Feature processing of tissues and / or cells

Abstract

The technology described herein provides methods and systems related to feature processing of tissues and / or cells (e.g., to provide histological information describing the tissues and / or cells). In particular, methods and systems are described that classify, segment, and / or analyze target classes in images of (e.g., representing) tissues and / or cells. The technology finds use in validating the safety and quality of the objects. Aspects of the technology use artificial intelligence machine learning to classify, segment, and / or analyze ovarian follicles from images of human donor ovarian tissue.

Description

FEATURE PROCESSING OF TISSUES AND / OR CELLSCROSS-REFERENCE TO RELATED APPLICATION

[0001] This application claims the benefit of U.S. Provisional Patent Application No. 63 / 562,992, filed March 8, 2024, which is incorporated by reference herein in its entirety.GOVERNMENT SUPPORT

[0002] This invention was made with government support under HD104173 awarded by the National Institutes of Health. The government has certain rights in the invention.Fil l !)

[0003] The technology described herein provides methods and systems related to feature processing of tissues and / or cells (e.g., to provide histological information describing the tissues and / or cells). In particular, some embodiments provide methods and systems to classify, segment, and / or analyze target classes in images of (e.g., representing) tissues and / or cells. In some embodiments, the technology finds use in validating the safety and quality of therapeutics and treatments. In some embodiments, artificial intelligence machine learning is used to classify, segment, and / or analyze ovarian follicles from images of human donor ovarian tissue.BACKGROUND

[0004] Cancer patient survival rates in pre-pubertal girls and young women have increased due to increasing effectiveness of cancer treatments. However, many cancer treatments are gonadotoxic and cause premature ovarian insufficiency (POI), loss of fertility, and loss of ovarian endocrine function. For these patients and others who suffer from reduced ovarian endocrine function, a lack of functioning ovaries causes irregular pubertal development, which may include associated irregular bone development, lipid metabolism, and cognitive development. Hormone replacement therapy (HRT) is commonly used as an off-label treatment for patients having reduced ovarian endocrine function. However, HRT is typically targeted and approved for post-menopausal patients. Thus, HRT is not always appropriate to deliver or recapitulate appropriate physiologic hormone levels for all patients (e.g., pre-pubertal girls) because HRT was not developed for all patient classes, including pre-pubertal girls, young women, or patients who are not candidates for controlled ovarian stimulation. Ovarian tissue cryopreservation (OTC) and subsequent autografting is another potential treatment option for these patients, but OTC has associated risks such as introducing cancer cells into the patientbody from autologous ovarian tissue transplantation, particularly from patients with leukemia or lymphoma. In addition, OTC has a high rate of tissue rejection and / or allergic reaction.

[0005] An alternative treatment for reduced ovarian endocrine function is implanting allogenic ovarian tissue within immune-isolating capsules to minimize and / or eliminate rejection while allowing nutrients, waste products, and hormone signal molecules to diffuse across the immune-isolating capsules. See, e.g., U.S. Pat. Nos. 10,918,673 and 11,786,560, each of which is incorporated herein by reference.

[0006] Some conventional solutions comprise use of cryopreserved ovarian tissues from deceased human donors. These cryopreserved tissues have a high quality and high follicle survival rates, even with operation and transportation time. However, the success rates of treatments using cryopreserved ovarian tissues are largely determined by follicle density (FD) in the tissue, which is currently determined by manual analysis of histologic sections and thus suffers from inherent variability among those performing the manual analysis. Unfortunately, development of imaging and analysis technologies (e.g., imaging and analysis technologies for cancer research and treatment) has typically excluded technologies for imaging and analyzing ovarian tissue. As a result, follicle identification and classification is difficult, inefficient, and often inaccurate.

[0007] Thus, new technologies are needed to identify follicles in human donor ovarian tissue and to quantify FD, e.g., with increased efficiency and / or accuracy. Such technologies may be useful in validating the safety and quality of human donor ovarian tissue (e.g., as indicated by anti-Mullerian hormone concentration, follicle density, number of follicles, morphological ovarian changes associated with age or disease, and developmental stage of follicles).SUMMARY

[0008] The technology described herein provides methods and systems related to feature processing of tissues and / or cells (e.g., to provide histological information describing the tissues and / or cells). In particular, some embodiments provide methods and systems to classify, segment, and / or analyze target classes in images of (e.g., representing) tissues and / or cells. In some embodiments, the technology finds use in validating the safety and quality of therapeutics and / or treatments. In some embodiments, artificial intelligence machine learning is used to classify, segment, and / or analyze ovarian follicles from images of human donor ovarian tissue.

[0009] In some embodiments, the technology provides a method (e.g., a computer-based method) comprising: a) imaging an ovarian or ovarian-associated tissue (e.g., a processedovarian tissue) to provide an image of the ovarian tissue; b) annotating a number of follicles of the image of the ovarian tissue; c) counting a number of follicles of the image of the ovarian tissue; and d) providing a follicle density of the ovarian tissue. In some embodiments, methods further comprise processing an ovarian tissue to provide a processed ovarian tissue. In some embodiments, methods comprise measuring a number of follicles of the image of the ovarian tissue.

[0010] The present disclosure is not limited to particular ovarian or ovarian associated tissues. Examples include but are not limited to, a whole ovarian cross-section (e.g., ovarian mesenteric cross section), human ovarian cortex strips, corpus luteums, corpus albicans, blood vessels, lymph vessels, or stromal cells.

[0011] In some embodiments, processing an ovarian tissue comprises one or more of the following steps: a) decertifying an ovarian cortex of an ovarian tissue to provide a histological sample; b) fresh-fixing the histological sample to provide a fresh-fixed sample; c) washing the fresh-fixed sample (e.g., in water (e.g., deionized water)) to provide a washed sample; d) dehydrating the washed sample to provide a dehydrated sample; e) orienting the dehydrated sample perpendicular to the ovarian cortex surface and embedding the dehydrated sample in paraffin to provide a paraffined sample; f) serially sectioning the paraffined sample to provide a plurality of (e.g., four or more (e.g., 4, 5, 6, 7, 8, 9, 10, or more)) sectioned samples; g) adding the plurality of sectioned samples to a slide; and / or h) staining the slide.

[0012] In some embodiments, imaging an ovarian tissue (e.g., a processed ovarian tissue) comprises: a) whole slide imaging the ovarian tissue with a high-resolution scanner to provide a whole slide image file; and b) importing the whole slide image file into a first software.

[0013] In some embodiments, annotating a number of follicles of the image of the ovarian tissue comprises: a) selecting a tissue section region of interest from the whole slide image file using the first software; b) generating a tissue section number: i) for the tissue section region of interest; and ii) related to a follicle depth and position; c) annotating the tissue section number to provide an annotated tissue section: i) by color coding; ii) by positional location fidelity; and iii) by classifying the one or more follicle; and d) exporting the annotated tissue section from the first software to a second software.

[0014] In some embodiments, measuring a number of follicles of the image of the ovarian tissue comprises measuring, using the second software: (i) a distance from the cortex to the centroid of the follicle; and / or (ii) a tissue section area. In some embodiments, counting anumber of follicles of the image of the ovarian tissue comprises counting, using the second software, a plurality of follicles of the image of the ovarian tissue.

[0015] The described imaging, annotating, measuring and counting; and providing steps can be performed manually (e.g., by an operator) or automated (e.g., using a computer processor and computer software) or a combination thereof.

[0016] In some embodiments, the method further comprises determining one or more patient variables selected from, for example, age, body mass index, ovary volume, ovary volume, weight, ethnicity, Kidney Donor Profile Index (KDPI), process duration, Cold Ischemic Time (CIT), blood type, and cause of death (CoD).

[0017] In some embodiments, the technology also provides systems (e.g., a computer-based system) comprising: a) one or more imaging device; b) one or more communication component; c) one or more first software (e.g., as described herein); and d) one or more second software (e.g., as described herein). In some embodiments, the system (e.g., the computer- based system) is configured to detect a follicle density in an image of an ovarian tissue and to perform actions on the image of the ovarian tissue.

[0018] In some embodiments, the one or more communication component comprises an input device and / or an output device.

[0019] In some embodiments, the one or more first software and / or the one or more second software provide one or more protocols for: a) imaging an ovarian tissue to provide an image of the ovarian tissue; b) annotating one or more follicles of the image of the ovarian tissue; c) counting a number of follicles of the image of the ovarian tissue; and / or d) providing a follicle density of the ovarian tissue.BRIEF DESCRIPTION OF THE DRAWINGS

[0020] The patent or application file contains at least one drawing executed in color. Copies of this patent or patent application publication with color drawing(s) will be provided by the Office upon request and payment of the necessary fee.

[0021] These and other features, aspects, and advantages of the present technology will become better understood with regard to the following drawings.

[0022] FIG. 1 is a workflow diagram providing steps of a method described herein (e.g., in some embodiments provided in the software utilized). Follicle measurements included follicle depth, centroid, and tissue section area.

[0023] FIG. 2 is a whole-slide scanned .SVS image showing four serial tissue section ROIs viewed in QuPath. This is an example of a whole slide scan image from Donor 17, Strip 2,slide 11. The red outlined ROIs highlight serial sections 1-4 (from right to left). The tissue in the bottom leftmost position was the experimental tissue section used in the Examples. Note: For cost and labor reduction, multiple tissue samples from one donor may be embedded, sectioned, stained, and whole slide scanned simultaneously on the same slide, as illustrated by this image showing a second set of serial sections from a separate tissue strip sample.

[0024] FIGS. 3A-3E: Ovarian follicle classifications and tissue and follicle annotations as seen in the QuPath software. FIG. 3A is the preantral ovarian follicle classification definitions. FIG. 3B is the outdated follicle classification system used by the control operator for thirteen donor samples. FIG. 3C is the updated follicle classification system used by the control operator for four donor samples and all donor samples by the comparison operator. FIG. 3D is an image showing follicle annotations overlayed upon an image of an H&E stained ovarian tissue strip in QuPath. FIG. 3E shows a digital image output by QuPath in the form of a binary colorized 8-bit .PNG follicle annotation.

[0025] FIGS. 4A-4C is an image of a tissue that has been filtered and thresholded in Imagel. FIG. 4A shows total outline. FIG. 4B shows cortex Outline. FIG. 4C shows QuPath output imported into ImageJ.

[0026] FIGS. 5A-5B show images of tissue with follicle annotations that were filtered, thresholded, and the distance to the ovarian surface calculated. FIG. 5A is a greyscale image showing the centroids of each follicle, represented as the brightest point. FIG. 5B shows a distance map with centroid points and cortex outline. FIG. 5C shows the cortical surface outline.

[0027] FIG. 6 is a plot showing total follicle count per slide by control vs comparison operator. Each dot represents a single slide overall follicle count (n=99) by the operator. Mean and SEM are plotted. A column chart with SEM of the control vs comparison operator follicle counts per slide with p<0.0001.

[0028] FIGS. 7A-7D are the labelling and results in Ilastik after training with seven images. FIG. 7A shows the Ilastik interface. FIG. 7B shows the original H&E-stained image with labels of Follicle and Tissue (which includes the background and anything that is not “Follicle”). FIG. 7C shows Live Update Probability results. FIG. 7D shows Live Update Segmentation results.

[0029] FIGS. 8A-8C are images as processed by Fiji. FIG. 8A shows Ilastik .TIF image export viewed in Fiji. FIG. 8B shows image thresholded, filtered, watershed, particle range defined and counted. FIG. 8C shows Fiji results from counting and analyzing the follicles.

[0030] FIG. 9 is an image showing merged Ilastik export (green) output overlayed on the original .TIF file (red). The output was modified to assign the channel to green and selecting a glasbey lookup table for visualization in RGB.

[0031] FIGS. 10 A- 10C are plots showing total follicle count per slide by operator compared with artificial intelligence (Al). Each dot represents a single sections total follicle count by each method. Mean and SEM are plotted with p<0.05. FIG. 10A is a column chart showing Al compared with control operator follicle counts per section. FIG. 10B is a column chart showing the control operator compared with Al compared with comparison operator follicle counts per section. FIG. 10C is a residual plot of FIG. 8B with comparison operator contributing to most of the dots above the x-axis and the Al contributing to most of the plots below the x-axis, with the control operator on the x-axis.

[0032] FIGS. 11A-11C are screen shots showing labelling and results in Ilastik Al after a first study in which the software was trained with 10 images.

[0033] FIGS. 12A-12C are screen shots showing labelling and results in Ilastik Al after a second study in which the software was trained with 10 images.

[0034] FIGS. 13A-13D are screen shots showing labelling and results in Ilastik Al after a third study. FIG. 13 A shows the Ilastik interface. FIG. 13B shows the original H&E-stained image with labels of Follicle, Tissue, and Background. FIG. 13C shows Live Update Probability results. FIG. 13D shows Live Update Segmentation results.

[0035] FIGS. 14A-14C are screen shots showing labelling and results in Ilastik Al after a fourth study in which the software was trained with 3 images.

[0036] FIGS. 15A-15C are screen shots showing labelling and results in Ilastik Al after a fifth study in which the software was trained with 7 images.

[0037] FIGS. 16A-16C are screen shots showing labelling and results in Ilastik Al after a sixth study in which the software was trained with 7 images. The same process as the fourth study was used, though some of the follicle labels from study five were erased.

[0038] FIGS. 17A-17F are images showing the artificial intelligence and automated image processing workflow. The ultimate output of the workflow is a follicle count. Each image represents a visual processing step in the artificial intelligence and automated workflow. An adjacent slide was used to train the Al (Steps B & C) and then used for the second and third Al training and analysis (Steps A, D, & F). The Pixel Classification (Step D) was based on previous training using the test image results from the Density Counting (Step C), which in turn involved discarding the labeled image for Steps E & F. FIG. 17A shows, in QuPath, thenative whole scan .SVS images and shows the four H&E-stained tissue sections on the slide which are cropped into individual .TIF files. FIG. 17B shows, in Fiji, the .TIF files converted to .HDF5 files via Ilastik plugin. FIG. 17C shows, in Ilastik, the .HDF5 files imported into a Density Counting project, and a combination of selecting correct features (indicative of follicles) and manually labeling (orange arrowhead, and red or green overlays) employed. Performance of the trained Al is judged visually by the operator ‘trainer’ or ’user’ by the correctly identified features (white arrowhead E> , pale yellow highlight) and incorrectly identified features (red arrowhead ► , pale yellow highlight). Improvements in training occur sequentially in steps I, and ii. Once performance is satisfactory to the user, the trained program is applied, with output options specified by the user. FIG. 17D shows the output from the Density Counting project used to train a Pixel Classification program. The same iterative training is employed to achieve improved identification results. The test images that were labeled were discarded. FIG. 17E shows, in Fiji, an automated macro to apply the Robust Automatic Threshold Selection plugin used to reduce noise in the test images. FIG. 17F shows, in Ilastik, the Object Classification using size and simple threshold criteria to determine follicle and tissue count per image section. First, the program overlays identified objects (white overlays), which have not yet been labeled (dark green arrowhead ►). Then, users can train which labels are correct and obtain a more accurate classification (orange arrowhead, and blue or yellow overlays). The trained program is run on the final test image set, with output options specified to obtain a .TIF output map with follicle and tissue labels and .CSV spreadsheet with follicle and tissue count and selected feature measurements. This last Ilastik processing is done to minimize the number of missed follicles (cyan arrowhead), and with the aim of maximizing correctly identified tissue (yellow arrowhead), and correctly identified follicles (green arrowhead).

[0039] FIG. 18 shows the Ilastik object Classification feature settings.

[0040] FIGS. 19A-19B shows representative H / E-stained deceased human donor ovarian cortex sections of each donor analyzed and follicle growth stages. FIG. 19A shows deceased human donor ovarian cortex H&E-stained sections taken at 20x. Complete WSI .TIF tissue images are cropped (2500x2500microns) in the middle of the tissue section to better visualize the cortex. Scale bars are 150 microns and 40 microns. Donor age and donor number are noted and referenced in Table 2. Donors: i) 1, ii) 2iv) 4, v) 5, vii) 7, viii) 8, x) 10, xi) 11 each had one sample assessed. Samples from donors iii) 3, vi) 6, ix) 9, and xii) 12 were taken from multiple sites on the same ovary, see FIGS. 22-23. FIG. 19B shows deceased human donor ovariancortex H&E-stained sections taken at 20x, with representative follicle examples from the sample population according to the following follicle classifications: i) Primordial, ii) Transitional Primordial, iii) Primary, iv) Transitional Primary, v) Secondary, vi) Multilayer, vii) Antral. Scale bars i-iv are 20-micron, v & vi are 50-micron, and vii is 200-micron. Table 1 for pre-antral ovarian follicle classification definitions.

[0041] Table 1. Pre-antral Ovarian Follicle Classification Definitions

[0042] FIGS. 20A-20B show donor primordial follicle density from “comparison” operator manual counts, exploring age and primordial follicle density relationships. FIG. 20A shows a histogram of each donor’s primordial follicle density. Donors are arranged by increasing age (16-37 y / o). Each dot represents the primordial FD per section sampled. Six to eighteen sections were analyzed per donor. FIG. 20B shows age vs primordial follicle density. Each dot represents the average primordial follicle density per donor (n = 12 donors). Six to eighteen sections were analyzed per donor. Statistical Analysis: Kruskal- Wallace with Dunn’s multiple comparison test *p<0.05, **p<0.01, ***p<0.001, and ****p<0.0001. Error Bars: Mean ± standard error of the mean are plotted.

[0043] FIGS. 21 A-21C show age and BMI vs primordial follicle density statistics. FIG. 21A shows BMI vs primordial follicle density. Each dot represents the average primordial follicle density per donor (n=12 donors). Six to eighteen sections were analyzed per donor. FIG. 21B shows BMI vs primordial follicle density with groups: underweight / BMI = <18.5 kg / nr (n = 2donors, 18 sections), normal / BMI = 18.5 - 24.9 kg / m2(n = 3 donors, 30 sections), overweight / BMI = 25 - 29.9 kg / m2(n = 3 donors, 18 sections), and obese / BMI = >30 kg / m2(n = 4 donors, 36 sections). FIG. 21C shows age vs primordial follicle density with donor age groups: 16-20 y / o (n = 4 donors, 30 sections), 24-29 y / o (n = 4 donors, 30 sections), and 33-37 y / o (n = 4 donors, 42 sections). Each dot represents the primordial FD per section sampled. The mean and SEM with P<0.05 are shown on graph. Each dot represents the primordial FD per section sampled. Statistical Analysis: Kruskal-Wallace with Dunn’s multiple comparison test *p<0.05, **p<0.01 , ***p<0.001 , and ****p<0.0001 . Error Bars: Mean + standard error of the mean are plotted.

[0044] FIGS. 22A-22E show primordial follicle depth from the ovarian surface. The primordial follicle distance from the ovarian surface “depth” was assessed from 17 sections (FIGS. 22A-22B) and within the samples from the same-donor (FIGS. 22C-22E). For FIGS. 22A-22B, total primordial follicles, n = 562. FIG. 22 A shows age vs primordial follicle depth with donor age groups: 16-20 y / o (n = 339 follicles, 30 sections, 4 donors), 24-29 y / o (n = 195 follicles, 30 sections, 4 donors), and 33-37 y / o (n = 28 follicles, 42 sections, 5 donors). Each dot represents the primordial FD per section sampled. FIG. 22B shows BMI vs primordial follicle depth: underweight / BMI = <18.5 kg / m2(n = 257 follicles, 18 sections), normal / BMI = 18.5 - 24.9 kg / m2(n = 205 follicles, 30 sections), overweight / BMI = 25 - 29.9 kg / m2(n = 25 follicles, 18 sections), and obese / BMI = >30 kg / m2(n = 75 follicles, 36 sections). The mean and SEM with P<0.05 are shown on graph. Each dot represents the primordial FD per section sampled. Statistical Analysis: Friedman test with Dunn’s multiple comparison were performed on primordial follicle depth grouped by age or BMI. Mann- Whitney tests were performed on the primordial follicle depth per section. For FIGS. 22C-22E, -a or -b denotes different sample locations within the same ovary of the donor. These donors, with two different samples assessed per donor, were chosen to span a large age range (18-37 y / o). FIG. 22C shows donor 3: a) n = 77 follicles, b) n = 180 follicles. FIG. 22D shows donor 6: a) n = 12 follicles, b) n = 120 follicles. FIG. 22E shows donor 12: a) n = 2 follicles, b) n = 5 follicles. *p<0.05, **p<0.01 , ***p<0.001, and ****p<0.0001. Error Bars: Mean ± standard error of the mean are plotted.

[0045] FIGS. 23A-23C show the same donor primordial follicle density statistics. The primordial follicle density of donors with multiple samples taken from the same ovary were assessed to determine intra ovary-location variation. For FIGS. 23A-23C, -a or -b denotes different sample locations within the same ovary of the donor. Two different samples were assessed per donor. FIG. 23 A shows donor 3 (18 y / o). FIG. 23B shows donor 6 (25 y / o). FIG.23C shows donor 12 (37 y / o). Mann- Whitney tests were performed on the average primordial FD per section. *p<0.05, **p<0.01, ***p<0.001, and ****p<0.0001. Error Bars: Mean ± standard error of the mean are plotted; n = 6 sections per sample.

[0046] FIGS. 24A-24H show operator and Al follicle count performance. FIG. 24 A shows control operator, comparison operator, Al follicle counts per tissue section, and Al follicle counts per tissue section. Each dot represents a single tissue section’s follicle count (n=77) by the chosen method. FIG. 24B shows correlation between the average manual (both operators) follicle count per tissue section vs the Al average follicle count per tissue section. Simple linear regression (red): Y = 0.826X - 0.612, R2= 0.9661. Each dot represents a single tissue section follicle count (n=77). FIG. 24C shows control operator, comparison operator, and Al average follicle counts per donor. Each dot represents a single donor follicle count (n=l 1 donors with 2 -15 tissue sections counted per donor) by the chosen method. FIG. 24D shows correlation between the comparison and control operators’ combined average follicle count per donor vs the Al average follicle count per donor. Simple linear regression (red): Y = 1.190X + 1.370, R2= 0.9796. Each dot represents the average follicle count per donor (n=l 1) FIG. 24E shows average manual follicle count per donor and Al follicle count per donor shown with upper SD and y-axis log 10 (n = 11). FIG. 24F shows a Bland- Altman plot with the control operator follicle count per tissue section represented on the x-axis, and the difference in controlcomparison operator follicle count per tissue section represented on the y-axis. Gray dashed lines represent the 95% limit of agreement, and the mean is represented by the solid black line. Each dot represents a single tissue section follicle count (n=77). FIG. 24G shows a Bland- Altman plot with the control operator follicle count per tissue section represented on the x-axis, and the difference in control operator- Al follicle count per tissue section represented on the y- axis. Gray dashed lines represent the 95% limit of agreement, and the mean is represented by the solid black line. Each dot represents a single tissue section follicle count (n=77). FIG. 24H shows a histogram of manual follicle count by control operator binned by 20, and average Al accuracy compared to the control operator. In order of bin, n (in follicle) = 43, 4, 6, 5, 0, 4, 1, 1. Statistical Analysis: Friedman test with Dunn’s multiple comparison test *p<0.05, **p<0.01, ***p<0.001, and ****p<0.0001. Error Bars: Mean + standard error of the mean are plotted.

[0047] FIG. 25 shows deep learning to estimate follicle density from patient features. FIG. 25A shows a representation of workflow to prepare and augment histology training and test sets, identify highly important features in determining follicle density, and weight features for follicle density estimation using sparse histological samples. FIG. 25B shows weighting of thetop 13 patient metadata features in correlation to follicle density in the training set. FIG. 25C shows receiver-operator characteristic curve with area under the curve (AUC) indicated for the confidence vs. precision of these weighted features in determining the follicle density of the test dataset. FIG. 25D shows principal component analysis of these 13 -weighted features, classified by spacing coefficient and intra-class correlation (sampling need). FIG. 25E shows model-estimated follicle density (solid line) compared to manually counted follicle samples (points) Shaded grey area represents 90% confidence interval and dotted lines represent 90% prediction interval for model estimated follicle density.Detailed Description

[0048] The technology described herein provides methods and systems related to feature processing of tissues and / or cells (e.g., to provide histological information describing the tissues and / or cells). In particular, some embodiments provide methods and systems to classify, segment, and / or analyze target classes in images of (e.g., representing) tissues and / or cells. In some embodiments, the technology finds use in validating the safety and quality of the objects. In some embodiments, Al machine learning is used to classify, segment, and / or analyze ovarian follicles from images of human donor ovarian tissue.1. Definitions

[0049] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art. For example, any nomenclatures used in connection with, and techniques of, cell and tissue culture, molecular biology, immunology, microbiology, genetics and protein and nucleic acid chemistry and hybridization described herein are those that are well known and commonly used in the art.

[0050] In case of conflict, the present document, including definitions, will control. Preferred methods and materials are described below, although methods and materials similar or equivalent to those described herein can be used in practice or testing of the present disclosure. Further, unless otherwise required by context, singular terms shall include pluralities and plural terms shall include the singular.

[0051] All publications, patent applications, patents and other references mentioned herein are incorporated by reference in their entirety. The materials, methods, and examples disclosed herein are illustrative only and not intended to be limiting.

[0052] The terms “comprise(s)”, “include(s)”, “having”, “has”, “can”, “contain(s)”, and variants thereof, as used herein, are intended to be open-ended transitional phrases, terms, or words that do not preclude the possibility of additional acts, steps, structures, components, orelements. The singular forms “a”, “and”, and “the” include plural references unless the context clearly dictates otherwise. The present disclosure also contemplates other embodiments “comprising”, “consisting of”, and “consisting essentially of’ the embodiments or elements presented herein, whether explicitly set forth or not.

[0053] For the recitation of numeric ranges herein, each intervening number therebetween with the same degree of precision is explicitly contemplated. For example, for the range of 6- 9, the numbers 7 and 8 are contemplated in addition to 6 and 9, and for the range 6.0-7.0, the number 6.0, 6.1 , 6.2, 6.3, 6.4, 6.5, 6.6, 6.7, 6.8, 6.9, and 7.0 are explicitly contemplated.

[0054] As used herein, the terms “about”, “approximately”, “substantially”, and “significantly” are understood by persons of ordinary skill in the art and will vary to some extent on the context in which they are used. If there are uses of these terms that are not clear to persons of ordinary skill in the art given the context in which they are used, “about” and “approximately” mean plus or minus less than or equal to 10% of the particular term and “substantially” and “significantly” mean plus or minus greater than 10% of the particular term.

[0055] Ranges can be expressed herein as from “about” one particular value and / or to “about” another particular value. When such a range is expressed, embodiments include from the one particular value and / or to the other particular value. Similarly, when values are expressed as approximations by use of the antecedent “about” it will be understood that the particular value forms an embodiment. It will be further understood that the endpoints of each of the ranges are significant both in relation to the other endpoint and independently of the other endpoint. It is also understood that there are a number of values disclosed herein, and that each value is also herein disclosed as “about” that particular value disclosed herein, and that each value is also herein disclosed as “about” that particular value in addition to the value itself. For example, if the value “10” is disclosed, then “about 10” is also disclosed.

[0056] The term “and / or”, when used in the context of a list of entities, is an inclusive “or” and refers to the entities being present singly or in combination.

[0057] As used herein, the term “subject” and “patient” are used interchangeably to refer to any vertebrate, including, but not limited to, a mammal (e.g., cow, pig, camel, llama, horse, goat, rabbit, sheep, hamster, guinea pig, cat, dog, rat, mouse, non-human primate (e.g., a monkey, such as a cynomolgus or rhesus monkey, chimpanzee, macaque, etc.), or a human). In some embodiments, the subject may be a human or a non-human. In some embodiments, the subject is a human. The subject or patient may be undergoing various forms of treatment.

[0058] The terms “administration of’ and “administering” a composition as used herein refer to providing a composition of the present disclosure to a subject in need of treatment. The compositions of the present disclosure may be administered by oral, parenteral (e.g., intramuscular, intraperitoneal, intravenous, ICV, intracistemal injection or infusion, subcutaneous injection, nebulization, or implant), by inhalation spray, nasal, vaginal, rectal, sublingual, or topical routes of administration and may be formulated, alone or together, in suitable dosage unit formulations containing conventional non-toxic pharmaceutically acceptable carriers, adjuvants, and vehicles appropriate for each route of administration.

[0059] As used herein, the terms “treat”, “treating”, and “treatment” are each used interchangeably to describe reversing, alleviating, or inhibiting the progress of a disease, condition, and / or injury, or one or more symptoms of such disease, condition, and / or injury, to which such term applies. Depending on the condition of the subject, the term also refers to preventing a disease or condition and includes preventing the onset of a disease or condition, or preventing the symptoms associated with a disease or condition. A treatment may be either performed in an acute or chronic way. The term also refers to reducing the severity of a disease or condition or symptoms associated with such disease or condition prior to affliction with the disease or condition. Such prevention or reduction of the severity of a disease or condition prior to affliction refers to administration of a treatment to a subject that is not at the time of administration afflicted with the disease or condition. “Preventing” also refers to preventing the recurrence of a disease or condition or of one or more symptoms associated with such disease or condition.

[0060] The term “assessing” includes any form of measurement and includes determining if an element is present or not. The terms “determining”, “measuring”, “evaluating”, “assessing”, and “assaying” are used interchangeably and include quantitative and qualitative determinations. Assessing may be relative or absolute. “Assessing the identity of” includes determining the most likely identity of a particular compound or formulation or substance and / or determining whether a predicted compound or formulation or substance is present or absent. “Assessing the quality of’ includes making a qualitative or quantitative assessment of quality e.g., through the comparisons of a determined value to a reference or standard of known quality.

[0061] The term “histology” and “histological” as used herein generally refers to microscopic analysis of the cellular anatomy and / or morphology of cells obtained from a multicellular organism including but not limited to plants and animals.

[0062] The terms “control”, “control assay”, “control sample”, and the like, refer to a sample, test, or other portion of an experimental or diagnostic procedure or experimental design for which an expected result is known with high certainty, e.g., to indicate whether the results obtained from associated experimental samples are reliable, indicate to what degree of confidence associated experimental results indicate a true result, and / or to allow for the calibration of experimental results. For example, in some instances, a control may be a “negative control” assay such that an essential component of the assay is excluded such that an experimenter may have high certainty that the negative control assay will not produce a positive result. In some instances, a control may be “positive control” such that all components of a particular assay are characterized and known, when combined, to produce a particular result in the assay being performed such that an experimenter may have high certainty that the positive control assay will not produce a positive result. Controls may also include “blank” samples, “standard” samples (e.g., “gold standard” samples), validated samples, etc.

[0063] The term “feature processing”, as used herein, generally refers to a classifying process of partitioning a digital image into multiple segments (sets of pixels, which in some instances may be referred to as a “feature”). Feature processing may be used to locate, classify, or otherwise isolate objects and / or boundaries (lines, curves, features, etc.) in an image. More precisely, feature processing is the process of including, excluding, or assigning a label to groups of pixels in an image corresponding to a feature such that pixels of a feature included in the image or with the same label share certain characteristics and / or correspond to a discrete biological structure (e.g., organelle, sub-cellular structure, etc.) in a tissue represented in the image.

[0064] General definitions of digital imaging terms not specifically provided herein may be found in various print and online sources known in the art, including but not limited to the Glossary at the Federal Agencies Digitalization Guidelines Initiative (FADGI) available online at www.digitizationguidelines.gov / glossary / .

[0065] The term “feature” as used herein refers to a group of pixels (e.g., a part or a pattern of an image) that share certain characteristics and / or correspond to a discrete biological structure (e.g., organelle, sub-cellular structure, etc.) in a tissue represented in the image.

[0066] As used herein, the term “object” refers to any material thing that can be seen and / or touched including, but not limited to, a tissue, a cell (e.g., an animal or a plant cell), and / or a cell component (e.g., parts within a cell (e.g., cell wall, cell membrane, cytoplasm, nucleus, and cell organelles)). In some embodiments, the object may be an animal cell or a plant cell. Insome embodiments, the object is a human tissue (e.g., a human ovary or human ovary tissue or human ovary cell).

[0067] As used herein, the suffix “-free” refers to an embodiment of the technology that omits the feature of the base root of the word to which “-free” is appended. That is, the term “X-free” as used herein means “without X”, where X is a feature of the technology omitted in the “X-free” technology. For example, a “calcium-free” composition does not comprise calcium, a “mixing-free” method does not comprise a mixing step, etc.

[0068] Although the terms “first”, “second”, “third”, etc. may be used herein to describe various steps, elements, compositions, components, regions, layers, and / or sections, these steps, elements, compositions, components, regions, layers, and / or sections should not be limited by these terms, unless otherwise indicated. These terms are used to distinguish one step, element, composition, component, region, layer, and / or section from another step, element, composition, component, region, layer, and / or section. Terms such as “first”, “second”, and other numerical terms when used herein do not imply a sequence or order unless clearly indicated by the context. Thus, a first step, element, composition, component, region, layer, or section discussed herein could be termed a second step, element, composition, component, region, layer, or section without departing from technology.

[0069] As used herein, the word “presence” or “absence” (or, alternatively, “present” or “absent”) is used in a relative sense to describe the amount or level of a particular entity (e.g., component, action, element). For example, when an entity is said to be “present”, it means the level or amount of this entity is above a pre-determined threshold; conversely, when an entity is said to be “absent”, it means the level or amount of this entity is below a pre-determined threshold. The pre-determined threshold may be the threshold for detectability associated with the particular test used to detect the entity or any other threshold. When an entity is “detected” it is “present”; when an entity is “not detected” it is “absent”.

[0070] As used herein, an “increase” or a “decrease” refers to a detectable (e.g., measured) positive or negative change, respectively, in the value of a variable relative to a previously measured value of the variable, relative to a pre-established value, and / or relative to a value of a standard control. An increase is a positive change preferably at least 10%, more preferably 50%, still more preferably 2-fold, even more preferably at least 5-fold, and most preferably at least 10-fold relative to the previously measured value of the variable, the pre-established value, and / or the value of a standard control. Similarly, a decrease is a negative change preferably at least 10%, more preferably 50%, still more preferably at least 80%, and mostpreferably at least 90% of the previously measured value of the variable, the pre-established value, and / or the value of a standard control. Other terms indicating quantitative changes or differences, such as “more” or “less,” are used herein in the same fashion as described above.

[0071] As used herein, a “system” refers to a plurality of real and / or abstract components operating together for a common purpose. In some embodiments, a “system” is an integrated assemblage of hardware and / or software components. In some embodiments, each component of the system interacts with one or more other components and / or is related to one or more other components. In some embodiments, a system refers to a combination of components and software for controlling and directing methods. For example, a “system” or “subsystem” may comprise one or more of, or any combination of, the following: mechanical devices, hardware, components of hardware, circuits, circuitry, logic design, logical components, software, software modules, components of software or software modules, software procedures, software instructions, software routines, software objects, software functions, software classes, software programs, files containing software, etc., to perform a function of the system or subsystem. Thus, the methods and apparatus of the embodiments, or certain aspects or portions thereof, may take the form of program code (e.g., instructions) embodied in tangible media, such as floppy diskettes, CD-ROMs, hard drives, flash memory, or any other machine-readable storage medium wherein, when the program code is loaded into and executed by a machine, such as a computer, the machine becomes an apparatus for practicing the embodiments. In the case of program code execution on programmable computers, the computing device generally includes a processor, a storage medium readable by the processor (e.g., volatile and non-volatile memory and / or storage elements), at least one input device, and at least one output device. One or more programs may implement or utilize the processes described in connection with the embodiments, e.g., through the use of an application programming interface (API), reusable controls, or the like. Such programs are preferably implemented in a high-level procedural or object-oriented programming language to communicate with a computer system. However, the program(s) can be implemented in assembly or machine language, if desired. In any case, the language may be a compiled or interpreted language, and combined with hardware implementations.

[0072] A “data processing unit”, as used herein, refers to any hardware and / or software combination that performs the functions required of it. For example, any data processing unit herein may be a programmable digital microprocessor such as available in the form of an electronic controller, mainframe, server or personal computer (desktop or portable). Where thedata processing unit is programmable, suitable programming can be communicated from a remote location to the data processing unit, or previously saved in a computer program product (such as a portable or fixed computer readable storage medium, whether magnetic, optical or solid state device based).

[0073] As used herein, the term “software” generally includes but is not limited to, one or more computer instructions and / or processor instructions that can be read, interpreted, compiled, and / or executed by a computer and / or processor. Software causes a computer, processor, or other electronic device to perform functions, actions and / or behave in a desired manner. Software may be embodied in various forms including routines, algorithms, modules, methods, threads, and / or programs. In different examples software may be embodied in separate applications and / or code from dynamically linked libraries. In different examples, software may be implemented in executable and / or loadable forms including, but not limited to, a standalone program, an object, a function (local and / or remote), a servelet, an applet, instructions stored in a memory, part of an operating system, and so on. In different examples, computer- readable and / or executable instructions may be located in one logic and / or distributed between multiple communicating, cooperating, and / or parallel processing logics and thus may be loaded and / or executed in serial, parallel, massively parallel and other manners.

[0074] Suitable software for implementing various components of example systems and methods described herein may be developed using programming languages and tools (e.g., Java, C, C#, C++, C, SQL, APIs, SDKs, Swift, NEXTStep, SmallTalk, Unix, Objective C, assembler). Software, whether an entire system or a component of a system, may be embodied as an article of manufacture and maintained or provided as part of a computer-readable medium. Software may include signals that transmit program code to a recipient over a network or other communication medium. Thus, in one example, a computer-readable medium may be signals that represent software / firmware as it is downloaded from a server (e.g., web server).

[0075] A “connection” by which two components of a system, e.g., an electrical system, a data system, a computer system, a circuitry system, etc., are connected will generally be an “operable connection”, or a connection by which entities are “operably connected”. The term “operable connection” and equivalents is one in which signals, physical communications, and / or logical communications may be sent and / or received. An operable connection may include a physical interface, an electrical interface, and / or a data interface. An operable connection may include differing combinations of interfaces and / or connections sufficient to allow operable control. For example, two entities can be operably connected to communicatesignals to each other directly or through one or more intermediate entities (e.g., processor, operating system, logic, software). Logical and / or physical communication channels can be used to create an operable connection.

[0076] The term “communication”, as used herein in relationship to “computer communication” or “data transfer”, refers to a communication between computing devices (e.g., computer, server, etc.) or component of a computer system (e.g., a memory store, a processor, a bus, an input device (e.g., a digital camera), an output device, etc.) and can be, for example, a network transfer, a file transfer, an applet transfer, an email, a hypertext transfer protocol (HTTP) transfer, and so on. A computer communication can occur across, for example, a wireless system (e.g., IEEE 802.11), an Ethernet system (e.g., IEEE 802.3), a token ring system (e.g., IEEE 802.5), a local area network (LAN), a wide area network (WAN), a point-to-point system, a circuit switching system, a packet switching system, and so on.

[0077] The terms “computer” or “computer component” or “component of a computer system”, as used herein, refers to a computer-related entity (e.g., hardware, firmware, software, and combinations thereof). Computer components may include, for example, a process running on a processor, a processor, an object, an executable, a thread of execution, a program, integrated circuitry, and a computer. A computer component(s) may reside within a process and / or thread. A computer component may be localized on one computer and / or may be distributed between multiple computers. The term “computer” as used herein generally includes a plurality of electrical and electronic components that provide power, operational control, and protection to the components and modules within the computer. For example, a computer can include, among other things, a processing unit (e.g., a microprocessor, a microcontroller, or other suitable programmable device), a memory, input units, and output units. The processing unit can include, among other things, a control unit, an arithmetic logic unit (“ALC”), and a plurality of registers, and can be implemented using a known computer architecture (e.g., a modified Harvard architecture, a von Neumann architecture, etc.). A “microprocessor” or “processor” refers to one or more microprocessors that can be configured to communicate in a stand-alone and / or a distributed environment, and can be configured to communicate via wired or wireless communications with other processors, where such one or more processor can be configured to operate on one or more processor-controlled devices that can be similar or different devices.

[0078] As used herein the terms “memory” and “data store” are used interchangeably and generally refer to a physical and / or logical entity that can store data, e.g., any memory storageof a computer and is a non-transitory computer readable medium. A data store may be, for example, a database, a table, a file, a list, a queue, a heap, a memory, a register, and so on. A data store may reside in one logical and / or physical entity and / or may be distributed between multiple logical and / or physical entities. A memory device may include a random-access memory (RAM), a read-only memory (ROM), an internal or external data storage medium (e.g., hard disk drive). Memory may be “permanent memory” (i.e. memory that is not erased by termination of the electrical supply to a computer or processor) or “non-permanent memory”. Computer hard-drive, CD-ROM, floppy disk, portable flash drive and DVD are all examples of permanent memory. Random Access Memory (RAM) is an example of non- permanent memory. A file in permanent memory may be editable and re- writable. The memory can include, for example, a program storage area and the data storage area. The program storage area and the data storage area can include combinations of different types of memory, such as a ROM, a RAM (e.g., DRAM, SDRAM, etc.), EEPROM, flash memory, a hard disk, a SD card, or other suitable magnetic, optical, physical, or electronic memory devices. The processing unit can be connected to the memory and execute software instructions that are capable of being stored in a RAM of the memory (e.g., during execution), a ROM of the memory (e.g., on a generally permanent bases), or another non-transitory computer readable medium such as another memory or a disc. “Memory” can include one or more processor- readable and accessible memory elements and / or components that can be internal to the processor-controlled device, external to the processor-controlled device, and can be accessed via a wired or wireless network. Software included in the implementation of the methods disclosed herein can be stored in the memory. The software includes, for example, firmware, one or more applications, program data, filters, rules, one or more program modules, and other executable instructions. For example, the computer can be configured to retrieve from the memory and execute, among other things, instructions related to the processes and methods described herein.

[0079] The terms “bit depth” and “color depth” are used interchangeably and refer, as used herein, to the number of bits used to represent each pixel in an image. The terms are used to represent bits per pixel and, at other times including e.g., when an image has multiple color channels, the total number of bits used multiplied by the number of total channels of an image. For example, a typical color image using 8 bits per channel is often referred to as a 24-bit color image (8 bitsx3 channels). Color scanners and color digital cameras may produce images at a variety of bit depths, including but not limited to e.g., 24-bit (8 bitsx3 channels) images, 36-bit(12 bitsx3 channels), 48-bit (16 bitx3 channels) images, etc. Grayscale image capture devices may also produce images at a variety of bit depths, though only in one channel, including but not limited to e.g., 1-bit (monochrome), 2-bit, 3-bit, 4-bit, 5-bit, 6-bit, 7-bit, 8-bit, 10-bit, 12- bit, 14-bit, 16-bit, etc. Individual grayscale images may be combined to generate a multichannel or color image where the resulting bit depth will depend on the bit depth of the individual grayscale images.

[0080] As used herein, the term “number” shall mean one or an integer greater than one (e.g., a plurality).2. Methods

[0081] Embodiments of the present disclosure include methods for collecting, processing, storing, and managing images and image results. In particular, the present disclosure provides methods to classify (e.g., define) and / or segment (e.g., distinguish) a target class (e.g., a cellular and / or subcellular feature of an object (e.g., an ovarian follicle)) in an image. In some embodiments, the technology provides methods to classify (e.g., define) and / or segment (e.g., distinguish) a plurality of target classes (e.g., a cellular and / or subcellular feature of an object (e.g., an ovarian follicle)) in an image when they are crowded together and / or overlapping). In some embodiments, the technology provides methods to analyze (e.g., quantify) target classes in images of (e.g., representing) tissues and / or cells. In some embodiments, the technology finds use in validating the safety and quality (e.g. anti-Mullerian hormone concentration, follicle density, number of follicles, morphological ovarian changes associated with age or disease, and developmental stage of follicles) of an object. In some embodiments, artificial intelligence machine learning is used to classify, segment, and / or analyze ovarian follicles from images of human donor ovarian tissue.

[0082] Embodiments of the present disclosure provide a method comprising: a) imaging an object to provide an imaged object; b) annotating a one or more features of the imaged object; c) measuring and counting the features of the imaged object; and d) providing a density of the feature (e.g., a follicle density).

[0083] In some embodiments, the object is produced by a method comprising a) removing the outermost layer of an object (e.g., decertifying an ovary (e.g., an ovary containing ovarian follicles in various stages of development and wherein each follicle contains an oocyte)) to provide a histological sample (e.g., a thin slice of object that is viewed under a microscope); b) fresh-fixing (e.g., fixing with formaldehyde containing fixatives) the histological sample to provide a fresh-fixed sample; c) washing the fresh-fixed sample in water (e.g., deionized water)to provide a washed sample; d) dehydrating the washed sample to provide a dehydrated sample; e) orienting the dehydrated sample perpendicular to the object surface (e.g., ovarian cortex surface) and embedding the dehydrated sample in paraffin to provide a paraffined sample; f) serially sectioning the paraffined sample to provide at least four sectioned samples; g) adding the at least four sectioned samples to a slide; and h) staining the slide.

[0084] In some embodiments, the object is obtained from a biobank (e.g., an archive of materials comprising preserved donated tissues from humans or animals).

[0085] In some embodiments, the histological sample comprises or is fresh tissue and / or cells on ice, formaldehyde fixed tissue and / or cells (e.g., formaldehyde fixed tissue in 30% sucrose), snap-frozen tissue and / or cells (e.g., frozen tissue in liquid nitrogen within 30 to 60 minutes of excision and then stored at -80 degrees Celsius), or optimal cutting temperature (OCT) embedded frozen blocks (e.g. OCT compound used to embed fresh tissue for frozen sectioning).

[0086] In some embodiments, the histological sample is fixed with formaldehyde containing fixatives and cryoprotected with 30% sucrose in 1 xPBS. In some embodiments, the histological sample is an immunohistochemistry (1HC) sample (i.e., a sample that comprises fluorescently labeled antibodies, or antibodies tagged by other means, that specifically bind to certain proteins in a cell to identify their presence in the cell and to determine whether the locations and amounts of these proteins are different in diseased vs healthy tissues).

[0087] In some embodiments, the methods of the present disclosure utilized ovarian tissue or ovarian-associated tissue. The present disclosure is not limited to particular ovarian or ovarian associated tissues. Examples include but are not limited to, a whole ovarian crosssection, human ovarian cortex strips, corpus luteums, corpus albicans, blood vessels, lymph vessels, or stromal cells.

[0088] In some embodiments, imaging comprises: a) whole slide imaging the object with a high-resolution scanner (e.g., any scanner capable of providing a high-resolution image) to provide a whole slide image file; b) importing the whole slide image file into a first software.

[0089] In some embodiments, the first software is an image analysis tool. In some embodiments, the first software is an automated image analysis tool. In some embodiments, the first software is capable of classifying (e.g., defining), segmenting (e.g., distinguishing target classes (e.g., cellular and / or subcellular features of an object (e.g., an ovarian follicle)) when they are crowded together and / or overlapping), and / or analyzing target classes in images. In some embodiments, the first software is capable of feature processing. In someembodiments, the first software is an open-source image analysis tool (e.g., ilastik software (e.g., the current ilastik version (1.0 and newer)(e.g., Stuart Berg et al., Ilastik: Interactive Machine Learning for (Bio)Image Analysis, 16 Nature Methods 1226-32 (2019), incorporated herein by reference))). In some embodiments, the first software is a trainable Al image analysis tool.

[0090] In some embodiments, the first software is used to train for automated classifying, segmenting, and / or analyzing target classes in images (e.g., defining, distinguishing, and / or analyzing an ovarian follicle) using a protocol (e.g., Pixel Classification project mode). In some embodiments, the first software is used for feature processing. In some embodiments, a whole slide image is taken in a first format (e.g., .SVS) and is then converted to a second format (e.g., .TIF) to maintain full resolution, to maintain other data, and / or to enable later use.

[0091] In some embodiments, annotating one or more feature of the imaged object comprises: a) selecting an object section region of interest from the whole slide image file using the first software; b) generating an object section number: i) for the object section region of interest; and ii) related to a feature depth and position; c) annotating the object section number to provide an annotated object section: i) by color coding; ii) by positional location fidelity; and iii) by classifying the one or more feature; and d) exporting the annotated object section from the first software to a second software.

[0092] In some embodiments, the producing convolution and pooling layers comprises training images (e.g., by drawing labels onto the image) and using a tool (e.g., suggest features tool) to predict a feature vs not a feature.

[0093] In some embodiments, representative labels for each label class are created for a feature and background using a paint-like brush tool to differentiate objects. In some embodiments, a protocol (e.g., Live Update (e.g., an application for the MSI® system to scan and download the latest drivers, BIOS, and utilities)) is selected to enable (i.e., train) a software (e.g., an Al software) to accurately predict features. In some embodiments, one or more protocols (e.g., Suggest Features) are used to enable (i.e., train) a software (e.g., an Al software) to select an assortment of sigma for each feature, to detect a labeled feature, to detect a feature depth from an object surface area (e.g., combining the outline and centroid points to obtain a distance map and using the Exact Euclidean Distance Transform in 3D to calculate the shortest distance between each point and the cortical surface) to produce a feature class discrimination, and to produce a segmentation of feature clusters. In some embodiments, protocols (e.g., for predicting features, for labeling features, for selecting sigma for each feature, for detectinglabeled features, for producing accurate segmentation) are created and used to optimize Al training. In some embodiments, a protocol is used to enable (i.e., train) a software (e.g., Ilastik) to differentiate a feature from the rest of an image.

[0094] In some embodiments, the feature is manually annotated. In some embodiments, the feature is semi-automated annotated. In some embodiments, the feature is annotated using Al.

[0095] In some embodiments, Al is used for interactive image classification, segmentation, and analysis of a feature. In some embodiments, Al is used for feature processing.

[0096] In some embodiments, AT is a modular software framework with one or more protocols for automated (e.g., supervised) pixel- and feature-level classification, for automated and semi-automated feature tracking, for semi- automated segmentation, and for feature counting without detection.

[0097] In some embodiments, Al analysis protocols enable targeted interactive processing (e.g., human assisted instruction during processing) of data subvolumes (e.g., part of filesystem with its own independent file / directory hierarchy and inode number namespace) and complete volume analysis in offline (e.g., not available on or performed using the internet or other computer network) batch mode (e.g., a method of performing a task automatically (e.g., without human assistance or intervention)). In some embodiments, measuring and counting the one or more features of the imaged object comprises: a) measuring, using the second software, (i) a distance from the cortex to the centroid of the feature; and (ii) an object section area; and b) counting, using the second software, the one or more feature.

[0098] In some embodiments, the second software is a measuring and counting tool. In some embodiments, the second software is an automated measuring and counting tool. In some embodiments, the second software is capable of measuring and counting target classes (e.g., cellular and / or subcellular features of an object (e.g., an ovarian follicle)) in 2D and 3D images. In some embodiments, the second software is used to measure and count target classes in images (e.g., measuring and counting an ovarian follicle) using a protocol. In some embodiments, the second software is capable of feature measuring and counting. In some embodiments, the second software is an open-source measuring and counting tool (e.g., Fiji software (e.g., the current Fiji version (e.g., v2.13.1 and newer))(e.g., an image processing package - a “batteries-included” distribution of ImageJ (e.g., an open source software for processing and analyzing scientific images), bundling many plugins which facilitate scientific image analysis)) (e.g., Johannes Schindelin et al., Fiji: an Open-Source Platform for Biological Image Analysis, 9 Nature Methods 676-682 (2012))).

[0099] In some embodiments, the annotation coordinates from a manual or an automated feature count are determined from a retained pixel distance and a micrometer relationship in the second software using a custom script to find the centroid of the feature.

[0100] In some embodiments, feature density is used to predict the quality of an object (e.g., the hormonal quality of an ovary). In some embodiments, feature density is associated with treatment of a deficiency (e.g., a high ovarian follicle density produces quality hormones for an ovarian tissue graft and encapsulating such an ovarian tissue graft in an immune-isolating capsule to protect the graft from rejection is a treatment for a female endocrine deficiency).3. Systems

[0101] Embodiments of the technology include systems comprising hardware (e.g., physical components of a system (e.g., keyboards, mice, pens, disk drives, iPads, printers, and flash drives, computer chips, motherboards, and internal memory chips)), software (e.g., instructions that provide instructions to hardware (e.g., operating system software (e.g., Apple iOS, Microsoft Windows, macOS, Google Android, Linux-based OS), application software (e.g., tools (e.g., word processing, form designer, spreadsheet, email, calculator, Adobe Photoshop, Microsoft Excel, Google Map), processes, humans, data (e.g., non-disputable raw facts or results of scientific observation or measurement), and network communications)))) for collecting, processing, storing, and managing images and image results. In particular, the present technology provides systems that perform methods to classify (e.g., define), segment (e.g., distinguish a target class (e.g., a cellular and / or subcellular feature of an object (e.g., an ovarian follicle)) when they are crowded together and / or overlapping), and / or analyze (e.g., quantify) target classes in images of (e.g., representing) tissues and / or cells and, in some embodiments, to validate the safety and quality (e.g. anti-Mullerian hormone concentration, follicle density, number of follicles, morphological ovarian changes associated with age or disease, and developmental stage of follicles) of the objects. In some embodiments, artificial intelligence machine learning is used to classify, segment, and / or analyze ovarian follicles from images of human donor ovarian tissue.

[0102] Embodiments of the present disclosure include a system comprising: a) first hardware (e.g., a one or more imaging device); b) one or more network communication component (e.g., an input device and / or an output device); c) a second hardware (e.g., one or more memory component); d) a first software (e.g., a one or more application software); e) a second software (e.g., a one or more application software).

[0103] In some embodiments, the system is configured to analyze data from a first software and a second software. In some embodiments, the system is configured to detect a density of a feature in an imaged object (e.g., an ovarian follicle) and perform actions on the imaged object.

[0104] In some embodiments, a one or more software comprises one or more protocols for: (i) imaging a processed object to provide an imaged object; (ii) annotating a one or more feature of the imaged object; (iii) measuring and counting the one or more feature of the imaged object; and (iv) providing a feature density of the object. In some embodiments, a first software alters protocols based on responses from a second software.4. Examples

[0105] It will be readily apparent to those skilled in the art that other suitable modifications and adaptations of the methods of the present disclosure described herein are readily applicable and appreciable, and may be made using suitable equivalents without departing from the scope of the present disclosure or the aspects and embodiments disclosed herein. Having now described the present disclosure in detail, the same will be more clearly understood by reference to the following examples, which are merely intended only to illustrate some aspects and embodiments of the disclosure, and should not be viewed as limiting to the scope of the disclosure. The disclosures of all journal references, U.S. patents, and publications referred to herein are hereby incorporated by reference in their entireties.

[0106] The present disclosure has multiple aspects, illustrated by the following non-limiting examples.SOFTWARE USED

[0107] Materials. Ilastik vl.4.6 software was used to process images using trained Al. Fiji v2.13.1 software (e.g., Schindelin, supra) was used to convert images to HDF5 file format and then further process the images and provide total follicle count. The Ilastik Plugin e.g., Berg, supra) for Fiji was used for import and export of data between Ilastik and Fiji. QuPath vO.2.2 software (Peter Bankhead et al., QuPath: Open Source Software for Digital Pathology Image Analysis, 7 Scientific Reports 16878 (2017), incorporated herein by reference) was used to obtain T1F file format for images of individual tissue sections.Example 1

[0108] Analysis of follicle density and distribution in human ovarian tissue using whole slide scans and a semi- automated image processing flow. Due to the heterogenous nature of ovarian tissue, with regard to follicle location, density, and follicle growth stage, there exists inherent subjectivity in identifying, classifying, and quantifying follicles. To test how thisaffects FD measurements, a single independent operator (also referred to as the comparison operator for the user- variation study) analyzed at least one ovarian tissue sample from 12 deceased human donors. Additionally, follicle counts were determined from 4 of the 12 donors using multiple tissue samples from different locations on either the same ovary or from the contralateral ovary to assess local and global heterogeneity, thus providing a grand total of 102 slides (and histologic sections) analyzed.

[0109] Because follicular growth exists on a spectrum and because definitions are evolving to increase the accuracy of capturing these stages, follicle classification remains a high-skill task. Accordingly, during the development of embodiments of the technology described herein experiments were performed to evaluate the variability of classification between two trained operators. In total, 99 of the same slides (and histologic sections) prepared from 17 ovarian tissue samples were reviewed and classified by two independent operators. The findings from the 102 slides the comparison operator analyzed were also characterized. FIG. 1 illustrates a workflow diagram of the key steps in the process of follicle density and distribution analysis.

[0110] Methods. Tissue selection, preparation, and imaging. Donor human ovarian tissue procured by the International Institute for the Advancement of Medicine (11AM) was delivered, then harvested, and the ovarian cortex was decertified by slicing 1-cm2squares of tissue that were ~0.7- 1.0 mm thick. Histological samples for the study were then obtained by cutting either 1 mm x 10 mm x 0.7-1.0 mm strips or 4 mm x 4 mm x 0.7-1.0 mm squares. These tissue samples (squares or strip) were fresh-fixed in Bouin’s solution, washed in dH2O, dehydrated in 50% ethanol for one hour, followed by 70% ethanol for one hour, oriented perpendicular to the cortex surface, embedded in paraffin, and serially sectioned (5 pm thick) maintaining order, with four sections per slide, spanning a total 1 mm of tissue. Every other slide was stained with hematoxylin and eosin (H&E) and whole slide imaged with a high-resolution scanner (Leica Aperio AT2, Leica Biosystems, Buffalo Grove, IL) at 20x, to produce 0.5034 microns / pixel resolution .SVS image files (FIG. 2).

[0111] QuPath. Six slides containing four tissue sections each (selecting every eighth section to be spaced 40 pm apart), were chosen for assessment. The slides were selected based on the following qualitative and semi-quantitative criteria: (i) sections were chosen that had been prepared from roughly halfway (0.5 mm) through sampled tissues; and (ii) sections were chosen that had low debris, minimal image distortion, and minimal presence of other artifacts throughout all six slides. The selected whole slide image files (e.g., in .SVS format) were imported into a ‘project’ file in the open-source image analysis software QuPath vO.2.2(Bankhead, P. (2017) “QuPath: Open source software for digital pathology image analysis” Scientific Reports 7, Article number: 16878). Custom QuPath scripts, which are programs that are either recorded and / or coded consisting of sequences of actions, were used to automate the selection of each individual tissue section via a region of interest allowing each section to be named and saved sequentially as individual lossless .TIF files. Once slides were screened for low artifacts, and those with a high number of imperfections were eliminated from potential evaluation, a random number generator was used to select a section number (indicating section depth and position) for the set of four single tissue sections per slide. The selected sections were manually annotated, following the pre-antral ovarian follicle classification definitions listed in FIG. 3A, using QuPath tools (a paint-like brush) and the follicles were classified by the user as shown in FIG. 3B (note that outdated follicle classifications are shown in FIG. 3B and that updated follicle classifications are shown in FIG. 3C). The annotated follicle regions in each tissue section, color coded by classification and maintaining positional location fidelity, were exported as binary colorized 8-bit .PNG image files via script as shown in FIG. 3E. Additional QuPath measurements for the .TIF image sections (capturing spatial metadata) were exported to a .CSV file using QuPath feature; the exported data included a slide / section identifier and data describing follicle classification, centroid, area, and perimeter.

[0112] ImageJ / Fiji. Open-source image processing software, Fiji / Image J (Schindelin, supra)) was used to obtain further measurements, specifically the distance from the cortex to the centroid of the follicle, and the tissue section area via a collection of publicly available and custom macros. The custom macros to obtain the tangential distance from the follicle centroid to the surface of the cortex was implemented by: (i) thresholding the image; (ii) creating a region of interest (ROI) and associated outline of the tissue; (iii) converting to a greyscale image; (iv) processing the greyscale image by a Gaussian Blur filter with a smoothing factor of 20 pixels; and (v) thresholding the image again and converting the image into a mask with all holes filled (FIGS. 4A-4C).

[0113] Analyze Particles, a built-in ImageJ functionality, was used and objects with a circularity of 0-0.6 were identified. This allowed for any identified features to be deleted in the ROI Manager (where all identified Regions of Interest were collated and selected for viewing or processing individually) if they were not tissue, such as artifacts. The macro prompted the user to use the pencil tool to erase a few pixels at the edge of the cortex. The other side of the ovarian tissue strip is the medulla, or the interior of the ovary. Once the cuts were made, three outlines were created (FIG. 4B). Only the cortex outline was retained, the medulla and wholeoutlines were deleted by the user. Next, the centroid of each annotation was calculated. The macro prompted the user to open the exported .PNG file from QuPath containing the colorized follicle annotations. The image was converted to a greyscale image and thresholded again as shown in FIG. 5A, with a range of 0-223 to locate the centroid by assigning high values to the pixels nearest the centroid and low values to the pixels furthest away in each annotation. The highest values were visualized as the white spots, with a gradient to gray or black.

[0114] Then, Find Maxima was used with a value of 10 and the output set to Point Selection, which turned the highest values (the white pixels) into points. The outline and centroid points were combined to obtain a distance map shown in FIG. 5B and the Exact Euclidean Distance Transform in 3D calculated the shortest distance between each point and the cortex.

[0115] Microsoft Access. Finally, the Image J follicle depth measures and QuPath exported measurements and calculations were imported into a Microsoft Access database where the individual follicle information was matched together by the filename and the follicle centroid (x , y coordinates) in microns. This allowed automated matching of follicle depth from cortex surface to the follicle classification. The customized MS Access database allowed all collected data to be viewed in a single Microsoft Excel spreadsheet, with three reports generated and saved in a specified location.

[0116] The first report, QryCompare, provided detailed information and a summary table of the follicle data per slide analyzed. It included the following information: total # of follicles in database, follicle class, number of follicles in that class, original filename, serial number (of the imported file), image name, region of interest, label, micron and pixel measurements (including the linear distance from the ovarian surface to the follicle centroid, and the centroid x, y coordinates), QuPath and Fiji Keys (used to match the file information together). The second worksheet of the QryCompare report is the QryCompare_Crosstab which provided the following information per slide: original file name, total number of follicles, and the total number of each follicle class.

[0117] The second report, Mismatch Summary provided information on slides that were unable to be paired. The filename, key x and y, and Fiji key x and y in the QryMismatches worksheet. The Percent Mismatches provided the original file name of slide, total number of matched annotations, total number of mismatched annotations, and the mismatched annotations as a percentage of total annotations. The final report, Match Summary, comprised all the successfully paired Fiji and QuPath files. The first worksheet is titled QryCompare and includes all the information found in the first report as well as an updated classification wherethe user can choose to pool QuPath labels together, or group in whatever way they choose. The second worksheet, MatchSummary, summarized the information from QryCompare by providing the original filename and the follicle count matched from that filename.

[0118] Statistical Analysis. Statistical analysis for operator comparison was performed in GraphPad Prism 7 using a Wilcoxon matched-pairs signed rank test of the QuPath output for average primordial FD per section. Mean and SEM with significant p<0.05 graphed. Mann- Whitney tests were performed for Donor 17, Donor 19, and Donor 23 average primordial FD per section and overall follicular depth. A one-way ANOVA was performed on Donor 23 via Kruskal-Wallis test and Dunn’s multiple comparison test for the same measurements. Mean and SEM with significant p<0.05 graphed.

[0119] Results and Discussion. Two different operators manually counted and analyzed ovarian tissue from twelve deceased human donors, each donor referred to as their Lab Block ID Number. The data from multiple sections from different locations on the same ovary were analyzed as pooled data (as in donors 17, 19, 20, and 23). The grand total of analysis was 102 slides from 17 ovarian tissue section samples by both operators. The attributes of the deceased human ovarian tissue donors such as age and BMI can be found in Table 2. The sampling scheme was included only six sections due to the time-consuming nature of follicle counting.

[0120] Table 2. Deceased human donor tissue attributes.

[0121] Operator Interactions. There are limitations to classifying follicles such as subjective differences between operators and lack of clearly defined states, which can be exacerbated by sub-optimal tissue fixation, sectioning artifacts, and staining inconsistencies. Follicle classification is especially challenging when follicles are in transitional states, are cross sectioned off-center, or are not clearly visible due to debris or poor histologic processing. These transitional states are poorly documented and defined. Due to the lack of a variety examples, operator training is difficult and classification is subjective. Some operators are more conservative than others in terms of the actual class of a follicle or if it should be counted as a follicle at all. The differences between healthy / normal and unhealthy / abnormal follicles are another added complexity that is subjective because (like the follicle class) it is difficult to discern what is abnormal morphology versus histologic processing artifacts.

[0122] To address some of the subjectivity between follicle classes, an updated follicle classification system (FIG. 1C) was used for donors 6, 17, 26, and 31 (n=24 sections). The outdated classification system (FIG. IB) classified follicles by primordial, primary, secondary, multilayer, and antral follicles (all with and without nucleus labels). The remaining donors were characterized by the control operator using the outdated classification system and by the comparison operator using the updated classification system. A label of “follicle” was included for structures that were identifiable as a follicle, but not clear as to which class they belong (e.g., typically, this was for follicles where the plane of section was highly off-center). The updated classification system included the categories of the outdated classification system as well as the additional transitional primordial, transitional primary, atretic antral, and abnormal (or unhealthy) follicle labels (all with and without nucleus). The updated classification system allowed for finer granularity in classifying the follicular growth state and added the capability to tag follicles which are abnormal / unhealthy.

[0123] The goal of the study was to determine if there were significant differences between two operators counting and classifying follicles. The “comparison” operator was trained and had been classifying follicles for approximately 1 year, and the “control” operator was trained and had been classifying follicles for approximately 3 years. It was expected that there would be slight differences between follicle classifications, but no discrepancies in overall follicle count. FIG. 6 depicts the overall follicle count per slide by each operator.

[0124] The calculated relative error (ABS(comparison value -control value) / control value) per slide was 43% between operators. Relative Accuracy of manual follicle counting was calculated by subtracting the absolute value of the per slide error value from one. Manual follicle count accuracy was 66.9% accurate when compared to the control operator counts.

[0125] The operators displayed significantly different follicle counts per section analyzed, with the comparison operator’s follicle count mean of 33.09 follicles / section and the control operator’s mean of 19.64 follicles / section, the comparison operator’s standard error of mean (6.62) was also almost double the control operator’s (3.15). Overall, the comparison operator and the control operator followed a similar trend in follicle counts. The control operator was generally more conservative in counting than the comparison operator.

[0126] Discussion. The main limitation of classifying follicles by hand is that it is time consuming. It is expected that operator variation is small enough to reproduce the same results, but it is accepted that this work has such discrepancies. The inherent heterogeneity of the ovarian follicles in the cortex also contributes to high or low follicle counts in the analyzed sections compared to the rest of the ovarian cortex.

[0127] The results (e.g., follicle counts per section analyzed) of operator interactions were found to be significantly different, as shown in FIG. 6. This is, in part, due to the limitations of follicle classification, such as subjective differences between operators (e.g., inclusion of follicles cross sectioned near an edge vs. exclusion of those follicles) and a lack of clearly defined states of follicle development, and is exacerbated by sub-optimal tissue fixation, sectioning artifacts, and staining inconsistencies. Follicle classification is especially challenging when follicles are in transitional states, are cross sectioned off-center, or are not clearly visible due to debris or poor histologic processing. These transitional states are poorly documented and defined. Operator training is also difficult, due to the lack of a variety examples and the subjectivity of classification (e.g., some operators may be more conservative than others when labeling a follicle to a class or whether a specific structure qualifies to be counted as a follicle at all). The differences between healthy / normal and unhealthy / abnormal follicles add further subjective complexity because, like the follicle class, it can be difficult to discern what is abnormal morphology versus what is a histologic processing artifact (which is also poorly documented). The results underscore a need for a robust training aid or alternative method of counting follicles.

[0128] The operator variations were statistically significant result, resulting in an average per-slide error rate of 43% and a relative accuracy of 66.9% when compared to the controloperator counts. These results underscore a need for a robust training aid or alternative method of counting follicles.Example 2

[0129] Investigation of the use of Al for automated, high-throughput ovarian follicle counting. Because there is demonstrated manual operator variation between the control and comparison operators, a validated automatic system for follicle counts was sought. Accordingly, during the development of embodiments of the technology described herein experiments were conducted to demonstrate the use of Al software to detect and accurately classify human ovarian follicles. Implementing well-trained Al, instead of utilizing a semiautomated workflow, to count and annotate follicles resulted in more efficient resource allocations and a higher level of accuracy.

[0130] Methods. TIF images of ovarian tissue sections were imported into the open-source trainable Al software Ilastik vl.4.6 (Berg (2019) “ilastik: interactive machine learning for (bio)image analysis” Nature Methods 16: 1226-1232) to train for automated follicle identification using the Pixel Classification project mode. Features were set at the largest sigma Equation (6) to begin with for Color / lntensity, Edge, and Texture. Labels were created of Follicle, Tissue, and Background using a paint-like brush tool to differentiate objects as shown in FIG. 7B. After marking a few areas representative for each label class, Live Update was selected to ensure the prediction of follicles as seen in FIG. 7C.

[0131] Suggest Features were then utilized to allow the software to select the most efficient assortment of sigma for each feature to have the most robust probability of detecting the labeled structure accurately (FIG. 7C) to be able to produce an accurate segmentation (FIG. 7D) in the least amount of computing time. The process of labeling additional features, determining if the training is being optimized via Live Update, utilizing Suggest Features, and importing additional images to train the program further is iterative and ongoing. The Ilastik classifier was trained to identify follicles versus the rest of the image (tissue and background) on seven representative whole slide scanned images. Batch Processing was used to employ the trained classifier on the same 102 images manually annotated to determine Al accuracy. The files were then exported in .TIF format for further processing in Fiji / ImageJ.

[0132] Workflow Integration. An ImageJ macro was created to import the files directly from a folder and an output was specified. Batch mode was turned used to process .TIF Ilastik export files in bulk. A glasbey lookup table (FIG. 8 A) and gaussian blur with sigma of 15 was employed. The image was then thresholded to black and white and converted into a mask, asshown in FIG. 8B. A Watershed filter was used to separate clustered follicles and Analyze Particles, with settings set to: size 1000-1000000, circularity 0.75 - 1.00, was used to display the summary of results during and after processing the batch (FIG. 8C). The results were then saved as a .CSV file per image analyzed as well as an executive summary of all images in the folder selected. This macro was based on the macro of the semi-automated process.

[0133] A new ImageJ macro was created to process and count the follicles from the exported Ilastik .TIF files. For follicle depth assessment, further Ilastik training to identify the cortex versus the medulla sides of the tissue strips is done to replicate the existing ImageJ workflow exactly and can be entirely automated. Currently, the operator manually selects the surface cortex in the semi-automated workflow (FIG. 4B).

[0134] Statistical Analysis. Using Graphpad Prism, a Wilcoxon matched pairs signed rank test of the follicle counts by the control operator and the Al was performed. A one-way ANOVA was performed on the follicle counts from the control operator, the comparison operator, and the Al via Friedman test and Dunn’s multiple comparison test. The error of follicle counts per section was calculated and averaged to obtain the average error per section. The relative accuracy of follicle counts per section was calculated as well, excluding sections with zero counts by the control operator. The time for manual follicle counting was estimated and the computing time for Al was exact.

[0135] Results and Discussion. With 102 sections to analyze, estimating 15 minutes per section scanning for follicles and exporting data and an additional 1 minute to annotate per follicle, the manual annotations took an estimated 89.7 hours in total, for an average time of 5.3 hours per tissue sample (set of SIX sections), 53 minutes per sections. Once trained, the analysis took approximately an hour to select inputs for the software used and to sort the export for the entire dataset. Ilastik was left to run overnight to export the whole dataset (5.5 hours in total, averaging 20 minutes per sample set). Running the Fiji macros for all samples took approximately 45 minutes and could be automated to run faster and batch all images together, but time was taken to ensure each sample set was outputting correctly. In the amount of time it took one operator to label and export data for 6 sections, all 102 sections were made into a binary mask and data was exported by automated software. As discussed, annotating just 6 sections manually involves providing a large commitment of labor. For a process which is reproduced with similar levels of accuracy, the Al quantified follicles in a quarter of the time it took an operator. The computational work was carried out on a 2022 MacBook Air with Apple M2, 8-core CPU, 8-core GPU, 8GB of RAM, and 256GB SSD. If a machine with moreRAM (and cores) or CPU capabilities was utilized (such as a supercomputer) the processing time could be shortened.

[0136] The Al is comparable to a human operator in identifying follicles. FIG. 9 was created in Fiji by assigning the red channel to the original .TIF file and assigning green to the Ilastik output after being processed through a glasbey lookup table.

[0137] This output is strikingly similar to that in FIG. 1C of the manually tagged follicles. Some clusters of follicles were not separated well and some follicles (particularly those with nuclei) were not fully captured when filtering in Fiji and were, thus, sometimes missed in the count. There are also small spots the software assigned as follicle erroneously and a small number made it past image fdtering and processing in this case.

[0138] The error of Al follicle counting was found to be 51.2% and accuracy was calculated to be 60% when compared to the control operator counts.

[0139] The results from the Wilcoxon matched-pairs signed rank test show a significant difference, but the distribution visually is similar. The average of the control operator follicle counts is 16.71 follicles / section with a SEM of 3.11, and the Al average follicle count is 12.94 follicles / section with a SEM of 2.21. To further test this, a one-way ANOVA was performed on all three follicle counting methods and significance was found between the comparison operator and the Al, as well as the comparison operator and the control operator, but none between the control operator and the Al. The comparison operator had a mean (32.43 follicles / section with SEM 7.04) of almost double the control operator (16.71 follicles / section with SEM 3.11).

[0140] As previously noted herein, use of an automated follicle count method results in four times faster processing than a manual follicle count method. In other words, in the amount of time it takes one operator to manually label and export data for six sections, 102 sections can be made into a binary mask and data exported, with similar levels of accuracy, by automated software. Thus, Al can reproduce the manual process in a quarter of the time (i.e., four times faster processing than a manual follicle count). Further, the time needed to complete the process shortens dramatically when a machine with more RAM (and cores) or CPU capabilities is used (such as a supercomputer).

[0141] A residual plot was produced of the Al, control operator, and comparison operator methods (FIGS. 10A-C). The Al slightly undercounted when compared to the control operator (x-axis) and the comparison operator tended to overcount when compared to the control operator, sometimes by as much as 200 follicles. In contrast, the largest discrepancy for the Alwas approximately 50 follicles. All methods were close in follicle count until around 50 follicles in any given slide, and then the results appeared subjective. Al performed better at identifying follicles than an operator trained and practicing for one year.

[0142] Conclusions. The expectation of a trained Al to identify and count follicles, in a fully automated manner, was met. The accuracy of the Al was comparable to a human operator and was four times as fast. The open-source software used enabled continuous community-wide improvement and high accessibility.

[0143] Al could be improved to increase accuracy (e.g., training the classifier on more images, particularly those with less than two follicles, to avoid overcounting, and those with more than fifty follicles, to avoid undercounting). Ilastik has an Auto-context feature that may enable only primordial follicles to be identified, lending more use for qualifying tissue. The unfdtered output from Ilastik can identify follicles, but a more robust filter would enable follicle clusters to be separated and for smaller misidentified objects to be filtered out. Training the Al to include the depth from the ovarian surface and follicular class discrimination and then modifying the Fiji macro to incorporate this input would be useful (e.g., it would enable a fully automated method of counting and classifying human ovarian follicles -with more accuracy and more speed than a human operator).EX MPLE 3

[0144] Follicle identification is a challenge for the field. While single cell structures can be captured by open-source (e.g., Fiji, Cell Profiler, QuPath) and paid image processing software, capture of multi-cell structures (e.g., ovarian follicles) is difficult due to the inherent variable composition of the structures. For example, QuPath and Cell Profiler Al can identify single cells, but identifying multicellular elements (e.g., follicles) is inaccurate and / or incorrect. Orbit Image Analysis needs significant computing power. Napari is difficult to implement. Tensor Flow is associated with a high computing power (NVIDIA GPU card with CUDA Compute Capability 3.5 or higher. 128 GB of RAM.) and is difficult to use.

[0145] In a first study, Ilastik Al was trained in two different training sessions first using 5 .TIF training images and subsequently using 10 .TIF training images. It was noted that training the classifier in this manner resulted in slow program performance due to large computing power requirements. Three labels were used in this initial classification trial: (i) follicle; (ii) tissue; and (iii) background. It was assumed that three labels would be adequate to obtain the follicle distance to the cortex surface. However, three labels decreased the Al speed. Two labels were subsequently used and found to be sufficient. During classifying of images, Ilastik Al wasnot capable of correctly classifying voids in the antrum of late-stage follicles and voids in less dense tissue, particularly around the edges of the tissue (e.g., FIGS. 11A-C). Ilastik Al identified small follicles well, but not their nuclei.

[0146] In a second study, the trained classifier produced in the first study was used to train Ilastik Al using images without large follicles (FIGS. 12A-C). This performed better for correctly identifying the tissue (instead of confusing it with “follicle”). Ten images without the larger follicles were used for training input.

[0147] In a third study, the trained classifier produced in the first study was used to train Ilastik Al and three labels were used: (i) follicle; (ii) tissue; and (iii) background (FIGS. 13A- D). Twenty-five images were used to train, but, during training, the program began to slow and crash and export failed. While this training worked well, it took too many training images to reach a level of proper identification. Additionally, the nucleus was still not identified. However, the discrimination of background vs tissue was excellent.

[0148] In a fourth study, 3 images were used for training, then 5, and then 7 using .HDF5 training images and using two labels (i.e., follicle and tissue) where the tissue and background were combined. As the training progressed, the classifiers were saved to use for visualizing their performance in classifying an image. FIGS . 14A-C illustrates the three classification trials - native tissue, probability, and segmentation. Ilastik Al confused the less dense tissue as follicle, while follicle nuclei and large follicle granulosa cells were confused for tissue. Ilastik Al inaccurately identified anything with an outline of dark cells as follicle. The classifier was good with this selection of tissue initially.

[0149] In a fifth study, the same process as in the fourth study was used, except every follicle was labeled. Seven images were classified and every follicle was labeled (FIGS. 15A-C). This study did not perform well as it was skewed toward follicle labeling since there were so many as the input to the training.

[0150] In a sixth study, the same process as in the fourth study was used, except some of the follicle labels from study five were erased. Seven images were classified (FIGS. 16A-C). This appeared to be less accurate than the fourth study. However, this sixth study took tissue pieces that were different into account. The follicle identification was strong and the excess was filtered out in Fiji.

[0151] Overall, everything input into Ilastik Al produced a bias. Training the Ilastik Al classifier with images comprising as much variety as possible helped to capture a normal spread of follicles and tissue morphology that would be present in images to be classified by the trainedclassifier. Types of variety included: (i) different stage follicles; (ii) follicles with and without nuclei; (iii) the surrounding squamous or granulosa cells; (iv) unhealthy and healthy follicles; (v) clusters of follicles; (vi) debris on slides; (vii) very dense or less dense tissue; (viii) more pink or more purple stained tissue; and (viiii) voids in the tissue (this was especially difficult to train as the Al thought every circular shape was a follicle).EXAMPLE 4

[0152] Tissue Acquisition and Processing. Donor human ovarian tissue was procured by the International Institute for the Advancement of Medicine (IIAM). The ovarian cortex was sliced using a template into 0.7-1 mm thick by 1cm2squares of tissue, which were further cut either into 1 mm x 10 mm x 0.7-1 mm strips or 4 mm x 4 mm x 0.7-1 mm squares. These tissue samples were fresh- fixed in Bouin’s solution for 18-24 hours, then washed in dHO for 1 hour, dehydrated in 50% ethanol for 1 hour, followed by 70% ethanol for 1 hour, oriented perpendicular to the cortex surface, embedded in paraffin, and serially sectioned (5 pm thick) with 4 sections per slide, spanning a total 1 mm of tissue. Every other slide was stained with hematoxylin and eosin (H&E) and whole slide imaged with a high-resolution scanner (Leica Aperio AT2, Leica Biosystems, Buffalo Grove, IL) at 20x magnification to produce 0.5034 microns / pixel resolution .SVS image files (FIG. 17A). The remainder of unstained slides and sections were archived for future use.

[0153] Image Evaluation. The semi-automated method involved two independent operators each manually annotating and counting follicles, the majority of which were primordial follicles (which were identified as having one or fewer layer of squamous granulosa cells), from 93 sections, sourced from 11 donors, at a consistent 40-micron spacing. These annotations were performed by two individuals: a “control” operator and a “comparison” operator. The “comparison” operator was trained by the “control” operator and had been classifying follicles for approximately 1 year, and the “control” operator had approximately 3 years of experience. The comparison operator analyzed 102 sections from 12 donors, which was used to generate the value of primordial FD and used to compare methods of counting. The comparison operator also assessed the distance of primordial follicles from the ovarian cortex surface.

[0154] The Al assessed the same 102 sections as the comparison operator; however, due to multiple rounds of training in Ilastik, one image per sample previously labeled was excluded after the second training round to minimize or eliminate bias in the final results. Thus, the comparative analysis across both operators and the Al method was limited to the 77 sections that were common to all methods of counting. The per-section and per-donor control operatorcounts were used as a benchmark for the comparison operator and Al and to calculate percent relative error or accuracy. Lastly, to investigate true positive and false positive follicle identification and to elucidate if primordial follicles were accurately identified, the comparison operator classifications were compared to the Al follicle labels.

[0155] Calculations in the Semi- Automated Process. The primordial follicle density (pFD, reported as follicles / mm3) in each section, sample, and donor were quantified, and are reported in Table 3.rn primordial follicles per section .. .

[0156] Primordial FD = - total section area*section thickness (1)

[0157] The tissue area was calculated by Fiji macro instructions, or custom programmed series of operations, and were reported as mm2. The section thickness is 5 microns, which was used with the calculated tissue area to generate tissue volume (reported as mm3). These values are reported in Table 3.

[0158] Table 3: Findings per sample (samples contained 6 sections, lettering (a, b, c) denotes different sample locations on a single donor ovary) from “comparison” operator counts

[0159] The distance from the ovarian cortex surface to a follicle centroid was generated by a custom Fiji macro, which calculated the shortest distance between each point and the cortical surface. Only primordial follicles were assessed from the comparison operator data.

[0160] Semi-Automated Process, Manual Follicle Counts.

[0161] Manual Identification of Follicles using QuPath. The open-source image analysis software QuPath, vO.2.2 (qupath.github.io) was utilized for image analysis. The whole slide .SVS image files were imported into a Project File. Custom QuPath scripts, programs consisting of sequences of actions that were recorded and / or coded, were used to automate the selection of each individual tissue section via a region of interest (ROI), allowing each section to be named and saved as cropped .TIF files sequentially without loss of resolution (see FIG. 17A). Six slides containing 4 tissue sections from each donor tissue sample were chosen for analysis, yielding a total of 102 tissue sections analyzed. Slides were selected from roughly halfway (0.5 mm) through the sectioned samples’ depth and had low debris, minimal image distortion, and minimal or no other artifacts. After slides with significant levels of imperfections were eliminated from the analysis, one of the four sections was randomly selected as the sampling starting point with a section sampling spacing of 40 pm, or every 8th section. The spacing avoided capturing the same primordial follicle twice (because they are~30 micron in diameter) and, thus, a primordial follicle with or without a visible nucleus was counted. For accurate representation, larger follicles were captured on multiple analyzed sections and were labeled with or without nucleus. The selected sections were manually annotated using QuPath tools and the follicles were classified by the user. The follicles were classified as primordial, transitional primordial, primary, transitional primary, secondary, multilayer, or antral according to the follicle classification definitions recommended by the Follicle Classification Committee at the National Institute of Child Health and Human Development (NICHD)-sponsored Ovarian Nomenclature 2021. Follicles whose class could not be unambiguously identified - such as when granulosa cells could not be clearly visualized as squamous, cuboidal, or layered - were generally categorized as “follicle”. The annotated follicle regions in each tissue section, were color coded by classification and maintaining positional location fidelity, were exported as binary colorized 8-bit .PNG image files via a custom QuPath script. Additional QuPath measurements for the .TIF image sections project (capturing spatial metadata) were exported to a .CSV file using an intrinsic QuPath feature and included slide and section IDs, follicle classification, centroid coordinates, area, and perimeter (measured in micrometers).

[0162] Follicle and Tissue Measurements using Fiji. The open-source image processing software Fiji / Image J, v2.13.1 (imagej.net) was used to measure the tangential distance from the surface of the cortex to the follicle centroid and the tissue section area utilizing image thresholding. First, the macros thresholded the image and created an ROI and associated outline of the tissue section. It was then converted to a greyscale image (8-bit) subjected to a Gaussian Blur filter with a smoothing factor of 20 pixels. The image was thresholded again and converted into a mask with all spaces in the tissue filled. The area of the tissue was then calculated, and the user was prompted to check the ROI(s).

[0163] A built-in Fiji functionality, Analyze Particles, was employed to identify objects with a circularity of 0.0-0.6. This analysis allowed for any extraneous artifacts not associated with the primary tissue section to be deleted in the ROI Manager, in which all ROI are collated and can be processed individually. The macro prompted the user to use the pencil tool to erase a few pixels in the ROI at the edge of the cortex surface to mark its location, essentially ‘cutting’ the ROI outline into a cortex surface line and a medulla border line, while maintaining the original section surface border line. Following this, only the cortex surface outline was retained; the medulla outline and whole tissue strip outline were deleted by the user. Next, the centroid of each annotation was calculated by the macro - the macro prompted the user to openthe exported .PNG file from QuPath containing the color-coded follicle annotations. The image was converted to a greyscale image and thresholded again. Each pixel was assigned a value within the pixel intensity range of 0-223 such that higher values were assigned to the pixels nearest the centroid and low values were assigned to the pixels furthest away in each annotation. The highest values were visualized as white spots, with a gradient to gray or black: thus, each follicle’s centroid was now closest to white. Then, Find Maxima was used with a tolerance (also known as prominence) value of 10 and the output set to Point Selection, which converted the highest values (the white pixels) into points (the centroid of each labeled follicle). The prominence value dictates the necessary difference between points to register as unique local maxima and was determined here as it visually produced optimal results. The outline and centroid points were then combined to obtain a distance map and the Exact Euclidean Distance Transform in 3D was used to calculate the shortest distance between each point and the cortical surface.

[0164] Collation of QuPath and Fiji Results using Microsoft Access. The measurements thus collected semi-automatically with Image J, including follicle depth, follicle centroid coordinates, and section ID, and measurements acquired manually with QuPath, were imported into a Microsoft Access database that collated the disparate measurements. The database first converted the follicle depth and centroid data from Image J into length expressed in units of microns. Then, it performed one-to-one matching of each follicle with QuPath data based on parameters that were collected separately by both methods, such as filename and follicle centroid coordinates. The customized Access database allows all collected data to be viewed in a single Microsoft Excel spreadsheet. Reports were generated from the Access database to compile information about follicles from each tissue section and donor sample (including the total number of follicles analyzed, follicle class, number of follicles per class, linear distance from the cortex surface to the follicle centroid, x- and y-coordinates of follicle centroid, etc.) and can be filtered per section, donor sample, or donor.

[0165] Fully Automated Process.

[0166] Converting File Type in Fiji / ImageJ. The same 102 cropped single tissue section .TIF images that were created and analyzed in the semi-automated method were also analyzed via the automated method. Ilastik, vl.4.6 (www.ilastik.org) is an open-source, trainable learning and segmentation Al software that depends upon user input in the form of drawing labels and utilizes existing machine learning algorithms so that the user does not need to code. Hierarchical Data Format version 5 (.HDF5) is the preferred format for Ilastik to run effectivelybecause HDF5 supports large, heterogeneous information (such as WSIs) and retains vital information such as resolution and spatial data. One of the few manual steps taken in the automated follicle count method was to convert files from .TIF format to .HDF5, as shown in Figure IB. A macro was created using the Ilastik plugin in Fiji (www.ilastik.Org / documentation / fiji_export / plugin#installation) to provide batch processing of all .TIF files once an output folder is specified. Once converted to HDF5, the images appear to be in greyscale in the Fiji software, but retain the color in the HDF5 format, visible in Ilastik.

[0167] Training the Artificial Intelligence Software Ilastik to Identify Human Ovarian Follicles.

[0168] Ilastik Density Counting Training. Ilastik was trained in automated follicle identification using the Density Counting Workflow, which implements a supervised Random Forest Regression algorithm. Seventeen representative .TIF images of ovarian tissue sections were imported as Raw Data, or training images, under the Input Data applet. The seventeen training images were from each sample subset (donors with multiple samples had images from all samples trained on), and visually spanned the range of pink to purple, had extreme or average follicle numbers as compared to our donor population (over 100 follicles to no follicles), and had varying levels of vasculature or voids in the tissue. The seventeen images were not manually counted or analyzed by the trained Al program so as to not bias the results of applying the trained Al program on the to-be processed 102 sections. In the Feature Selection applet, features were initially set at the largest sigma (the amount of information considered at once for evaluation, visualized by circle diameter size) for Color / Intensity, Edge, and Texture feature types.

[0169] To use the interactive counting feature under the Counting applet, dots of “Foreground” label were placed in the center of a few follicles. Each dot places a local gaussian blur to a chosen sigma value on the image. To change the amount of pixels that are considered into the “Foreground”, different sigma values may be used. After placing about 10 dots, to select the appropriate sigma, Live Update was turned on to see in real-time what area the sigma value was encompassing. A sigma of 16 was implemented to just touch the edges of primordial follicles, which was the follicle class of interest. Brush strokes were annotated on to the tissue stroma and background slide and labeled “Background”. Once additional labels were marked on all 17 training images, Suggest Features was used to allow the program to select the most efficient assortment of sigma values that maximizes the probability of detecting the labeled structure (Figure 1C). Choosing the most computationally efficient assortment, a GaussianSmoothing with a sigma of 10, Laplacian of Gaussian with a sigma value of 10, and Gaussian Gradient of Magnitude with sigma values of 5 and 10 were set. The export settings were chosen in the Density Export applet, to be probabilities in .TIF format, with the range normalized to 0-255 intensity values assigned to each pixel. The Random Forest Regression algorithm calculates the probability of a pixel being a Foreground or Background label. A pixel value of 255 correlates to 100% calculated certainty that it is Foreground, while a value of 0 corresponds to 100% Background. The process of labeling additional features, visually determining trained Al performance via Live Update, utilizing Suggest Features, and training on images to further improve identification was iterated until satisfactory performance was achieved, balancing accurate annotations with minimal computing time usage. Finally, for the Density Counting workflow, the unseen set of 102 HDF5 images were imported as Raw Data (this time as test images) and processed using the Batch Processing Mode for use in the next training step.

[0170] Ilastik Pixel Classification Training. The second round of training in Pixel Classification followed a similar process as the Density Counting. Pixel Classification uses the Random Forest supervised machine learning algorithm to perform classifications. The .TIF probabilities export of 17 test images (one from each sample) from the Density Counting workflow were used as training images in this Pixel Classification project. In the Training applet, labels were created to differentiate objects as either “Follicle” or “Background” (to identify follicles out of the rest of the image). After marking a few areas representative for each label, Live Update was selected to visually test the prediction of pixel assignment to “Follicle”, as seen in Figure ID. The “Suggest Features” tool was then utilized to optimize the prediction. Pixels were assigned a value according to the Random Forest algorithm, which calculates the chance of a pixel being “Background” or “Follicle”, producing a probability map, similar to the Density Counting workflow.

[0171] Once the Al correctly identified all the follicles, the Source was set to probabilities in .TIF format, with the range normalized to 0-255 intensity values assigned to each pixel, and the file export location was chosen in the Prediction Export applet. Because the first image of each of the 17 samples was trained on, these images retained the user generated labels. Thus, these images were no longer unseen to the Al and were not analyzed further, resulting in a final 5 test images per sample (85 images total). In the Batch Processing applet, the unseen 85 probability maps from the Density Counting workflow were imported as Raw Data, and all output .TIF files were processed in Fiji for the last time.

[0172] Image noise reduction in Fiji. All remaining 85 images that had been through the previous Cell Density and Pixel Classification Ilastik workflows, had an additional noisereduction filter applied called Robust Automatic Threshold Selection (RATS), available as a Fiji plugin (imagej.net / plugins / rats). This automatic threshold has three user-defined inputs: noise threshold, lambda factor, and minimum leaf size. The noise threshold was calculated by measuring the standard deviation of gray values in a few representative background portions containing no follicles calculating an average value of 25, lambda was set at the recommended at a value of 3, and the minimum leaf size which defaults to create 5 quadtrees was around 1000 on the images used for parameter testing. Once RATS was implemented, these images were put through a final round of Ilastik. The noise reduction is displayed in FIG. 17E.

[0173] Ilastik Object Classification Training. The Object Classification workflow is similar to that of Pixel Classification - it classifies groups of pixels (full objects) while simultaneously tabulating the count and selected measurements of these identified objects. The workflow is based on the Raw Data .H5 H&E image, and the corresponding .TIF Prediction Maps with RATS applied. Because this workflow is more computationally tasking, only six .TIF images were imported and used for Al training. Next, under the Threshold and Size Filter applet, a simple threshold level of 0.3 and a pixel size range of 1500 to 10,000 pixels were applied to the prediction maps to reduce small object noise and isolate follicles. Additionally, a smoothing sigma filter of 3.8 in both X- and Y-directions was applied to reduce pixel noise within the objects. For assistance in selecting appropriate parameters, Show Intermediate Results was turned on to see what was visually occurring at each step. Show Intermediate Results presented the smoothed image and the results before size filtering. A specific feature set was chosen to allow the Al to differentiate between follicle and non-follicle objects (FIG. 18). The features were chosen based on the visual characteristics of the follicles after the RATS was applied, with pixel intensity generally increasing towards the center of the follicle guiding the intensity features chosen, the size and circularity of the follicles. In the Object Classification applet, the objects were manually classified by the operator into “follicles” and “not follicles” for each of the training images by interacting with and labeling highlighted objects in the binary image layer. The Live Update feature was used to determine the efficacy of the manual annotations and adjust as needed. When a satisfactory accuracy level was achieved based on the six training images (FIG. 17F), the export image settings were configured.

[0174] Follicle Counts from Al. After the Ilastik Object Classification training export was initialized, the output included the export of the output image mask, which depicted predictionsfor the classified objects, and a table containing the objects predicted labels as well as selected information about their shape, intensity, and location. Upon completion of this training protocol, 85 images were assessed using batch processing, and the image (integer 64-bit .TIF files) and .CSV results were generated. The results contained objects identified by the Al program which were labeled as either “follicles” or “not follicles” based on the threshold and size filter settings, as well as the Al-assisted labeling step.

[0175] Statistical Analysis. The time to obtain export files from the Ilastik batch processing and the Fiji batch processing was noted to compare time between counting methods. This computational work was carried out on a 2022 MacBook Air with Apple M2, 8-core CPU, 8- core GPU, 8GB of RAM, and 256GB SSD. Statistical Analysis for operator comparison using the semi-automated workflow was performed in Prism 10 (Graphpad by Dotmatics, Boston, MA). To analyze the donor population age and BMI correlation, a simple linear regression was performed (n = 12 donors). To investigate the correlation between the averaged operator follicle count and the Al follicle count per-section and per-donor, a simple linear regression was performed, n = 77 sections, n = 11 donors (2-15 sections analyzed per donor). To analyze how closely the control operator and the comparison operator or the Al count matched on a per-section basis, the Bland-Altman method comparison was performed, 77 sections, n = 11 donors (2-15 sections analyzed per donor). To investigate the primordial FD and follicle depth of multiple samples from a shared donor Mann- Whitney tests were performed, n = 6 sections, n = 1 donor with 2 samples taken from the same ovary, 6 sections analyzed per sample for a total of 12 sections per donor. The relative error of follicle counts per section is shown in Equation (2) and is reported as % error, n = 59. The percent relative accuracy of follicle counts per section is shown in Equation (3) and is reported as % accuracy, n - 59 sections. Equations 2 and 3 exclude sections with zero counts by the control operator.

[0176] Calculated Realtive Error = 'Comparison Value-Control Value|Control Value

[0177] Relative Accuracy = 1 — | Calculated Relative Error |

[0178] The True Positive Rate was determined by the accurate identification of follicles by the Al as shown in Equation (4). Conversely, the False Positive Rate was determined by the Al incorrectly labeling a non-follicle structure as “follicle”, as shown in Equation (5). Additionally, from the true positive rate, the specificity of the Al follicle identification was able to be calculated, as shown in Equation (6).„ n ■ • n . ..[L0179]JTrue Positive Rate(v4)„ , „ . . „ Inorrectly Identified Non-Follicle as Follicle Count

[0180] False Positive Rate = - AI Count (5) True Positive

[0181] Sensitivity = - Manual Operator Count (6)

[0182] The time for manual follicle counting was estimated, approximating 15 minutes per section for 102 total sections to manually peruse for follicles and export data, and one minute per follicle (n = 1951) to annotate. On the other hand, the computing time for AI was exactly measured, including the time to select inputs for the software used and to sort the export for the entire dataset. The AI was run twice to determine the reproducibility of results. To analyze the effect of age or BMI on primordial FD calculated as seen in Equation (1) (and follicle depth), as well as to compare the different follicle counting methods, a Kruskal-Wallace test and Dunn’s multiple comparison test were performed, n = 102 sections, n = 17 samples (6 sections per sample). The p-values for statistical significance are represented with asterisks (*p<0.05, **p<0.005, ***p<0.0005, ****p<0.0001).

[0183] Results.

[0184] Histological analysis of ovarian cortical tissue from 12 donors. Ovarian cortical tissue froml2 donors was analyzed for primordial follicle density and distance from the ovarian surface to primordial follicles in the cortex, i.e., the follicle “depth”. The age of the donors spanned from 16 to 37 y / o, and BMI from 17.9 to 41.8 kg / m2. The age, BMI, and ethnicity of each donor are summarized in Table 4.

[0185] Table 4: Donor tissue attributes

[0186] The analyzed tissue samples varied qualitatively (based on visual assessment of WSIs) in stained colors, stroma density, number and size of capillaries, the clustering of the follicles, and varied quantitatively in the number and growth stage of follicles, and the depth of primordial follicles from the ovarian surface. Figure 2 shows representative images from WSI taken for each donor (2Ai-Axii) and a representative image for follicles at different development stages (FIG. 19B).

[0187] Donor Variation in Primordial Follicle Density. To determine whether there was a correlation between age and follicle density, primordial follicle count data from the comparison operator was analyzed. Primordial FD, as determined by Equation (1), was calculated per section and is reported as follicles / mm3. These per-section counts were binned per-donor and are plotted with increasing age, 16-37 y / o in FIG. 20 A. In ascending order, the donors were found to have the following primordial FD and standard deviation (in primordial follicles / mm3): donor 1 (455.7 ± 94.0), donor 2 (95.6 ± 66.3), donor 3 (875.0 ± 297.9), donor 4 (0.0 ± 0.0), donor 5 (11.0 ± 8.5), donor 6 (358.4 ± 291.4), donor 7 (0.0 ± 0.0), donor 8 (176.7 ±76.6), donor 9 (0.0 ± 0.0), donor 10 (117.9 ± 56.8), donor 11 (0.0 ± 0.0), and donor 12 (23.7 ±33.5). There is a general decline in primordial FD with increased age; however, there were several outliers to these trends, including a 20 y / o donor with 0 identified primordial follicles (by manual counts) and donor 10 (35 y / o) with a mean of 104.74 follicles / mm3. Further, donors 2 (18 y / o), 4 (20 y / o) and 5 (24 y / o) have low average primordial FD, on par with donors 8-12 (29-37 y / o donors). The heterogeneity in primordial FD values per section is exemplified in donors 3 (18 y / o), possessing a range of 951.8 primordial follicles / mm3, and in donor 6 (25 y / o) which has a range of 800.6 primordial follicles / mm3.

[0188] An average FD by donor was determined from all section primordial FD values associated with a donor plotted against age (FIG. 20B). Age appears to loosely correlate with donor population primordial FD as demonstrated by the analysis of comparison operator follicle counts (FIG. 21C). BMI also displayed a negative trend with primordial FD; donors inour sample population with lower BMI possessed greater primordial FD than donors with higher BMI (FIG. 21A-B).

[0189] To determine if age and BMI were confounding factors, the donor age and BMI were fit to a simple linear regression, (Equation 7, R2= 0.1826) suggesting a slight positive correlation between age and BMI of the analyzed donors.

[0190] Y = 0.3834X + 16.11 (7)

[0191] Each donor BMI (spanning 17.9 - 41.8 kg / m2) was also plotted against the calculated Primordial FD as provided by Equation (1) , but no correlation was identified due to a relatively small sample size (FIG. 21).

[0192] The distance of the primordial follicles from the ovarian cortical surface of all 12 donors was analyzed semi-automatically in Fiji. In line with previous findings of a positive correlation between age and the donor population revealed that primordial follicle depth (see, e.g., Schleedom et al.), a steady increase in mean primordial follicle depth (FIG.22) with age was found, increasing from an average depth of 472.2 microns below the cortex surface in donors 16-20 y / o, to 501.3 microns in donors 24-29 y / o, and reaching 537.6 microns in donors 33-37 y / o (FIG. 22A). Additionally, tissue taken from donors within a BMI range of 18.5-24.9 exhibited an average primordial follicle depth that was significantly closer to the surface of the ovary than in the other BMI groups (FIG. 22B).

[0193] Of the 102 total slides analyzed, 8 donors had one sample assessed, while 3 donors had 2 samples analyzed, and 1 donor had 3 samples investigated (with all donors with multiple samples taken from the same ovary) as in donors 3, 6, 9, and 12, chosen to cover a broad age range (18, 25, 32, and 37 y / o respectively). This allowed for assessment of intra-ovary location variability (FIG. 23) Of the samples that were taken from the 4 donors, one donor (donor 6) exhibited a significant difference in primordial FD from 9 land 626 follicles / mm3between sampled locations. Lastly, one donor (donor 3) showed a significant difference in primordial follicle depth from the ovarian surface from 653 and 475 micron between sampled locations (FIG. 22C-22E). A summary of primordial FD per sample and per donor is included in Table 3.

[0194] Artificial intelligence follicle count and variation compared to manual operator variation and overall follicle count. To determine if there were significant differences in the follicle counts for each method, a comparison was made between the two operators with different levels of experience, and between the operator counts (control operator, comparison operator, and average of the operators) and Al counts. The control operator analyzed 93 of the102 tissue sections analyzed by the comparison operator, and of those, the 77 sections that were unseen to the Al were compared between all methods. Although the operators also classified follicles in addition to counting them, only the comparison operator classifications were interpreted in this publication. The Al program does not yet have classification capabilities, but the “follicle” labels were compared to the comparison operator’s manual classifications to determine which classes of follicles were accurately identified by the Al.

[0195] The unit of measure for each method was the total and per-section follicle count. Within the 77 sections, the control operator counted 1 181 follicles, the comparison operator counted 1178 follicles, and the Al counted 927 follicles. The per section relative error, as seen in Equation (2), was 15.0% between comparison and control operators, which yielded 85.0% manual follicle count relative accuracy, see Equation (3), when compared to the control operator counts. The relative error of Al follicle counting per section was found to be 43.3%, and relative accuracy was 56.7% when compared to the control operator counts. To obtain an overall error and accuracy percentage, the total Al counts were compared to the total control operator counts. The overall relative error was found to be 21.5%, which resulted in 78.5% relative accuracy of follicle counts. The Al program was run twice to ensure the same results per section analyzed. The operators displayed no significant difference of follicle counts per section analyzed, with the comparison operator’s follicle count mean of 15.30 follicles / section and standard error of mean (SEM) of 2.96 (n = 1178 follicles) and the control operator’s mean of 15.34 follicles / section and SEM of 3.12 (n = 1181 follicles). The combined manual follicle count per tissue section is 15.32 follicles / section with a SEM of 3.03 (n = 1180 follicles), and the Al average follicle count is 12.04 follicles / section with a SEM of 2.55 (n = 927 follicles). To further explore this, a Friedman test with Dunn’s multiple comparison test was performed on all three follicle counting methods on a per section basis (FIG. 24A), and significance was found between both the comparison operator and the Al and the control operator and the Al, but none between the operators.

[0196] The correlation of the per-section average manual follicle counts to the Al follicle counts was high. The simple linear regression (FIG. 24B) revealed Equation (8) with an R2value of 0.9661.

[0197] Y = 0.826X - 0.612 (8)

[0198] Comparison of operators’ manual follicle counts demonstrated no significant difference in counts per section nor per donor average. The operators’ follicle counts per donor showed no statistical difference, with the comparison operator’s mean of 15.00 follicles / donorand SEM of 6.73, and the control operator’s mean of 15.45 follicles / donor and SEM of 7.42. The combined manual follicle count per donor is 15.22 follicles with a SEM of 7.07, and the Al average donor follicle count is 11.64 follicles with a SEM of 5.88. A Friedman test with Dunn’ s multiple comparison test was performed on all three follicle counting methods (FIG. 24C) exhibited only slight statistical difference between the comparison operator and the Al.

[0199] The correlation of the per-donor average manual counts to the Al follicle counts fits the line of regression closely. The simple linear regression (FIG. 24D) revealed Equation (9) with an R2value of 0.9796.

[0200] Y = 1.190X + 1.370 (9)

[0201] To further compare each follicle counting method per donor, the mean and SD of the Al and the average manual counts were investigated. Between the Al and manual operator average, 4 of 11 donors had statistical difference (FIG. 24E). There are 4 of 11 instances (donors 2, 3, 4, and 7) in which the SDs do not overlap, indicating a greater degree of mismatch in count than the other 8 donors. Of the donors displaying differences in counts, 1 / 2 did not have representative images trained on in the last Al training stage. Donor 3 was the only instance in which the Al counted a higher mean (2.6 follicles / section) than the operators did (0.2 follicles / section). The Al and the manual operator SD generally increase with follicle count.

[0202] Difference plots were generated compared to the control operator counts (with persection comparison operator counts subtracted from control operator counts) and showed good agreement. Data falling positively and negatively indicated that the comparison operator sometimes counted more or fewer follicles on a given section than the control operator (FIG. 24F). The 95% limits of agreement were at negative 8.13 and 8.21, resulting in a bias of 0.029. The Al difference plot displays good agreement between operator- and Al-generated counts until over 100 follicles / section, at which point the Al undercounted by around 30 follicles (FIG. 24G). The Al counted less follicles than the operator in 61 of 77 instances. The 95% limits of agreement were at negative 10.36 and 16.96, resulting in a bias of 3.299.

[0203] To further analyze Al accuracy, a histogram of the manual follicle count by the control operator, in bins of 20, and the average Al accuracy compared to the control operator count per bin was generated (FIG. 24H). Again, the sections in which the control operator counted 0 follicles were excluded. As visually shown, all bins reside in 74-90% accuracy, except for bin 1-20 follicles, which has an accuracy of 28.5%. Because bin 1-20 contains 43 / 59 of the follicle counts, it disproportionately affects analysis of the Al. Bin 81-100 follicles hasno associated value because the control operator did not count a section with this number of follicles.

[0204] Further investigation into the accuracy of the Al counts was conducted to obtain a more robust analysis of the data (FIG. 24). The final Al output images from the trained Object Classification program were overlaid onto the native H&E-stained tissue which allowed for visualization of follicle class to Al count. The true positive rate, see Equation (4), was 79.9% while the false positive rate, see Equation (5), was found to be 21.2% based on comparison operator counts. The sensitivity of the AT follicle identification was calculated, see Equation (6), and resulted in 73.4% sensitivity. The accuracy of primordial follicle identification was the highest at 73.3%, with a general decrease in accurate Al detection as the follicle class matured. The Al program identified follicles with nuclei and without nuclei at about the same rate, 53.0% and 48.7%, respectively. Follicles with an abnormal appearance were unable to be consistently identified with only 36.1% match.

[0205] Due to the general undercounting of Al values, a constant multiplier was empirically derived. A multiplier of 1.946 applied on a per-section basis contributes no significant difference between manual operator and Al values.

[0206] Processing times compared between automated and manual follicle counting methods. The total time needed for follicle identification was estimated for the manual process and recorded for the automated process. With 102 sections to analyze, estimating 15 minutes per section to manually peruse for follicles and export data, and an additional 1 minute per follicle to annotate (n = 1951), the manual annotations spanned an estimated total of 58.0 hours total, for an average time of 3.4 hours per tissue sample (set of 6 sections), or 34 minutes per section. Contrastingly, once trained, the automated follicle counting methods needed approximately 5 minutes to import the test dataset and select export settings per round of training. The first round of training using the Density Counting project took 37.4 hours in total (2.2 hours per sample, or 22 minutes per section). The second round of training utilizing the Pixel Classification project elapsed 5.1 hours in total (18 minutes per sample, or 3 minutes per section). The last round of training yielded all results in only 28 minutes for the 85 images it was ran on (approximately 1.5 minutes per sample of 5, or 20 seconds per section). In total, applying the three stages of trained Al programs took 43 hours for all 102 images, or 25 minutes per image. The same trained program and the same test images were run again to confirm the Al identifies follicles in a reproducible manner. On the same machine, follicle identification was one-third faster using the automated method than the manual method.Example 5

[0207] This example described analysis of entire cross sections of human ovaries. The entire section provides a more robust snapshot of the ovary, including spatial follicle density, and other anatomical structures.

[0208] The sample is processed as described in Examples 1-4 above with the exception that sections were cut with a feather blade (vs using a cutting guide) to be roughly 1 cm thick, then fixed in paraffin blocks (vs Bouin’s solution). The samples were sectioned, H&E stained, and sent to the imaging core for whole slide imaging. An SVS file with one tissue sections per slide (vs 4 cortex sections per slide) was obtained.

[0209] As with cortex strip follicle annotations, the built-in measurement feature in QuPath can be used to export the outline (whole tissue area), and manual annotations with follicle classifications, area of the follicle, and the centroids. The only change in the script is that the sample image is down-sized by lOx to get TIf file since the area is so large (cortex strip is down sample of lx). The resulting png output from Export Binary Script feeds into Fiji macro to calculate depth of follicles. Existing Fiji macros work for the whole tissue, no changes necessary, except for there is no erasing edges of the tissue outline to mark where the surface of the cortex is, since it is all cortex. The “mean” value in the table is the Euclidean distance of the follicle to the surface of the cortex.

[0210] With minimal adaptions, the semi-automated method works with human ovarian cortex strips and whole ovarian cross sections. The only changes required are sample processing (cutting and embedding), downsampling the images for export in QuPath, and the cortex surface selection in Fiji. The tissue area, follicle metadata, and depth are collected for analysis.

[0211] Using open-source, free software, whole ovarian cross sections are now able to be analyzed. Follicle density is the only direct measure of follicle reserve, and this methods allows for investigation of follicular distribution directly in a streamlined manner. Other anatomical structures such as corpus luteums and corpus albicans (remnants from ovulated follicles which still produce hormones), blood and lymph vessels, stromal cells, and more can be analyzed using this method. Information on follicle depth, specifically primordial follicle depth, may have negative correlations with age or diseased ovaries and can be analyzed using this method.Example 6

[0212] This example describes a data-driven pipeline that applies machine learning-based methods to detect follicles and measure follicle density while accounting for intra-donor variability and patient data.

[0213] Methods

[0214] Image Preprocessing and Segmentation. Histological images were standardized for brightness, contrast, and noise reduction using QuPath. Classical image processing methods were implemented using OpenCV, including thresholding, morphological operations, and watershed segmentation to detect and segment follicles. Additionally, deep learning-based follicle detection was performed using a Faster R-CNN architecture in PyTorch. Data augmentation techniques were applied to enhance model robustness, and model performance was evaluated using the Dice Coefficient and Jaccard Index. Automated follicle counts were normalized by total tissue area to account for variations in sample size.

[0215] Determination of Optimal Image Sampling. A hierarchical statistical approach was employed to determine the number of images required for accurate follicle density measurement. A subset of donor samples was selected for a pilot study to assess intra- and inter-donor variability. Mixed-effects models were constructed using the statsmodels library to quantify intra-donor variance, and intra-class correlation coefficients (ICC) were computed to determine the proportion of total variability attributable to donor-level differences. A simulation-based power analysis was performed, iteratively increasing the number of images per donor to identify the threshold at which follicle density estimates stabilized as a balance of biological sensitivity with computational accuracy.

[0216] Follicle Clustering Calculations. To analyze how follicle density was also correlated by distribution, the spatial clustering of follicles was quantified by Morisita’s index for up to 50 plot sample units per donor. Morisita’s index was defined as 15 =Q[( xi2 - N) / N(N - 1)], where Q is the number of sample units, xi to the number of follicles in each sample unit, and N = xi.3

[0217] Ensemble Feature Weighting for Predictive Modeling of Follicle Density. Demographic variables influencing follicle density were identified through feature selection and predictive modeling. Data preprocessing included imputation of missing values, encoding of categorical variables, and scaling of numerical features. Exploratory data analysis was performed using descriptive statistics and visualization techniques. Predictive modeling employed mixed-effects models and tree-based gradient boost machine algorithms (e.g.XGBoost). Model interpretability was assessed using SHapley Additive Explanations (SHAP) and Local Interpretable Model-agnostic Explanations (LIME). Dimensionality reduction techniques, including principal component analysis (PCA) and factor analysis, were applied to refine the selection of demographic predictors.

[0218] Statistical Reproducibility. Multiple means comparisons were performed by ANOVA with Newman- Keuls method. Linear regressions, including 95% confidence intervals (CI) and prediction intervals (PI), for correlations between (i) each individual donor patient feature to the estimated follicle density and (ii) the model-weighted patient feature index to the estimated follicle density for a range of number of patient slides.

[0219] Results and Significance

[0220] An overview of the framework and results in shown in FIG. 25. FIG. 25 A shows a representation of workflow to prepare and augment histology training and test sets, identify highly important features in determining follicle density, and weight features for follicle density estimation using sparse histological samples. FIG. 25B shows weighting of the top 13 patient metadata features in correlation to follicle density in the training set. FIG. 25C shows receiver-operator characteristic curve with area under the curve (AUC) indicated for the confidence vs. precision of these weighted features in determining the follicle density of the test dataset. FIG. 25D shows principal component analysis of these 13-weighted features, classified by spacing coefficient and intra-class correlation (sampling need). FIG. 25E shows model-estimated follicle density (solid line) compared to manually counted follicle samples (points) Shaded grey area represents 90% confidence interval and dotted lines represent 90% prediction interval for model estimated follicle density. All data indicated here resultant from 1000-fold cross validation with the LASSO regression at 70% training / 30% test data partition. Abbreviations: CNN - convoluted neural network, PCA - principal component analysis, BTRF - Bagged Tree Random Forest, FD - Follicle Density, ROC - Receiver- Operator Characteristic Curve, BMI - Body Mass Index, WIT - Warm Ischemic Time, KDPI - Kidney Donor Profile Index, CIT - Cold Ischemic Time, CoD - Cause of Death.

[0221] This integrated framework facilitated a comprehensive and reproducible analysis of ovarian tissue repositories, with implications for reproductive biology and personalized medicine. Specifically, unsupervised weighting of age, BMI, follicle clustering, stromal density, and other patient metadata enables increasingly accurate FD predictions while simultaneously stratifying donors by intra-ovary follicle variance to decrease required histology.

[0222] All publications and patents mentioned in the above specification are herein incorporated by reference. Various modifications and variations of the described method and system of the invention will be apparent to those skilled in the art without departing from the scope and spirit of the invention. Although the invention has been described in connection with specific preferred embodiments, it should be understood that the invention as claimed should not be unduly limited to such specific embodiments. Indeed, various modifications of the described modes for carrying out the invention that are obvious to those skilled in the relevant fields are intended to be within the scope of the following claims.

Claims

CLAIMSWE CLAIM:

1. A method comprising: a) imaging an ovarian tissue or ovarian-associated tissue to provide an image of the ovarian tissue; b) annotating one or more follicles of the imaged ovarian tissue; c) measuring and counting the one or more follicles; and d) providing a follicle density and / or follicle count.

2. The method of claim 1, wherein one or more of said imaging, annotating, measuring and counting or providing is performed by a computer processor and computer software.

3. The method of claim 1, wherein one or more of said imaging, annotating, measuring and counting or providing is performed manually by an operator.

4. The method of any one of the preceding claims, wherein one or more of said imaging, annotating, measuring and counting or providing comprises training a computer processor and computer software using one or more training images.

5. The method of any one of the preceding claims, wherein the ovarian tissue is processed by a method comprising: a) decertifying an ovary to provide a histological sample; b) fresh-fixing the histological sample to provide a fresh-fixed sample; c) washing the fresh-fixed sample in deionized water to provide a washed sample; d) dehydrating the washed sample to provide a dehydrated sample; e) orienting the dehydrated sample perpendicular to ovary cortex surface and embedding the dehydrated sample in paraffin to provide a paraffined sample; f) serially sectioning the paraffined sample to provide at least four sectioned samples; g) adding the at least four sectioned samples to a slide; and h) staining the slide.

6. The method any one of the preceding claims, wherein the ovarian tissue is a whole ovarian cross-section.

7. The method of any one of the preceding claims, wherein the ovarian tissue is human ovarian cortex strips.

8. The method of any one of the preceding claims, wherein the ovarian-associated tissue is selected from corpus luteums, corpus albicans, blood vessels, lymph vessels, and stromal cells.

9. The method of any one of the preceding claims, wherein imaging the ovarian tissue comprises: a) whole slide imaging the ovarian tissue with a high-resolution scanner to provide a whole slide image file; b) importing the whole slide image file into a first software.

10. The method of any one of the preceding claims, wherein annotating a one or more follicle of the imaged ovarian tissue comprises: a) selecting a tissue section region of interest from the whole slide image file using the first software; b) generating a tissue section number: i) for the tissue section region of interest; and ii) determining follicle depth and position for each of the regions of interest; c) annotating the tissue section number to provide an annotated tissue section: i) by color coding; ii) by positional location fidelity; and iii) by classifying the one or more follicle; and d) exporting the annotated tissue section from the first software to a second software.

11. The method of any one of the preceding claims, wherein measuring and counting the one or more follicle of the imaged ovarian tissue comprises: a) measuring, using the second software, (i) a distance from the cortex to the centroid of the follicle; and (ii) a tissue section area; and b) counting, using the second software, the one or more follicles.

12. The method of any one of the preceding claims, wherein annotating a one or more follicle of the imaged ovarian tissue comprises: a) selecting a tissue section region of interest from the whole slide image file using the first software; b) generating a tissue section number: i) for the tissue section region of interest; and ii) related to a follicle depth and position; and c) annotating the tissue section number to provide an annotated tissue section: i) by color coding; ii) by positional location fidelity; and iii) by classifying the one or more follicle.

13. The method of any one of the preceding claims, wherein measuring and counting the one or more follicle of the imaged ovarian tissue comprises: a) measuring, using the first software, (i) a distance from the cortex to the centroid of the follicle; and (ii) a tissue section area; and b) counting, using the first software, the one or more follicle.

14. The method of any one of the preceding claims, further comprising determining one or more patient variables selected from age, body mass index, ovary volume, ovary volume, weight, ethnicity, Kidney Donor Profile Index (KDPI), process duration, Cold Ischemic Time (CIT), blood type, and cause of death (CoD).

15. An information system comprising: a) one or more imaging devices; b) one or more network communication components; c) one or more first software; and d) one or more second software; wherein the information system is used for detecting a follicle density in an ovarian tissue and performing actions on the ovarian tissue.

16. The system of claim 15, wherein the one or more network communication component comprises an input device and an output device.

17. The system of claim 15 or 16, wherein the one or more application software comprises one or more protocols for: a) imaging an ovarian tissue to provide an image of theovarian tissue; b) annotating one or more follicles of the image of the ovarian tissue; c) counting a number of follicles of the image of the ovarian tissue; and d) providing a follicle density of the ovarian tissue.

Images