) further locating tumoral cells and / or HEV cells via a trained model and wherein the maturity of the TLS is determined by the density of FDC within B-cell zones, mDC within T-cell zones and the densities of high endothelial venules within tumoral stroma, preferably wherein the models in the step c) has been trained by supervised learning on a training dataset formed by images of tumor tissue sections previously stained with: B cell, FDC, T cell, mDC markers, HEV and / or tumoral cell markers and manually segmented into tissue, zone of cells or cells classes of interest.

8. The method according to any one of claims 1 to 7 wherein said tumor tissue section is further stained with a nucleic marker, preferably DAPI, and preferably wherein said method further comprises a step of segmenting each cell in said tissue via a trained model, more preferably wherein the model has been trained by supervised learning on a training dataset formed by images of tumor tissue sections previously stained with nuclei markers and segmented into cell and nuclei.

9. The method according to any one of claims 1 to 8, comprising exporting the count, surface, coordinates or mean intensity of each tissue, cells or zone of cells.

10. The method according to any one of preceding claims wherein said tumor tissue is obtained from a patient suffering from a cancer, preferably selected from the group consisting of lung cancers, colorectal cancers, cervical cancers and breast cancers, preferably a non-small-cell lung cancer.5111. An in vitro method for prognosing the survival outcome of a patient suffering from a cancer comprising detecting the tertiary lymphoid structures in a tumor tissue section previously obtained from said patient by a computer-implemented method according to any one of claims 1 to 10, preferably wherein a higher TLS density, preferably a higher mature TLS density compared to a predetermined reference value is indicative of a higher survival outcome in said patient.

12. An in vitro method for evaluating the response to a cancer therapy of a patient suffering from a cancer comprising detecting the tertiary lymphoid structures in a tumor tissue section previously obtained from said patient by a computer-implemented method according to any one of claims 1 to 10, preferably wherein a higher TLS density, preferably a higher mature TLS compared to a predetermined reference value is indicative of a higher response to a cancer therapy in said patient.

13. A computer program product comprising code instructions which, when executed by a computer, cause said computer to implement a method according to any of the preceding claims.

14. A non-transitory computer readable storage medium having stored thereon code instructions which, when executed by a computer, causes said computer to implement a method according to any of claims 1 to 12.

15. A computing system, comprising at least one memory and one or more processors, configured to carry out the method according to any of claims 1 to 12.

Images