Systems and methods for generating normalized images of biological tissue sections

Abstract

The present disclosure provides methods, systems, and computer-readable media for stain normalization image processing of digitized biological tissue images. A method includes identifying a plurality of stain channels from an input image of a biological tissue section, generating a plurality of images for each stain channel based on relative stain concentration, and outputting stain-normalized images based on the generated images. A method for constructing scale invariant images of biological tissue sections includes identifying nuclei within an image, setting a length scale based on nuclei size, and scaling parameters for other morphological structures using the length scale. A method for identifying morphological structures includes generating a stain channel from an input image, determining a lower bound, creating a segmentation map, and identifying a morphological structure from the segmentation map. Systems and computer-readable media for performing these methods are also provided.

Description

SYSTEMS AND METHODS FOR GENERATING NORMALIZED IMAGES OF BIOLOGICAL TISSUE SECTIONS CROSS-REFERENCES TO RELATED APPLICATIONS

[0001] This application claims priority to U.S. Application No. 63 / 622,894, titled “Systems and Methods for Generating Normalized Images of Biological Tissue Sections,” and U.S. Application No. 63 / 622,925, titled “Systems and Methods for Morphological Feature Extraction from Digital Images of Biological Tissue Sections,” both filed on January 19, 2024, each of which is hereby incorporated by this reference in its entirety. BACKGROUND

[0002] Histopathological examination of sections of biological tissue plays a pivotal role in the diagnosis and prognosis of numerous medical conditions (e.g., carcinoma, dysplasia, benign tumors, celiacs disease, ulcerative colitis, etc.). Conventionally, this examination involves the preparation of thin tissue sections, staining the sections with specific dyes, and then manually scrutinizing slides with the prepared tissue sections under a microscope to identify and / or characterize morphologic structures within the tissue sections. SUMMARY

[0003] In some examples, the techniques described herein relate to a computer- implemented method including generating, from an input image of a biological tissue section, a stain channel indicating a stain of interest, determining, from the stain channel, a lower bound, creating a segmentation map based on the lower bound, and identifying a morphological structure of the biological tissue section from the segmentation map. In some examples, the stain channel may include a range of pixel intensity values corresponding to a range of stain concentrations, themethod may further include generating multiple images for the stain channel, each of which captures a different subrange in the stain channel's range of pixel intensity values, the multiple images may include a stain-high image that captures a subrange of pixel intensity values that are higher than the pixel intensity values of any of the other subranges in the range of pixel intensity values, and the lower bound may include pixel intensity values from the stain-high image.

[0004] In some examples, creating the segmentation map from the lower bound may include applying a correction term to the lower bound and creating the segmentation map from the corrected lower bound. In these examples, the method may further include generating, from the input image, multiple stain channels, including the stain channel indicating the stain that stains for nuclei, where each stain channel within the multiple stain channels corresponds to a different stain applied to the biological tissue section, generating, from the multiple stain channels, an upper bound, and generating the correction term based on the upper bound. In some examples, (1) each stain channel within the multiple stain channels may include a different range of pixel intensity values corresponding to a different range of stain concentrations, (2) the method may further include generating multiple images for each stain channel within the multiple stain channels, (3) each of the multiple images may include a stain-background image, a stain-high image, and a stain- low image, each of which captures pixel intensity values within a different subrange of the corresponding stain channel's range of pixel intensity values, (4) the stain-background image may capture a subrange of pixel intensity values corresponding to a background level of a stain indicated by the corresponding stain channel, (5) the stain-high image may capture a subrange of pixel intensity values that are higher than the pixel intensity values in the subranges of the stain- background image and the stain-low image, and (6) the stain-low image may capture a subrange of pixel intensity values that are higher than the pixel intensity values in the subrange of thebackground image and lower than the pixel intensity values in the subrange of the stain-high image. In some examples, each of the multiple images may further include one or more stain-medium images including a subrange of pixel intensity values that are between the pixel intensity values in the subranges of the stain-high images and the stain-low images. In some examples, the upper bound may include pixels based on one or more of the images generated for each of the multiple stain channels except for the stain-background images.

[0005] In some examples, the stain channel may indicate a stain that stains for nuclei. In some such examples, the stain may include at least one of a hematoxylin stain, or a fluorescent stain that binds to DNA.

[0006] In some examples, the techniques described herein relate to a computer- implemented method including identifying, from an input image of a biological tissue section, multiple stain channels, each indicating a different stain applied to the biological tissue section, normalizing, based on the multiple stain channels, each pixel intensity value for the input image of the biological tissue section to one or more normalized values to yield one or more stain- normalized images of the biological tissue section, and identifying a morphological structure of the biological tissue section from the one or more stain-normalized images of the biological tissue section.

[0007] In some examples, the biological tissue section may be a section of intestinal tissue, and the morphologic structure may be a villus-crypt pair. In some examples, the multiple stain channels may include a hematoxylin channel, corresponding to a hematoxylin stain, and an eosin channel, corresponding to an eosin stain.

[0008] In some examples, the method may further include identifying, from the input image and / or a stain-normalized image generated from the input image, multiple nuclei, and settinga length scale based on a size of one or more of the multiple nuclei. In some examples, the method may further include determining a quantitative histological measure of the morphologic structure using the length scale (e.g., a villus-height to crypt-depth ratio of a villus-crypt pair).

[0009] In some examples, identifying the morphologic structure may include determining that the morphological structure satisfies a quality metric. In these examples, the method may further include, in response to determining that the morphological structure satisfies the quality metric, generating a user interface, including a display of the morphological structure, that visually draws attention to the morphological structure, visually indicates that the morphological satisfies the quality metric, and / or is prioritized in a queue of user interfaces, where each user interface in the queue may include an image of a different section of the biological tissue. In some such examples, the method may further include determining that an image of an additional section of the biological tissue either doesn't include any instance of the morphological structure or includes an additional instance of the morphological structure that doesn't satisfy the quality metric, in response to determining that the morphological structure in the section of the biological tissue satisfies the quality metric, adding the input image to a digital queue of usable images for a pathologist to review, and in response to determining that the image of the additional section of the biological tissue either doesn't include any instance of the morphological structure or includes an additional instance of the morphologic structure that doesn't satisfy the quality metric, precluding the image of the additional section from the digital queue of usable images.

[0010] In some examples, identifying the morphological structure from the one or more stain-normalized images may include generating a segmentation map of the one or more stain- normalized images and identifying the morphologic structure from the segmentation map. In some examples, the method may further include identifying a lower bound of the stain-normalized imageand generating the segmentation map may include segmenting the lower bound to generate the segmentation map and / or applying a correction term to the lower bound and segmenting the corrected lower bound to generate the segmentation map.

[0011] In some examples, the techniques described herein relate to a computer- implemented method of stain normalization image processing for digitized biological tissue images including identifying, from an input image of a biological tissue section, multiple stain channels, each indicating a different stain of a multiple stains applied to the biological tissue section and including a range of pixel intensity values corresponding to a range of concentrations for that stain, generating, for each stain channel within the multiple the stain channels, a multiple images based on relative stain concentration, where each image within the multiple images captures pixel intensity values within a different segment of the stain channel's range of pixel intensity values, and outputting one or more stain-normalized images based on one or more of the multiple images generated for each stain channel.

[0012] In some examples, the range of pixel intensity values for each stain channel may include a pixel intensity value for each pixel of the input image that represents a stain intensity, in the pixel, corresponding to concentration of the stain indicated by the stain channel. In some examples, the range of pixel intensity values within each stain channel includes a range of grayscale intensity values.

[0013] In some examples, identifying the multiple stain channels from the input image of the biological tissue section can include deconvolving color values for the input image of the biological tissue section into the multiple stain channels. In some examples, at least one image, within the multiple images generated for each stain channel, may capture one or more dynamics of pixel intensity levels across the pixel intensity values of the image (e.g., spatial dynamicsindicating variations in pixel intensity across two or more regions of the image, contrast dynamics indicating a variation between one or more of the relatively highest pixel intensity values within the image and one or more of the relatively lowest pixel intensity values within the image, brightness and saturation dynamics corresponding to relative quantities of pixel intensity values designated as a high pixel intensity value and pixel intensity values designated as low pixel intensity values, and / or distribution dynamics indicating a distribution of the pixel intensity values captured by the image).

[0014] In some examples, the multiple images generated for a particular stain channel, within the multiple stain channels, may include a stain-background image, a stain-high image, and a stain-low image. In these examples, the stain-background image may capture a segment of pixel intensity values corresponding to a background level of a stain indicated by the particular stain channel, the stain-high image may capture a segment of pixel intensity values that are higher than the pixel intensity values in each of the other segments of the particular stain channel's range of pixel intensity values, and the stain-low image may capture a segment of pixel intensity values that are above the pixel intensity values of the segment captured in the stain-background image and below the pixel intensity values of the segment captured in the stain-high image. Additionally, in some examples, the multiple images generated for the particular stain channel may include a stain- moderate image and the stain-moderate image may capture a segment of pixel intensity values that are above the pixel intensity values of the segment captured in the stain-low image and below the pixel intensity values of the segment captured in the stain-high image.

[0015] In some examples, the particular stain channel may represent a hematoxylin channel corresponding to a hematoxylin stain and / or an eosin channel corresponding to an eosin stain. In some examples, the input image may be captured via a camera of a microscope.

[0016] In some examples, the techniques described herein relate to a computer- implemented method for constructing scale invariant images of biological tissue sections, the method including identifying multiple nuclei of cells within an image of a section of biological tissue, setting a length scale based on a size of one or more of the multiple nuclei, and scaling, for a morphological structure other than a nuclei that is represented in the image, a parameter related to length using the length scale that is based on the size of the one or more nuclei.

[0017] In some examples, setting the length scale based on the size of one or more of the multiple nuclei may include setting the length scale based on an average size of the multiple nuclei. In some examples, the image of the section of biological tissue may represent a stain- normalized image in which pixel intensity values corresponding to stain concentrations of a stain have been normalized, where the stain is a stain, applied to the section of biological tissue, that stains for cell nuclei (e.g., a hematoxylin stain, a fluorescent stain that binds to DNA, a Feulgen stain, a methyl green stain, or a Giemsa stain).

[0018] In some examples, the image may represent a stain-high normalized image including pixel intensity values that are higher than pixel intensity values, corresponding to stain concentrations of the stain, of one or more additional stain-normalized images. In some examples, the section of biological tissue may include a section of intestinal tissue, the morphological structure may include a villus-crypt pair, and the parameter related to length may include a villus- height to crypt-depth ratio for the villus-crypt pair. In some examples, the image may be captured via a camera of a microscope at an unknown level of magnification.

[0019] In some examples, the techniques described herein relate to a computer- implemented method including determining, from an input image of an intestinal tissue section, multiple stain channels including a hematoxylin channel, corresponding to a hematoxylin stainapplied to the intestinal tissue section, and an eosin channel, corresponding to an eosin stain applied to the intestinal tissue section, normalizing, based on the multiple stain channels, each pixel intensity value for the input image of the intestinal tissue section to one or more normalized values, to yield one or more stain-normalized images of the intestinal tissue section, and identifying one or more villus-crypt pairs of the intestinal tissue section from the one or more stain- normalized images of the intestinal tissue section. In some examples, identifying the one or more villus-crypt pairs includes may include identifying a villus-crypt pair by identifying a central line, corresponding to a portion of the intestinal tissue, from which both a first line, corresponding to a villus, is projected in a first direction and a second line, corresponding to a crypt, is projected in a second direction, and the central line may correspond to a center of intensity identified for pixels of at least one of a lower bound or a corrected lower bound generated from the one or more stain- normalized images.

[0020] In some examples, the techniques described herein relate to at least one computer-readable storage medium having encoded thereon executable instructions that, when executed by at least one processor, cause the at least one processor to carry out one or more aspects of the methods just described. In some examples, the techniques described herein relate to an apparatus including at least one processor, and at least one computer-readable storage medium having encoded thereon executable instructions that, when executed by the at least one processor, cause the at least one processor to carry one or more aspects of the methods just described.

[0021] Method and systems of the presently disclosed embodiments were isolated or otherwise manufactured in connection with the examples provided below. Other features and advantages will be apparent from the detailed description, and from the claims.

[0022] In some examples, the techniques described herein relate to a computer- implemented method of stain normalization image processing for digitized biological tissue images including identifying, from an input image of a biological tissue section, multiple stain channels, each indicating a different stain of a multiple stains applied to the biological tissue section and including a range of pixel intensity values corresponding to a range of concentrations for that stain, generating, for each stain channel within the multiple the stain channels, a multiple images based on relative stain concentration, where each image within the multiple images captures pixel intensity values within a different segment of the stain channel's range of pixel intensity values, and outputting one or more stain-normalized images based on one or more of the multiple images generated for each stain channel.

[0023] In some examples, the range of pixel intensity values for each stain channel may include a pixel intensity value for each pixel of the input image that represents a stain intensity, in the pixel, corresponding to concentration of the stain indicated by the stain channel. In some examples, the range of pixel intensity values within each stain channel includes a range of grayscale intensity values.

[0024] In some examples, identifying the multiple stain channels from the input image of the biological tissue section can include deconvolving color values for the input image of the biological tissue section into the multiple stain channels. In some examples, at least one image, within the multiple images generated for each stain channel, may capture one or more dynamics of pixel intensity levels across the pixel intensity values of the image (e.g., spatial dynamics indicating variations in pixel intensity across two or more regions of the image, contrast dynamics indicating a variation between one or more of the relatively highest pixel intensity values within the image and one or more of the relatively lowest pixel intensity values within the image,brightness and saturation dynamics corresponding to relative quantities of pixel intensity values designated as a high pixel intensity value and pixel intensity values designated as low pixel intensity values, and / or distribution dynamics indicating a distribution of the pixel intensity values captured by the image).

[0025] In some examples, the multiple images generated for a particular stain channel, within the multiple stain channels, may include a stain-background image, a stain-high image, and a stain-low image. In these examples, the stain-background image may capture a segment of pixel intensity values corresponding to a background level of a stain indicated by the particular stain channel, the stain-high image may capture a segment of pixel intensity values that are higher than the pixel intensity values in each of the other segments of the particular stain channel's range of pixel intensity values, and the stain-low image may capture a segment of pixel intensity values that are above the pixel intensity values of the segment captured in the stain-background image and below the pixel intensity values of the segment captured in the stain-high image. Additionally, in some examples, the multiple images generated for the particular stain channel may include a stain- moderate image and the stain-moderate image may capture a segment of pixel intensity values that are above the pixel intensity values of the segment captured in the stain-low image and below the pixel intensity values of the segment captured in the stain-high image.

[0026] In some examples, the particular stain channel may represent a hematoxylin channel corresponding to a hematoxylin stain and / or an eosin channel corresponding to an eosin stain. In some examples, the input image may be captured via a camera of a microscope.

[0027] In some examples, the techniques described herein relate to a computer- implemented method for constructing scale invariant images of biological tissue sections, the method including identifying multiple nuclei of cells within an image of a section of biologicaltissue, setting a length scale based on a size of one or more of the multiple nuclei, and scaling, for a morphological structure other than a nuclei that is represented in the image, a parameter related to length using the length scale that is based on the size of the one or more nuclei.

[0028] In some examples, setting the length scale based on the size of one or more of the multiple nuclei may include setting the length scale based on an average size of the multiple nuclei. In some examples, the image of the section of biological tissue may represent a stain- normalized image in which pixel intensity values corresponding to stain concentrations of a stain have been normalized, where the stain is a stain, applied to the section of biological tissue, that stains for cell nuclei (e.g., a hematoxylin stain, a fluorescent stain that binds to DNA, a Feulgen stain, a methyl green stain, or a Giemsa stain).

[0029] In some examples, the image may represent a stain-high normalized image including pixel intensity values that are higher than pixel intensity values, corresponding to stain concentrations of the stain, of one or more additional stain-normalized images. In some examples, the section of biological tissue may include a section of intestinal tissue, the morphological structure may include a villus-crypt pair, and the parameter related to length may include a villus- height to crypt-depth ratio for the villus-crypt pair. In some examples, the image may be captured via a camera of a microscope at an unknown level of magnification.

[0030] In some examples, the techniques described herein relate to a computer- implemented method including generating, from an input image of a biological tissue section, a stain channel indicating a stain of interest, determining, from the stain channel, a lower bound, creating a segmentation map based on the lower bound, and identifying a morphological structure of the biological tissue section from the segmentation map. In some examples, the stain channel may include a range of pixel intensity values corresponding to a range of stain concentrations, themethod may further include generating multiple images for the stain channel, each of which captures a different subrange in the stain channel's range of pixel intensity values, the multiple images may include a stain-high image that captures a subrange of pixel intensity values that are higher than the pixel intensity values of any of the other subranges in the range of pixel intensity values, and the lower bound may include pixel intensity values from the stain-high image.

[0031] In some examples, creating the segmentation map from the lower bound may include applying a correction term to the lower bound and creating the segmentation map from the corrected lower bound. In these examples, the method may further include generating, from the input image, multiple stain channels, including the stain channel indicating the stain that stains for nuclei, where each stain channel within the multiple stain channels corresponds to a different stain applied to the biological tissue section, generating, from the multiple stain channels, an upper bound, and generating the correction term based on the upper bound. In some examples, (1) each stain channel within the multiple stain channels may include a different range of pixel intensity values corresponding to a different range of stain concentrations, (2) the method may further include generating multiple images for each stain channel within the multiple stain channels, (3) each of the multiple images may include a stain-background image, a stain-high image, and a stain- low image, each of which captures pixel intensity values within a different subrange of the corresponding stain channel's range of pixel intensity values, (4) the stain-background image may capture a subrange of pixel intensity values corresponding to a background level of a stain indicated by the corresponding stain channel, (5) the stain-high image may capture a subrange of pixel intensity values that are higher than the pixel intensity values in the subranges of the stain- background image and the stain-low image, and (6) the stain-low image may capture a subrange of pixel intensity values that are higher than the pixel intensity values in the subrange of thebackground image and lower than the pixel intensity values in the subrange of the stain-high image. In some examples, each of the multiple images may further include one or more stain-medium images including a subrange of pixel intensity values that are between the pixel intensity values in the subranges of the stain-high images and the stain-low images. In some examples, the upper bound may include pixels based on one or more of the images generated for each of the multiple stain channels except for the stain-background images.

[0032] In some examples, the stain channel may indicate a stain that stains for nuclei. In some such examples, the stain may include at least one of a hematoxylin stain, or a fluorescent stain that binds to DNA.

[0033] In some examples, the techniques described herein relate to a computer- implemented method including identifying, from an input image of a biological tissue section, multiple stain channels, each indicating a different stain applied to the biological tissue section, normalizing, based on the multiple stain channels, each pixel intensity value for the input image of the biological tissue section to one or more normalized values to yield one or more stain- normalized images of the biological tissue section, and identifying a morphological structure of the biological tissue section from the one or more stain-normalized images of the biological tissue section.

[0034] In some examples, the biological tissue section may be a section of intestinal tissue, and the morphologic structure may be a villus-crypt pair. In some examples, the multiple stain channels may include a hematoxylin channel, corresponding to a hematoxylin stain, and an eosin channel, corresponding to an eosin stain.

[0035] In some examples, the method may further include identifying, from the input image and / or a stain-normalized image generated from the input image, multiple nuclei, and settinga length scale based on a size of one or more of the multiple nuclei. In some examples, the method may further include determining a quantitative histological measure of the morphologic structure using the length scale (e.g., a villus-height to crypt-depth ratio of a villus-crypt pair).

[0036] In some examples, identifying the morphologic structure may include determining that the morphological structure satisfies a quality metric. In these examples, the method may further include, in response to determining that the morphological structure satisfies the quality metric, generating a user interface, including a display of the morphological structure, that visually draws attention to the morphological structure, visually indicates that the morphological satisfies the quality metric, and / or is prioritized in a queue of user interfaces, where each user interface in the queue may include an image of a different section of the biological tissue. In some such examples, the method may further include determining that an image of an additional section of the biological tissue either doesn't include any instance of the morphological structure or includes an additional instance of the morphological structure that doesn't satisfy the quality metric, in response to determining that the morphological structure in the section of the biological tissue satisfies the quality metric, adding the input image to a digital queue of usable images for a pathologist to review, and in response to determining that the image of the additional section of the biological tissue either doesn't include any instance of the morphological structure or includes an additional instance of the morphologic structure that doesn't satisfy the quality metric, precluding the image of the additional section from the digital queue of usable images.

[0037] In some examples, identifying the morphological structure from the one or more stain-normalized images may include generating a segmentation map of the one or more stain- normalized images and identifying the morphologic structure from the segmentation map. In some examples, the method may further include identifying a lower bound of the stain-normalized imageand generating the segmentation map may include segmenting the lower bound to generate the segmentation map and / or applying a correction term to the lower bound and segmenting the corrected lower bound to generate the segmentation map.

[0038] In some examples, the techniques described herein relate to a computer- implemented method including determining, from an input image of an intestinal tissue section, multiple stain channels including a hematoxylin channel, corresponding to a hematoxylin stain applied to the intestinal tissue section, and an eosin channel, corresponding to an eosin stain applied to the intestinal tissue section, normalizing, based on the multiple stain channels, each pixel intensity value for the input image of the intestinal tissue section to one or more normalized values, to yield one or more stain-normalized images of the intestinal tissue section, and identifying one or more villus-crypt pairs of the intestinal tissue section from the one or more stain- normalized images of the intestinal tissue section. In some examples, identifying the one or more villus-crypt pairs includes may include identifying a villus-crypt pair by identifying a central line, corresponding to a portion of the intestinal tissue, from which both a first line, corresponding to a villus, is projected in a first direction and a second line, corresponding to a crypt, is projected in a second direction, and the central line may correspond to a center of intensity identified for pixels of at least one of a lower bound or a corrected lower bound generated from the one or more stain- normalized images.

[0039] In some examples, the techniques described herein relate to at least one computer-readable storage medium having encoded thereon executable instructions that, when executed by at least one processor, cause the at least one processor to carry out one or more aspects of the methods just described. In some examples, the techniques described herein relate to an apparatus including at least one processor, and at least one computer-readable storage mediumhaving encoded thereon executable instructions that, when executed by the at least one processor, cause the at least one processor to carry one or more aspects of the methods just described.

[0040] In some examples, the techniques described herein relate to at least one computer-readable storage medium having encoded thereon executable instructions that, when executed by at least one processor, cause the at least one processor to carry out one or more aspects of the methods just described. In some examples, the techniques described herein relate to an apparatus including at least one processor, and at least one computer-readable storage medium having encoded thereon executable instructions that, when executed by the at least one processor, cause the at least one processor to carry one or more aspects of the methods just described.

[0041] Method and systems of the presently disclosed embodiments were isolated or otherwise manufactured in connection with the examples provided below. Other features and advantages will be apparent from the detailed description, and from the claims. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] FIG. 1 is a block diagram of a tissue analysis system by which some embodiments may operate.

[0043] FIGS.2–3 and FIGS.5–8 are flow charts of processes 200, 300, 500, 600, 700, and 800, respectively, that may be implemented in some embodiments for an automated biological tissue analysis framework.

[0044] FIG.4 is an exemplary block diagram showing an aspect of an exemplary stain normalization process.

[0045] FIG.9 is an exemplary implementation of a computing device that may be used in a system implementing techniques described herein.

[0046] FIGS. 10, 21, 35, and 40–43 are exemplary graphs describing features of one implementation of the disclosed automated biological tissue analysis framework.

[0047] FIGS.11–20, 22–34, and 36–37 are exemplary images of biological tissue used to illustrated exemplary implementations of the disclosed automated biological tissue analysis framework.

[0048] FIGS.38–39 are exemplary user interfaces that may be provided as part of one exemplary implementation of the disclosed automated biological tissue analysis framework.

[0049] While the above-identified drawings set forth presently disclosed embodiments, other embodiments are also contemplated, as noted in the discussion. This disclosure presents illustrative embodiments by way of representation and not limitation. Numerous other modifications and embodiments can be devised by those skilled in the art which fall within the scope and spirit of the principles of the presently disclosed embodiments. DETAILED DESCRIPTION

[0050] Biological tissue sections exhibit a diverse range of structures, textures, and cell types. Identifying and / or analyzing morphologic structures within tissue architectures, including complex tissue architectures, uses imaging techniques and processing software, and to achieve high reliability desired for use cases in healthcare (e.g., diagnostics) may need to be capable of distinguishing subtle variations with high precision. The inventors have recognized and appreciated that one challenge facing the high reliability of this analysis and / or identification is staining variability in tissue sections that are presented for histopathological analysis. Biological tissue samples may be stained using various dyes for various purposes, which may include to highlight specific cellular components for analysis, enhance contrast, or other purposes. Unfortunately, conventionally, staining protocols may vary between different histopathologicallaboratories, directly leading to varying stain intensity and / or stain quality. Even within the same laboratory, variations in staining may exist due to factors such as batch variability, fluctuations in laboratory conditions, variable equipment calibration, and more. Such variation introduces unpredictability in the appearance of digitized images of tissue sections. Such variation in appearance present challenges to image processing techniques, including in identification or analysis and automation of histopathological tasks. Some automation techniques may be configured to analyze images for certain features and / or patterns as a part of extracting information and / or accurately segmenting digitized images of tissue sections, and variation in appearance can undermine such techniques’ ability to reliably identify features and / or patterns.

[0051] Conventional color normalization methods (used for image processing in other contexts) are ineffective for normalizing colors within images of biological tissue sections. Such methods typically normalize colors by deconvolving the colors of an image into three color channels: red, green, and blue. However, staining protocols used to prepare biological tissue samples predominantly manifest in the red channel, resulting in a disproportionately heightened representation within this channel. As a result, conventional methods fail to yield a useful result when normalizing images of stained tissue sections into red, green, and blue channels.

[0052] In contrast, some techniques disclosed herein include normalizing pixel values for an image of biological tissue using a novel type of channel (e.g., a channel capable of meaningfully delineating stain concentration variance within a stained image). Instead of normalizing pixel values by deconvolving the colors of an input input into color channels (e.g., splitting an RGB into red, green, and blue channels), these disclosed techniques can include deconvolving (e.g., digitally unmixing) the colors of an input image (e.g., an RGB input image) into separate stain channels (e.g., with one stain channel for each stain that was used to process abiological tissue captured in the input image). Each stain channel may include a range of pixel values (corresponding to stain concentrations). In some examples, a stain channel may be separated into multiple normalized images. Each normalized image may capture a different segment of the range of pixel values. These techniques will be described in greater detail below in connection with FIGS. 2–43.

[0053] Another challenge recognized and appreciated by the inventors is variation in magnification. Tissue sections may be imaged at various levels of magnification, introducing further unpredictability in the appearance of digitized images of tissue sections. Conventional solutions involve the use of a ruler integrated with a microscope, which enables pathologists to take measurements in real time as they are observing biological tissue sections. However, such solutions are ineffective when measurements are taken post-observation and information regarding the degree of magnification applied to a tissue section is unavailable. In contrast, some techniques disclosed herein include generating a normalized length scale for an image (derived from entities within an image) that may be generated without requiring information relating to the level of magnification applied to the image. Because of the variance in the size of morphologic structures, generating a length scale from intrinsic features of an image of biological tissue was previously considered infeasible. However, the inventors recognized and appreciated nuclei as a morphologic feature that (e.g., on average) is of a consistent size (e.g., across different individuals and / or samples). Thus, as will be described in greater detail below, some of the disclosed techniques include generating a normalized length scale based on a determined size of one or more nuclei within an image.

[0054] Another challenge recognized and appreciated by the inventors is the challenge of computationally determining a tissue boundary. Histopathological images often containbackground noise, artifacts, or non-tissue elements, which can complicate the accurate delineation of tissue boundaries. Identifying tissue boundaries can also be complicated by tissue heterogeneity (e.g., variations in composition and / or density). Responding to this challenge, some of the techniques described herein include novel approaches to identifying morphologic structures and / or features (e.g., from stain and scale normalized images that correct for noise, artifacts, non-tissue elements, etc.). These approaches can improve the uniformity and accuracy of morphological measurements and / or feature extraction across diverse tissue sample, reducing measurement discrepancies that may arise from human error or variations in manual techniques.

[0055] In addition, or as an alternative, to identifying morphologic structures and / or features within an image of a biological tissue section, in some examples the disclosed tissue analysis system may analyze each image in a set of images (e.g., where each image corresponds to a different section of a biological tissue sample) to identify well-oriented images. In some such examples, the tissue analysis system may select well-oriented images for a histopathologist’s review and / or prioritize well-oriented images in a queue generated for a histopathologist’s review.

[0056] As will be explained throughout this application, features described herein improve the functioning of a computer. For example, by enabling machine learning models to handle variations encountered in clinical practice, some techniques described herein can enable a computer to analyze histopathological images. By enabling machine learning models to accurately identify tissue boundaries within an image of a biological tissue section, processes and features described herein may enable accurate and consistent computerized detection of morphological structures and features. Additionally, by enabling trained models to adapt to real-world histopathological scenarios, image processing can be improved.

[0057] Described below are examples of ways in which techniques described herein may be implemented. It should be appreciated that these examples are merely illustrative, that embodiments are not limited to operating in accordance with the specific examples shown in the figures and discussed below, and that other embodiments are possible. The following examples are put forth so as to provide those of ordinary skill in the art with a complete disclosure and description of how to make and use the tissue analysis system and some aspects and embodiments herein and are not intended to limit the scope of the various aspects and embodiments herein.

[0058] FIG. 1 illustrates a block diagram of a system 100 with which some embodiments may operate for analyzing tissue. System 100 may, in some cases, perform techniques described below in connection with FIGs.2–43. As depicted in FIG.1, system 100 can include a computing device 102. Computing device 102 may be any type or form of device that may perform functions directed at automated histopathological tissue analysis. Additionally, system 100 can include a tissue analysis facility 104 that, for example, may be configured to perform one or more of the acts described in connection with FIGs. 2–43. In some examples, the tissue analysis facility 104 may be configured to (1) transform an input image 106 of a section of biological tissue 108 into a normalized image 110 and / or (2) algorithmically analyze the normalized image 110 (e.g., to identify a morphological structure and / or feature, such as morphological structure 112 and / or a feature of the morphological structure 112). The input image 106 may represent an image of the biological tissue section 108 (e.g., captured via a microscope 114). Each of these elements will be described presently in greater detail, in connection with FIGs. 2–43.

[0059] FIG.2 describes an example of an integrated framework describing how one or more features of the illustrative frameworks of FIGs. 3–8 may work together, in someembodiments. FIGs. 3–4 illustrate a framework for generating a stain-normalized image of a biological tissue section that may be used in some embodiments. FIG.5 shows a process that may be used in some embodiments for generating a scale-normalized image of a biological tissue section. FIG. 6 is a flowchart of an illustrative process of some embodiments for generating a segmentation map from a lower bound determined from a stain channel indicating a stain that stains for nuclei. FIG.7 illustrates some techniques for identifying a morphological structure from a stain-normalized image, that may be used in some embodiments. And FIG.8 illustrates a process that may be used in some embodiments for identifying a villus-crypt pair from a stain-normalized image of a section of intestinal tissue. Each of these frameworks will now be described, in turn. INTEGRATED HISTOPATHOLOGICAL PROCESSING FRAMEWORK

[0060] FIG.2 provides a flow chart of a process 200 that may be implemented in some embodiments by a system, such as system 100 in FIG.1, to process and / or analyze an image of a section of biological tissue. Processes 300, 500, 600, 700, and 800, provided in the flow charts of FIGS.3, 5, 6, 7, and 8, provide exemplary embodiments of the various aspects of process 200. The processes of FIGS. 2–3 and 5–8 can be implemented in some embodiments by tissue analysis facility 104, described in connection with FIG.1, on a computer-readable storage media.

[0061] Process 200 begins at step 210 with a staining process in which multiple stains are applied to a section of biological tissue (e.g., biological tissue section 108). Each of the stains may stain for a different cell type or structure and may stain a different color. The section of biological tissue may be stained with any stain or combination of stains (e.g., designed to stain for any type or form of microscopic entity). In one specific example, the staining process may include staining the section of biological tissue with hematoxylin and eosin. In this example, the hematoxylin and eosin staining may help distinguish between nuclei (stained blue by hematoxylin)and cytoplasm (stained pink by eosin). In some examples, the staining process may include staining the section of biological tissue with one or more stains that stain for a particular cellular structure (e.g., mitochondria, the Golgi apparatus, etc.). In additional or alternative examples, the staining process may include an immunochemistry (IHC) staining process in which antibodies labeled with different stains are used to help detect the presence (or absence) of specific proteins with the section of biological tissue. While systems and techniques are described herein in connection with images of fixed tissue sections, it should be appreciated that the systems and techniques described herein (e.g., for inferring structural information) can be used with images of any biological tissue, including tissues that do not result from sectioning.

[0062] In conventional workflows, after staining a biological tissue section, the stained section is manually analyzed (e.g., by a pathologist viewing the stained section under a microscope). The following steps may be used in connection with conventional workflows (e.g., to assist a pathologist) and / or as an automated alternative to conventional workflows. In the clinical setting, the staining process described at step 210 results in stained tissue sections that are subject to a great deal of inter-laboratory and / or intra-laboratory variation in staining. As mentioned previously, such variability may arise from a variety of factors (e.g., differing laboratory protocols, reagent and batch variations, etc.).

[0063] After staining the section of biological tissue, at step 220, an input image of the stained section is captured (e.g., input image 106). In some examples, the input image may represent a digital image that is represented in a color model, such as an RGB (Red-Green-Blue) color model. In these examples, each pixel of the input image may be defined by multiple channels, one channel for each color in the color model, and each pixel may include an intensity value for each of the multiple channels. In some examples, the input image may represent an RGB imagedefined by three color channels (a red color channel, a green color channel, and a blue channel) and each of the pixels in the input image may include three intensity values (one for each of the three color channels).

[0064] The input image may be captured using any technique designed for capturing images of biological tissue section. In one implementation, the stained section of biological tissue may be placed on a stage of a microscope such as microscope 114 (e.g., below the objective lenses of the microscope) and the input image may be captured by a camera coupled to the microscope. In some examples, the input image may be captured as part of an automated image scanning process. For example, the input image may be captured using a specialized slide scanning microscope designed for automatic slide scanning. Variable equipment (e.g., microscope) calibrations may affect the pixel values of the input image captured by the equipment (e.g., affecting the brightness, contrast, color balance, etc., of the input image), introducing further variability in the appearance of stains in a section of biological tissue (e.g., in addition to the inter- laboratory and intra-laboratory stain variability described previously).

[0065] In FIG. 2, steps 210 and 220 are included in process 200. However, in some examples, protocol 200 may begin with receiving the input image of the biological tissue section, which may have been previously prepared (e.g., as described at steps 210 and 220).

[0066] While variability in tissue staining or equipment calibrations do not generally affect a pathologist’s ability to manually analyze a stained tissue section or an image of a stained tissue section, such variability is an obstacle to automating the tissue section analysis, as trained models may benefit from consistent inputs to work effectively. To solve for stain variability (or calibration variability that affects the appearance of stains in an image), a stain-normalization process may be applied to the input image of the stained biological tissue section at step 230,yielding a stain-normalized image of the biological tissue section (e.g., that is invariant to inter- laboratory staining protocol variance, reagent and batch variance, equipment calibration variance, etc.). The disclosed tissue analysis facility may apply any type of stain normalization process to the stained biological tissue section. FIG. 3 provides one such exemplary process, which will be described in greater detail later on.

[0067] In addition to the stain variability described above, images of stained biological tissue sections may be subject to scale variability (e.g., based on an amount of magnification applied to an image and / or to a tissue section when an image is captured). Such variability is an obstacle to automating tissue section analysis (e.g., especially in instances in which size-relevant features or measurements are to be detected by a trained model). Responding to this, at step 240, the tissue analysis facility may apply a magnification normalization process to the input image, yielding a magnification-normalized image of the biological tissue section. The tissue analysis facility may apply any type of scale normalization process to the stained biological tissue section. FIG. 4 provides one such exemplary process, which will be described in greater detail later on. Step 240 describes the scale normalization process as being applied to the input image. This may include, in some examples, the scale normalization being applied to the stain normalized image generated from the input image at step 220.

[0068] After generating a normalized image of the biological tissue section (e.g., normalized for stain and / or scale), the disclosed system may analyze the normalized image to identify one or more morphologic structures and / or features within the biological tissue section. The disclosed tissue analysis facility may analyze the normalized image using any type or form of trained model and / or protocol. FIGS. 6–8 will provide some exemplary processes for analyzing a normalized image.STAIN NORMALIZATION FRAMEWORK

[0069] Process 300 in FIG. 3 is directed to identifying, for an image of a biological tissue section, pixel intensity contributions for each of multiple stains used to stain the biological tissue section. The result of this process may be used in a variety of ways (e.g., as will be discussed in connection with FIGS.5–8). Process 300 begins with step 310, where the system (e.g., the tissue analysis facility 104 in FIG. 1) identifies, from an input image (e.g., the input image 106) of a biological tissue section (e.g., the biological tissue section 108), multiple stain channels. Each stain channel may indicate a different stain applied to the biological tissue section and may include a range of pixel intensity values corresponding to a range of concentrations for that stain. In some examples, the range of pixel intensity values for each stain channel may include a pixel intensity value for each pixel of the input image.

[0070] In some examples, the input image may represent a digital image that uses a color model, such as the RGB color model, to represent colors. In these examples, as mentioned in connection with FIG.2, each pixel of the input image may be defined by multiple color values, one for each color included in the color model. Each color value may indicate an intensity of the color indicated by the value. As a specific example, in examples in which the input image uses the RGB color model, each pixel of the input image may be defined by three color values: a red value (indicating an intensity corresponding to a red color for the pixel), a green value (indicating an intensity of a green color for the pixel), and a blue value (indicating an intensity of a blue color for the pixel). In some examples, the digital image may be associated with multiple color channels, one for each color of the color model (e.g., a red channel, a green channel, and a blue channel in examples in which the color model is an RGB color model), where a color channel includes, for each pixel of the digital image, the color value corresponding to the color indicated by the colorchannel. In some such examples, the system may use these color channels to generate the disclosed stain channels and / or may generate and / or use the disclosed stain channels instead of generating and / or using the color channels. It should be appreciated that, while for ease of description examples are discussed herein in the context of RGB color models and color channels, embodiments are not limited to any particular color model and may operate with other color models and color channels (e.g., CMYK color models having four color channels for cyan, magenta, yellow, and black, or other color models with other color channels, which may include other additive and / or subtractive color models). In some examples, a range of pixel intensity values for a stain channel may include or represent a range of grayscale intensity values.

[0071] FIG. 4 depicts an exemplary additional block diagram of the system 100 in which the system identifies a first stain channel 400 and a second stain channel 402 from the input image 106. In the embodiment depicted in FIG.4, the first stain channel 400 indicates a first stain 404 and captures a range of pixel intensity values 406 for the input image 106 corresponding to stain concentrations for the first stain 404. The second stain channel 402 indicates a second stain 408 and captures a range of pixel intensity values 410 for the input image 106 corresponding to stain concentrations for the second stain 408.

[0072] The system may identify stain channels corresponding to any stain, depending on which stains were applied to the biological tissue section (e.g., as described at step 210 of FIG. 2). In one embodiment, the biological tissue section may have been stained with hematoxylin and eosin. In this embodiment, the system may identify a hematoxylin channel, corresponding to the hematoxylin stain, and an eosin channel, corresponding to the eosin stain. Additional examples of stains for which a stain channel may be identified include, without limitation, Alcian blue stains, immunohistochemical stains, Toluidine blue stains, Masson’s trichrome stains, fluorescent stainsthat binds to DNA (e.g., 4’,6-Diamidino-2-Phenylindole, Propidium Iodide, Hoechst Stains, etc.), silver stains such as Grocott’s methenamine silver, Warthin-Starry stains, Giesma stains, etc.

[0073] The system may identify stain channels from the input image in a variety of ways. In some examples in which the input image uses a color model, the system can identify the stain channels by deconvolving the color values of the input image (e.g., the multiple color values of each pixel) into the stain channels (e.g., into values that fall within each of the stain channels). In some examples, this may involve deconvolving a first number of pixel values (e.g., one for each color in the color model) to a second number of pixel values (e.g., one for each stain used to stain the biological tissue section). As a specific example, the input image may represent an RGB input image and the biological tissue section may have been stained with two stains: hematoxylin and eosin. In this specific example, the system may, for each pixel of the input image, deconvolve three color values (a red channel, a green channel, and a blue channel) into two stain values (a hematoxylin value and an eosin value), yielding two values for each pixel (one for each stain channel) instead of three values for each pixel (one for each color channel) and two stain channels instead of three color channels. In some examples, the system may rely on a novel color deconvolution algorithm that produces stain unmixing results much faster (e.g., 100x faster) than conventional deconvolution protocols, permitting real-time analysis of stain unmixing results. This algorithm is described in greater detail below within the section of this application labeled “Staining Invariance” (e.g., in connection with the discussion of FIG. 10).

[0074] At step 320, the system may generate, for each of the stain channels identified at step 310, multiple images based on relative stain concentration. Each of the images generated for a stain channel may capture pixel intensity values within a different segment (e.g., subrange or threshold class) of the stain channel’s range of pixel intensity values. In some examples, thesegments may be organized based on the relative pixel intensity. In one example, the range may be divided into three segments: a stain-background segment, a stain-high segment, and a stain-low segment. The stain-background segment may include values indicative of a background level of the stain indicated by the stain channel. The stain-high segment may include pixel intensity values at the high end of the range (e.g., pixel intensity values that are higher than the pixel intensity values of the stain-low segment). And the stain-low segment may include pixel intensity values at the low end of the range (e.g., pixel intensity values that are above the values indicative of the background level of the stain and that are low relative to the pixel intensity values of the stain-high segment). In some examples, the range may further include one or more stain-moderate (e.g., stain- medium) segments, with a subrange of pixel intensity values that fall between the subranges of the stain-high segment and the stain-low segment.

[0075] In some examples, one or more (e.g., each) of the images generated for a stain channel may be binarized (e.g., to yield a black-and-white image). An image may be binarized in a variety of ways. In one example, an image may be binarized by assigning values of the image that are at and / or above a threshold value to a first binary value (e.g., 1) and values of the image that are at and / or below the threshold value to a second binary value (e.g., 0). In some examples, the threshold value for an image may be dynamically determined based on the values of the image. For example, the threshold value may represent a value corresponding to a median and / or mode of the values of the image. Binarizing the images may result in a variety of benefits. In some examples, binarizing the images may help yield crisply delineated borders (e.g., boundaries), which may be used for the distance transform and / or segmentation processes that will be described later (e.g., in connection with FIGS. 5–8). Processes for generating the various segments of therange are described in greater detail below (e.g., in connection with the section of this application labeled “Staining Invariance.”)

[0076] FIG. 4 depicts an embodiment in which four images are generated for the first stain channel 400 (image 412 with a stain-high segment of range 406, image 414 with a stain- medium segment of range 406, image 416 with a stain-low segment of range 406, and image 418 with a stain-background of range 406) and four image are generated for the second stain channel 402 (image 420 with a stain-high segment of range 410, image 422 with a stain-medium segment of range 410, image 424 with a stain-low segment of range 410, and image 426 with a stain- background of range 410). Exemplary images, generated for a hematoxylin channel and an eosin channel of an input image shown in FIG. 22, are illustrated in FIGS. 23–30 as discussed in the section of this application labeled “Section S3. ViCE’s binarization across a diverse dataset.”

[0077] In some examples, an image (e.g., each of the images) generated for a stain channel may capture one or more dynamics of the varying levels of pixel intensity across the pixel intensity values of the image. As one example, the image may capture spatial dynamics indicating variations in pixel intensity across two or more regions of the image. Once identified, these spatial dynamics may be used in a variety of ways. For example, spatial dynamics may be used for segmentation (e.g., to partition an image into distinct homogeneous regions), mathematical morphological transformations (e.g., to enhance or suppress various features in an image using operations such dilation, erosion, opening, and / or closing), and / or morphological feature identification (e.g., to extract and analyze shape-related features from an image).

[0078] As another example, the image may capture contrast dynamics indicating a variation between one or more of the relatively highest pixel intensity values within the image and one or more of the relatively lowest pixel intensity values within the image. Once identified, thesecontrast dynamics may be used in a variety of ways. For example, contrast dynamics may be used for stain normalization (e.g., the images generated for a stain channel may be normalized by shifting the values of an image based on the relative contrast in each image).

[0079] As another example, the image may capture brightness and / or saturation dynamics corresponding to relative quantities of pixel intensity values designated as a high pixel intensity value and pixel intensity values designated as low pixel intensity values. Once identified, these brightness and / or saturation dynamics may be used in a variety of ways. For example, brightness and / or saturation dynamics may be used when performing a distance transform (e.g., to identify the distance of pixels in the image to a nearest edge). The distance transform may be used in a variety of contexts (e.g., for nuclei parameterization as will be discussed in connection with the steps of FIG.5).

[0080] As another example, the image may capture distribution dynamics indicating a distribution of pixel intensity values captured by the image. Once identified, these distribution dynamics may be used in a variety of ways. For example, distribution dynamics may be used for stain normalization and / or feature identification.

[0081] Returning to FIG. 3, at step 330, the system may output one or more stain- normalized images (e.g., normalized image 110 in FIG. 1) that are based on one or more of the images generated at step 320 for each stain channel. In some examples, the one or more stain- normalized images may include or represent an image of a lower bound or corrected lower bound (e.g., with one or more of the features described below in connection with FIG. 6). Additionally or alternatively, the one or more stain-normalized images may represent or include one or more of the images generated for each stain channel at step 320. In one embodiment, one or more of theimages generated at step 320 may be digitally assembled to yield the one or more stain-normalized images. SCALE NORMALIZATION FRAMEWORK

[0082] Process 500 in FIG. 5 begins with step 510, where the system (e.g., the tissue analysis facility 104 in FIG. 1) may identify nuclei across multiple cells within an image of a section of biological tissue (e.g., an input image such as the input image 106 and / or a normalized image generated from the input image such as normalized image 110 in FIG. 1). The system may identify the nuclei in a variety of ways. In some examples, the system may identify the nuclei from one or more stain channels identified from an input image and / or one or more stain-normalized images (e.g., a stain-high image) generated for the one or more stain channels (e.g., as described in connection with step 320 of FIG. 3). In one such example, the system may identify the nuclei from a stain channel that indicates a stain that stains for nuclei (e.g., a hematoxylin stain, a fluorescent stain that binds to DNA, a Feulgen stain, a methyl green stain, a Giemsa stain, etc.) and / or from one or more stain-normalized images generated for the stain channel that indicates the stain that stain for nuclei (e.g., a hematoxylin-high image, a Giemsa-high image, etc.).

[0083] In certain examples, the system may identify the nuclei (e.g., from the one or more stain-normalized images) using mathematical morphology. In one specific example, the system may (1) identify (e.g., using mathematical morphology) discrete structures within the one or more stain-normalized images, (2) determine a distribution of relative sizes for the discrete structures, (3) determine a peak distribution within the distribution of relative sizes, and (4) determine that the structures within the peak distribution are nuclei. This process for identifying nuclei may be effective because nuclei may be of a relatively uniform size.

[0084] At step 512, the system may set a length scale based on a size of one or more of the nuclei. The system may determine the size of the nuclei in a variety of ways. In examples in which a level of magnification applied to the biological tissue is known, the size of the nuclei may be measured using area and / or direct measurements. In some examples (e.g., in examples in which a level of magnification is not known), the system may determine the size of the nuclei using an intensity map (e.g., representing the distribution of pixel intensity values across an image associated with the nuclei such as the one or more stain-normalized images discussed in connection with step 510). The term “intensity map” may refer to any type or form of image that represents the distribution of intensity values across an image. The intensity map may visualize the distribution of intensity values in a variety of ways. In some examples, the intensity map may visualize varying magnitudes of intensity values with varying shades (e.g., of color or grayscale) and / or with varying elevations. For example, the intensity map may represent a heat map in which the magnitudes of the intensity values are visualized with different shades of color and / or a geographic hill in which the magnitudes of the intensity values are visualized with different elevations. In one embodiment, an intensity map may simply refer to a set (a distribution) of intensity values corresponding to an image.

[0085] The intensity values, represented in the intensity map, may correspond to measures of intensity for a variety of characteristics determined for a pixel. In some examples, the intensity values may represent a distance from a designated (e.g., nearest) boundary (e.g., where the highest intensity values correspond to pixels that are the farthest from a boundary and the lowest intensity values correspond to pixels that are the closest to the boundary). In some such examples, the intensity map may be generated for the one or more stain-normalized images by applying a distance transform to the one or more stain-normalized images. The term “distancetransform” may refer to an operation that assigns pixel values (e.g., each pixel value) within an image (e.g., a binary stain-normalized image) to the nearest pixel with a different value (e.g., distance to the nearest background pixel), resulting in a grayscale image in which each pixel’s intensity value corresponds to its distance from the pixel’s nearest boundary (e.g., edge).

[0086] One specific example of identifying nuclei size from an intensity map will be described in greater detail below in connection with the section of this application labeled “Section S4: Determining ViCE’s Length Scale” in connection with FIGs. 31–35. As will be described in this section, in this specific example the system identifies pixels of relatively high Hematoxylin concentration from a raw input (e.g., RGB) image. FIG. 31 depicts an exemplary raw input image of a tissue section stained with Hematoxylin and FIG. 32 depicts an exemplary H-high image generated for a Hematoxylin stain channel identified for the raw input image. Then, the system reduces noise of the image (e.g., to fill in small gaps or holes in the image of the tissue). The system may reduce the noise of an image using any type of operation. In this specific example, the system reduces noise by performing a dilation operation on the H-high image followed by an erosion operation using a cross-shaped kernel. After reducing noise in the H-high image, the system performs a distance transform (e.g., calculating the distance of each pixel within the H- high image to a nearest edge) to generate a distance map (e.g., depicted in FIG. 34). The distance map may represent a local skeleton of the H-high image (e.g., corresponding to the pixels with the maximum distance values generated by the distance transform). In some examples, the distance values of the distance map may be plotted as a histogram (e.g., as depicted in FIG.35). Then, the most common distance (e.g., peak or maximum) identified in the histogram may be determined to be the average width of nuclei. One advantage of this specific example is that the system may determine the size of nuclei without completely demarcating individual nuclei. This advantagemay be especially useful in images that capture a series of connected nuclei. By measuring minimum distance to an edge (e.g., boundary), the calculation may calculate an average size for a series of nuclei, even if the nuclei are connected.

[0087] After determining the size of one or more of the nuclei captured by the input image, the system may set the length scale in a variety of ways. In some examples, the system may set the length scale based on an average size of the nuclei. For example, the system may set the length scale based on a mean size of multiple (e.g., all) of the identified nuclei, a median size for the identified nuclei, and / or a mode size of the identified nuclei. In one example, the average size may be determined using distance values of a distance map (e.g., plotted as a histogram), as described in the specific example corresponding to FIGs. 31–35.

[0088] After setting the length scale, the length scale may be implemented by functionalizing mathematical morphology, image analysis, and / or image processing parameters related to pixel length and pixel count to perform morphological operations and image analysis and / or processing operations that are invariant across a range of input image scales, resolutions, magnifications, and / or other parameters. At step 514, the system may scale, for a morphological structure (other than a nuclei) that is represented in the image, a parameter related to length using the length scale set at step 512. As a specific example, the image of the biological tissue section may represent an image of a section of intestinal tissue, the morphological structure may represent a villus-crypt pair, and the parameter related to length may be a villus-height to crypt-depth ratio for the villus-crypt pair. By using a normalized length scale for input image, some examples of the disclosed framework may dynamically scale mathematical morphology, image analysis, and image processing parameters. SEGMENTATION FRAMEWORK

[0089] Process 600 in FIG. 6 begins with step 610, where the system (e.g., the tissue analysis facility 104 in FIG. 1) generates, from an input image of a biological tissue section (e.g., input image 106), a stain channel (e.g., first stain channel 400 in FIG. 4) indicating a stain of interest. In some examples, the stain channel may indicate a stain that stains for nuclei. In these examples, the stain channel may indicate any type or form of stain that stain for nuclei (e.g., a hematoxylin stain, a fluorescent stain that binds to DNA, a Feulgen stain, a methyl green stain, a Giemsa stain, etc.). In some examples, the stain channel may represent a set of multiple stain channels (e.g., first stain channel 400 and second stain channel 402 in FIG.4).

[0090] In some examples, the system may generate the stain channel using one or more of the stain normalization features described as part of the stain normalization framework described in connection with FIGS.3 and 4. As described in connection with FIGS.3 and 4, such features may include (1) identifying multiple stain channels for an input image (e.g., first stain channel 400 and second stain channel 402) and (2) generating a set of images for each of the stain channels (e.g., images 412–426 in FIG.4). Each image for a stain channel may include a different segment of the stain channel’s range of pixel intensity values. As shown in FIG. 4 and described in connection with FIG. 4, a set of images generated for a stain channel may include a stain-high image, a stain-medium mage, a stain-low image, and a stain-background image.

[0091] At step 620, the system may determine a lower bound from the stain channel (i.e., the stain channel indicating the stain that stains for nuclei). In some examples, the lower bound may represent pixel intensity values from a stain-high image generated for the stain channel (e.g., image 412 generated for first stain channel 400 as described in connection with FIG. 4). In one embodiment, the lower bound may represent all pixel intensity values from the stain-high image generated for the stain channel.

[0092] In some examples, at step 620, the system may also determine an upper bound. In these examples, the system may determine the upper bound from multiple stain channels identified at step 610 (e.g., each of the stain channels). In one such example, the upper bound may represent pixel intensity values from the images generated for the multiple stain channels. In one specific example, the upper bound may represent all of the pixel intensity values captured by all of the images generated for each of the stain channels, except for any stain-background images generated for the stain channels.

[0093] At step 630, the system may create a segmentation map based on the lower bound. Then, at step 640 the system may identify a morphological structure and / or feature of the biological tissue section from the segmentation map.

[0094] The system may create the segmentation map from the lower bound in a variety of ways. In some examples, the system may create the segmentation map from a corrected lower bound. In these examples, the system may (1) apply a correction term to the lower bound to determine a corrected lower bound and (2) create the segmentation map from the corrected lower bound. In one example, this correction term may be based on the upper bound discussed at step 620 and further described in connection with the section of this application labeled “Upper bound tissue boundary estimate.” The system may create the segmentation map in a variety of ways, as will be discussed in greater detail in connection with FIG. 7 and again in connection with the section of this application labeled “Upper bound tissue boundary estimate.” ADDITIONAL SEGMENTATION FRAMEWORK

[0095] Process 700 in FIG. 7 beings with step 710, where the system (e.g., the tissue analysis facility 104 in FIG. 1) identifies, from an input image of a biological tissue section (e.g., input image 106), multiple stain channels (e.g., first stain channel 400 and second stain channel402 in FIG.4), each indicating a different stain applied to the biological tissue section. The systems may identify the stain channels in a variety of ways (e.g., as explained in connection with step 310 of FIG. 3 and with FIG. 4). Then, at step 720, the system may normalize, based on the stain channels, each pixel intensity value for the input image one or more normalized values to yield one or more stain-normalized images of the biological tissue section. The system may normalize the pixel intensity values in a variety of ways (e.g., as explained in connection with steps 320 and 330 of FIG. 3 and FIG. 4). In addition, or as an alternative, to normalizing the pixel intensity values, the system may normalize the scale of the input image (e.g., using one or more of the features or processes described in connection with FIG. 5) to yield one or more scale-normalized images of the biological tissue section. Finally, at step 730, the system may identify a morphological structure of the biological tissue section from the one or more stain-normalized (and / or scale-normalized) images of the biological tissue section (e.g., normalized image 110 in FIG.1).

[0096] The system may identify the morphological structure in a variety of ways. In some examples, the system may generate a segmentation map from the stain-normalized image and may identify the morphological structure from the segmentation map. In one example, the system may generate the segmentation map based on the lower bound and / or the corrected lower bound (e.g., which may have been generated as described in connection with the steps of FIG.6).

[0097] In certain examples, the system may segment the lower bound and / or corrected lower bound by identifying an intensity center of the lower bound and / or corrected lower bound. This intensity center may represent a distance intensity (e.g., based on a distance value corresponding to a pixel’s distance from a boundary). In some examples, the intensity center may represent the center of a morphologic structure or feature. In other examples, a morphologicstructure or feature may be identified based on its position relative to the intensity center (e.g., its proximity to the intensity center). The generating and use of intensity center is described in greater detail below in the sections of this application labeled “ViCE’s ridge-based skeletonization,” “Stem: Smoothing and villus branch cuts,” “Head: Villus height correction term,” and “Determining ViCE’s length scale.”

[0098] In addition to, or instead of, identifying the morphological structure from a segmentation map, in some examples, the system may determine a quantitative histological measure of a morphological structure using a segmentation map. In some such examples, the system may rely on a segmentation map to identify various points of reference within a morphological structure from which to measure. Additionally or alternatively, the system may rely on a normalized length scale (e.g., such as the normalized length scale described in connection with step 520 of FIG.5) in determining the quantitative histological measure. A specific example of determining a quantitative histological measure of a morphological structure will be described in connection with step 830 of FIG.8.

[0099] In certain examples, identifying the morphological structure may include determining whether the morphological structure satisfied a quality metric (e.g., determining that the morphological structure is well-oriented). Exemplary quality control metrics are discussed in greater detail below in connection with the section labeled “Quality control metrics.” In some such examples, the system may, in response to determining that the morphological structure satisfies the quality control metric, generate a user interface, with a display of the morphological structure, that (1) visually draws attention to the morphological structure and / or (2) visually indicates that the morphological structure satisfies the quality metric. In some examples, the user interface may be placed in a digital queue of user interfaces, each of which includes a display of a differentsection of the biological tissue. In these examples, the system may prioritize the user interface within the queue in response to determining that the morphological structure displayed in the user interface satisfies the quality metric.

[0100] In some examples, the system may also determine that an image of an additional section of the biological tissue either does not include any instance of the morphological structure or includes an additional instance of the morphological structure that does not satisfy the quality metric. In these examples, the system may preclude the image of the additional section from a digital queue of user interfaces (e.g., the queue of interfaces selected and / or prioritized for a pathologist’s review). VILLUS-CRYPT SEGMENTATION FRAMEWORK

[0101] Process 800 in FIG. 8 begins with step 810, where the system (e.g., the tissue analysis facility 104 in FIG. 1) identifies, from an input image of an intestinal tissue section (e.g., the input image 106 in FIG. 1), a hematoxylin stain channel (e.g., first stain channel 400), indicating a hematoxylin stain applied to the intestinal tissue section and an eosin stain channel (e.g., second stain channel 402 in FIG. 4), indicating an eosin stain applied to the intestinal tissue section. The systems may identify the stain channels in a variety of ways (e.g., as explained in connection with step 310 of FIG.3 and with FIG.4). Then, at step 820, the system may normalize, based on the stain channels, each pixel intensity value for the input image to a normalized value to yield one or more stain-normalized images of the intestinal tissue section. The system may normalize the pixel intensity values in a variety of ways (e.g., as explained in connection with steps 320 and 330 of FIG. 3 and with FIG.4). In addition, or as an alternative, to normalizing the pixel intensity values, the system may normalize the scale of the input image (e.g., using one ormore of the features or processes described in connection with FIG.5) to yield a scale-normalized image of the intestinal tissue section.

[0102] Finally, at step 830, the system may identify one or more villus-crypt pairs of the intestinal tissue section from the one or more stain-normalized (and / or scale-normalized) images of the biological tissue section (e.g., normalized image 110 in FIG. 1). In some examples, the system may identify the villus-crypt pairs by identifying a central line, corresponding to a section of the intestinal tissue, from which both a first line, corresponding to a villus, and a second line, corresponding to a crypt, are projected in different (e.g., opposite) directions. In some examples, the central line may correspond to an intensity center (e.g., with one or more of the intensity center features described above in connection with FIG. 7). In one embodiment, this intensity center may correspond to a center of intensity for a lower bound and / or a corrected lower bound generated from one or more of the one or more stain-normalized images.

[0103] In addition to, or instead of, identifying the villus-crypt pair, in some examples the system may determine a quantitative histological measure of the villus-crypt pair. For example, the system may determine a villus-height to crypt-depth ratio of the villus-crypt pair. EXEMPLARY TISSUE ANALYSIS FACILITY

[0104] This section provides details relating to an example of a tissue analysis facility (referred to in this section as ViCE) that may implement some of the techniques described herein. It should be appreciated that the example below is merely illustrative of ways in which techniques described herein may be implemented, that other embodiments are possible, and that embodiments are not limited to operating in accordance with the example below.

[0105] ViCE codifies the human interpretation of images of stained biological tissue (e.g., hematoxylin (H) and eosin (E) images). Let us say there exists a human observer whose interpretation of HE images can be modeled by the following: I(x,y)|input = {IH(IRGB), IE(IRGB)} = {IH, IE} { (IH, IE, ℓrelative)}where ISis a transformation of IRGBthat results in an image with intensity values corresponding toinferred stain S concentration values, and ( ) represents the observer’s interpretation of inferredstain concentrations relations with a length scale, ℓrelative. An observer’s interpretation is staining protocol and scale agnostic and therefore nonnormative. Similarly, scaling is proportional to the size of objects within each image and therefore is magnification-invariant. Staining Invariance

[0106] Physical tissue properties determine stain affinity (As). Staining protocol parameters (λ) like dye concentration and dye set-time directly influence stain concentration (Is). In one embodiment, assuming independence between stain affinity and an effective staining protocol parameter (λeff), these dynamics may be modeled as follows:where ||Is(x,y)|| denotes min-max normalization of the respective stain’s concentration obtained through color deconvolution. FIG. 10 depicts an exemplary relationship between concentration and stain affinity.

[0107] A proportionality between stain affinity and concentration follows:

[0108] In one embodiment, it is assumed that such a proportionality can be adequately described by a positive linear translation (α) on and a translation (β) of stain affinity:

[0109] In this example, a function (F) invariant to positive affine transformations, by definition, therefore maps the LHS and RHS of Equation (2) to sets with one-to-one correspondence:

[0110] Otsu’s thresholding method (Õ) satisfies this, where (n, x) denotes the resulting histogram of image I(x,y) and c, the number of threshold classes: (n, x) = hist(I(x,y))

[0111] Therefore, the staining protocol independent term, As(x,y), can be approximated via the staining protocol dependent term, Is(x,y): Õ(hist(exp(||Is(x,y)||)); c) Õ(hist(As(x,y)); c).

[0112] In one embodiment, the disclosed framework sets the number of threshold classes c to be four. In examples in which two dyes are used for the sample, this produces eight binary images that capture relative stain concentration information resilient against staining protocol discrepancies between lab technicians and laboratories and form the basis of all segmentation maps. These images are labeled as the following: {{H-high, H-mid, H-low, H-bkgd}, {E-high, E-mid, E-low, E-bkgd}}. Objects formed from the collection of adjacent, similarly mapped pixels can then be grouped and described by their H-class and E-class and morphological features. ViCE’s General Model

[0113] ViCE’s general approach can be described as the following: I(x,y)input= IRGB{ IH(IRGB), IE(IRGB) } = { IH, IE} {H(AH), E(AE) } { H( (IH)), E( (IE)) } = { H, E }segmentation maps and feature extractions where (Z(x,y)) = exp( min-max feature scaling of Z(x,y) ) = = exp( [ Z(x,y) - min{Z(x,y)} ] / [ max{Z(x,y–)} - min{Z(x,y)} ] ),

[0114] In some examples, given RGB input channels, ViCE uses its improved colordeconvolution method to effectively “de-stain” the RGB input channels into HE channels. Theindividual HE channels, {IH, IE}, have intensities proportional to stain concentration.

[0115] In some examples, ViCE extracts staining information stable across stainingprotocols by applying a function, ( ), where ( (S)) = (AS) and (S) AS. In these examples,ViCE may preferentially utilize functions agnostic to translation and scale transformations. A tetraclass threshold on (S) via a Kittler-Illingworth minimum error thresholding method (e.g., Otsu’s method) satisfies these preferences, while also providing intuitive four classes, interpreted as high, medium, low, and background stain concentration classes. Otsu’s threshold results are translation and scale agnostic because translations and uniform scalings do not affect relative intra- class variance and therefore will not change the class of any pixel. Therefore, within the confines of our model according to some embodiments, the sets of binary images: {S-high, S-mid, S-low, S-bkgd} and {AS-high, AS-mid, AS-low, AS-bkgd}, are equivalent and contain staining information that is independent of staining protocol.

[0116] In the embodiment focused on in this section, applying this process to both stain channels results in a set eight binary images, i.e. {{H-high, H-mid, H-low, H-bkgd}, {E-high, E- mid, E-low, E-bkgd}}. In some examples, every ensuing method within ViCE uses information related to stain concentration with information from these eight binary images in lieu of any color deconvolution result and raw RGB data. In some examples, every segmentation map can bederived entirely from these eight images and Mathematical Morphology (MM) methods (e.g., with an exception, in one embodiment, for the crypt segmentation map, which may use IHwithin a new MM filter found within ViCE).

[0117] In some examples, with {H-high} and granulometric methods, ViCE estimates the diameter of the average nuclei, , within each image to define a length scale, i.e. [ℓ] . The instant application identifies nuclei as a basis for defining length scale due to their relative invariance in size.

[0118] By parameterizing every variable related to length by , every operation within ViCE can be inherently scaled with image resolution, magnification, and compression. Most notably, the size of every structuring element within every MM methods can be functions of .

[0119] Due to the use of MM methods, especially ones involving iterative operations with structuring elements of increasing size, e.g. alternating sequential filters, the length scale need not be exact. Although the ideal value for is the average nuclei diameter, one range of acceptable values is the following: {pixel width(nucleus) 3 px ≤ < pixel width(crypt) ÷ 3}, or{ pixel width(nucleus) 3 px ≤ 2 × pixel width(nucleus)} (e.g., { | ℕ ≡ 1(mod 2) 3 ≤ 2×width(nucleus) })

[0120] Through the extensive use of MM methods, ViCE generates segmentation maps, extracts features, and evaluates tissue quality from staining protocol independent images of relative stain concentration levels with parameterized operations including variables that individually scale by the size of each image’s nuclei. Given a set of HE images with ranges of varied and unknown magnifications, resolutions, imaging devices and methods, and tissue stainingprotocols, ViCE can automatically deploy quantitative histology methods via scale and staining protocol agnostic methods. Mathematical Morphology

[0121] Mathematical morphology allows one to process and analyze digital images based on the size and shape of geometrical structures. Morphological operations used in morphological filters, feature extractions, and smoothing techniques, use structuring elements of a priori determined size, shape, and offset. These operations tend to simplify image data while preserving their essential shape characteristics.

[0122] ViCE primarily uses disk-shaped structuring elements where a radii (e.g., every radii) directly scales with the radius of the average nucleus found in each image. This allows for direct physical interpretations of all morphological techniques, especially since many morphological methods derive from the erosion and dilation of an object by a structuring element. Having clear digital to physical analogs can be a powerful tool for exploratory analysis. Tissue Segmentation

[0123] As mentioned previously, in some examples, one or more (e.g., all) of the segmentation maps described herein may originate from a set of binary images representing classes of relative stain concentration: {{H-high, H-mid, H-low, H-bkgd}, {E-high, E-mid, E-low, E- bkgd}}. In these examples, every pixel can be mapped to a unique H- and E- class. Objects formed from the collection of adjacent, similarly mapped pixels can then be grouped and described primarily by these classes. In essence, in one embodiment ViCE converts RGB images into eight binary images and one length constant. This binarization allows for improved performance, especially with MM methods. ViCE’s Improved Color Deconvolution Method

[0124] Color deconvolution “de-stains” RGB image channels by producing an image whose channels correspond to the specific stains with pixel values proportional to stain concentration at that particular pixel location, instead of red, green, and blue channels and their respective intensities (Ruifrok et al.2001; Landini et al. 2020).

[0125] ViCE introduces a beneficial color deconvolution method that produces results at least two orders of magnitude faster than the currently available MATLAB and ImageJ methods. Through matrix manipulations alone, the disclosed framework can reduce the number of function callbacks by a number equal to the number of image pixels, thereby greatly reducing execution time without precision loss relative to the original method. Given a set of 100 GB images, a five- hundred-fold improvement in execution time between methods was observed and scales by ~log(numPixels). Precise Tissue Segmentation Can Be Unnecessary

[0126] The initial segmentation map is that of the tissue from the background. Direct segmentation often relies on the nontrivial partitioning of non-viable tissues of varying size, shape, and stain concentration from tissue. In some examples, ViCE instead approaches tissue segmentation indirectly by using upper bound (bwUB) and lower bound (bwLB) estimates of the true tissue boundary throughout its methods. bwUB is the segregation map that only excludes pixels within both {H-bkgd} AND {E-bkgd}, thereby including all pixels with any signal in either stain channels. Through MM methods, holes from tissue damage are filled, but may also fuse adjacent objects together. These methods may enable bwUB to be a superset of the true boundary. bwLB is the segregation map that is the smoothed amalgamation of nuclei, obtained through morphological methods and derived from {H-high} due to the nucleus’s strong H signal. Because nuclei are within cells, this method ensures bwLB to be a subset of the true boundary.

[0127] In some examples, ViCE defines the boundary in terms of bwUB and bwLB. While the perimeter of bwUB significantly overlaps with the true boundary, bwUB often fuses adjacent villi. While the villus length measured through bwLB will always be lower than the true value, bwLB does not fuse adjacent villi brushed against one another. Fusing within bwLB uses overlapping enterocyte nuclei, which are located near the basal membrane, formed during slide preparation. Although bwLB systematically underestimate villus heights, the objects commonly fused in bwUB do not pass quality control procedures, thereby removing all adjacent villi within contact of one another. Lollipop Stems, Heads, and Bases

[0128] In some examples, ViCE depicts all VC candidates as lollipops. Lollipop stems outline villus skeletons obtained through ViCE’s ridge-based skeletonization of the smoothed bwLB, bwSmooth. Lollipop heads result from applying necessary villus height correction terms, whose values equate to lollipop head radii. Lollipop bases correspond to the crypt search window from which each lollipop’s best local crypt is determined. ViCE’s Ridge-Based Skeletonization

[0129] ViCE uses ridge-based skeletons derived from the input image’s distance transform. The distance map’s ridge lines are then computed by taking the Laplacian of the Gaussian (LoG) of the distance map, which approximates a second derivative. Pixels with negative values are pixels where travel in any direction within the distance map results in downhill travel. We therefore binarize the image through a negative values threshold. After pruning spurious branches, ViCE’s ridge-based skeletonization method is complete.

[0130] ViCE’s skeletons remain homotopic with standard methods and provide unique additional utilities. By mapping any of its values to the input image’s distance transform its firstderivative, direct measurements of structure width and local width changes. Skeletons derived from other methods cannot guarantee accurate width measurements and will underestimate width values, if not located exactly at the ridge. Additionally, these skeletons are less unlikely to obtain measurements of local width variations, as any deviation from its ridges yields no useful information. 4.3.2 Stem: Smoothing and Villus Branch Cuts

[0131] Skeletonizing smooth objects create fewer spurious branches than rough objects. As such, ViCE skeletonized a smoothed bwLB, and thus producing bwSmooth. After skeletonization, ViCE automatically separates potential villi from the raw skeleton. Because ViCE’s skeletons can directly map the rate of change of its distance map, a natural threshold appears, where ||grad(dist(image))|| > 0.5.

[0132] The smoothing method includes two Savitz-Golay filters applied to the ordered address values of bwLB’s boundary’s row and column values respectively. The filters include successive 2nd-degree polynomial fits along a frame length of forty nuclei or roughly four crypt widths. Fitting with such low order polynomials across forty nuclei long frames effectively smooths jagged segments that are less than a dozen nuclei wide into continuous regions. This smooths out divots and cracks formed from goblet cells, poorly oriented tissue, and damaged tissue along bwLB’s perimeter. Additionally, this method does not significantly impact villi as both the villus tip, the region where the villus base merges into the tissue base, and the space between adjacent villi include parabolic curves, which minimizing the residual from the filter’s 2nd order polynomial fittings. Head: Villus Height Correction Term

[0133] ViCE then finds the villus height correction terms by mapping the pixel addresses of villus tips onto the distance transform of bwUB. This provides the pixel distance of the villus tip location to the nearest edge within the upper bound estimate, bwUB. This distance is the villus height correction term and is depicted as the radius of the lollipop head. Stand: Demarcation Parameters

[0134] After finding the lollipop stems and heads, its base is found through MM and trigonometry. This produces every lollipop’s demarcation width and demarcation angle, used in determining the width and angle of the crypt search windows used when finding the best crypt near every lollipop. Crypt Maps

[0135] Generating crypt maps may involve distinguishing between identically stained nuclei within crypts, extracellular matrix, and enterocytes. To do so, ViCE uses its 2D extension of the 1D MM filtering (MMF) method used by the original authors to improve lunar penetrating radar data (Zhang et al., 2019). Similar to bandpass filters, ViCE’s MMF extracts information specific to a range defined by the scale of two structuring elements of different size. This filter distinguishes between identically stained nuclei by disproportionately affecting nuclei surrounded by H-signal versus individual nuclei. This partitions crypt nuclei from extracellular and enterocyte nuclei. bwCrypt = f(imgHE,0) - f(imgHE, 3 / 2 × d–nuc), f(img,r) = 1 / 2 × [open(close(img,SE(r))) + close(open(img,SE(r)))] SE(r) = disk–shaped structuring element of radius r open(img,SE) = dilate( erode(img,SE),SE) close(img,SE) = erode( dilate(img,SE),SE)

[0136] Crypt segmentation maps are then found with disk-shaped structuring elementradii of [0, 1.5 × ^- ] and additional MM smoothing and filtering methods, detailed inSupplementary Materials. VC candidates are then cropped and rotated and individually processed. VC pairs may involve the selection of the “best” nearby crypt. This is interpreted as the longest crypt whose center of mass is within the villus’ search window. Quality Control Metrics

[0137] As mentioned above, quality control scores are based on seven physical evaluation metrics. Each metric outputs one of the following: Good, Okay, and Poor. From these outputs, a final Quality result is determined, where Good Quality lollipops correspond to well- oriented VC pairs. The individual evaluation metrics are described below. 1. Tissue Damage Metric ○ Lollipop border and apical enterocyte overlap percentage. ■ The lollipop border is the region between the upper bound (bwUB) and lower bound (bwLB) estimate of tissue boundary. Apical enterocytes contain moderate to strong signals. The border of a true villus ought to contain enterocytes. ○ Villus tip quality ■ Measure of intactness of villus tip 2. Orientation Metric ○ Threshold of max crypt and search window overlap percent ■ Crypts should overlap vertically relative to its search window. Well- oriented crypts should have overlap percentages > 90% ○ Min crypt aspect ratio■ Crypts should span the mucosal layer. Good crypts have aspect ratios > 3 ○ Min crypt area ■ If the size of a VC pair’s best local crypt is unreasonably small, the VC pair is not well-oriented. 3. Software Bug Catch ○ Min search window depth ■ Sometimes there’s roughly no overlap between the search window and the tissue boundary, usually due to the VC pair not being a VC pair. ○ Initial coarse quality control ■ ViCE initially overestimates the number of VC candidates because it’s easier to delete an object than to go back and find something else ■ Villus height is larger than the width of its lollipop Table S1. Workflow Generalization Table S1. Generalization of ViCE’s workflowSection S1. Outputs

[0138] When ViCE analyzes an image on its default settings, it will automatically generate and save figures summarizing evaluation and quantification results, as well as cropped,rotated, and indexed close-ups of every VC pair labeled fit for histological quantification. To better illustrate ViCE’s capabilities, the following figures include data generated from one input image. FIG. 11 is an exemplary uniform background correction result. FIG. 12 is an exemplary tissue quality evaluation summary. FIG. 13 depicts an exemplary contrast enhancement through color model manipulation of color deconvolution results and FIG. 14 depicts an exemplary crypt segmentation map. FIG. 15 depicts exemplary color deconvolution results (left) combined hematoxylin (red) and eosin (green), and gray scale color deconvolution results of (middle) eosin and (right) hematoxylin. FIG.16 depicts exemplary tissue boundary estimates: (left) upper bound, (middle) lower bound, and (right) smoothed lower bound estimate. FIG. 17 depicts an exemplary stain concentration density map of (left) eosin and (right) hematoxylin. FIG.18 depicts an example of an automatically generated close-up figure of a Good Quality lollipop, highlighting villus mask and “ideal” crypt choice. The figure title contains the lollipop index, quality score, villus height, crypt depth, and their ratio.

[0139] FIG. 19 provides examples of the automatically generated and saved intermediate results: (left) crypt segmentation map, (middle) crop and rotation result, and (right) lollipop mask representing the villus mask.

[0140] FIG. 20 provides an exemplary tissue quality evaluation summary. Section S2: Color Deconvolution Improvement

[0141] Color deconvolution (CD) is a method that separates and quantifies immunohistochemical staining information from color images (Ruifrok et al. 2001). We first transform RGB images into images with intensities proportional to stain concentration by adapting previously developed color deconvolution algorithms (Ruifrok et al. 2001, Landini et al. 2020). We significantly improve upon these algorithms by reducing the number of function callbacks bythe number of pixel elements through matrix manipulations. This can reduce execution time by two or more orders of magnitude without loss in precision, as shown in the graph depicted in FIG. 21. Section S3. ViCE’s Binarization Across a Diverse Dataset

[0142] This section refers to various figures. FIG. 22 shows raw images (e.g., RGB input images). FIG.23 shows a high relative hematoxylin signal image (corresponding to the raw images of FIG. 22). FIG. 24 shows a moderate relative hematoxylin signal image (corresponding to the raw images of FIG. 22). FIG. 25 shows a low relative hematoxylin signal image (corresponding to the raw images of FIG. 22). FIG. 26 shows a background relative hematoxylin signal image (corresponding to the raw images of FIG.22).

[0143] FIG. 27 shows a high relative eosin signal image (corresponding to the raw images of FIG. 22). FIG. 28 shows a moderate relative eosin signal image (corresponding to the raw images of FIG. 22). FIG. 29 shows a low relative eosin signal image (corresponding to the raw images of FIG. 22). FIG. 30 shows a background relative eosin signal image (corresponding to the raw images of FIG.22). Section S4: Determining ViCE’s Length Scale

[0144] In some examples, length related parameters may directly scale by the estimated average nucleus diameter. Nuclei are easily identifiable due to DNA’s strong affinity to hematoxylin (H), producing distinct purplish blue objects. FIG. 31 shows a raw image (e.g., an input RGB image). FIG. 32 shows an image where we have identified pixels of relatively high H concentration. FIG. 33 shows an image where we observe that the centers of nuclei typically contain pixels from H-strong.

[0145] In some examples, we then attempt to find nuclei diameter by mapping the skeleton of the binarized image H-strong with its distance transform. This method can be sensitive to salt and pepper noise. We can resolve this with dilation of a [3x3] box-shaped kernel followed by an erosion of a [3x3] cross-shaped kernel. FIG. 34 shows an exemplary distance transform.

[0146] By collecting the distance transform pixel values at each skeletonized pixel, we have effectively collected the pixel distance of local ridgelines to their nearest edge. FIG. 35 depicts a chart with regional max values from distance transform.

[0147] Since the objects within H-strong are expected to be objects composed of adjacent nuclei, it follows that the most common width should be the width of individual nuclei. Therefore, we estimate the most common nucleus radius as the argument that gives the maximum value from the histogram above. Finally, in some examples, we calculate the diameter as 2*radius + 1. Section S5 Upper Bound Tissue Boundary Estimate

[0148] We determine this segmentation map by morphologically refining a segmentation map derived from basic observations. For example, let us define background pixels as the following: {H(x,y), E(x,y)} = {H , E }

[0149] It follows that any pixel that is not a background pixel can be an element of an upper bound tissue boundary estimate, bwUB. This method, however, can produce low-utility segmentation maps composed of non-distinct blobs.

[0150] Observation 1: Commonly found damaged tissue typically contains frayed extracellular matrix and cytoplasm. Eosin stains these objects. Hematoxylin does not stain these objects.

[0151] Let us collect pixels where {H(x,y)} = {Hlow| Hmid| Hhigh} and eosin is sufficiently intense.

[0152] Objects are removed and holes are filled if they are within a set range parameterized by the expected nucleus diameter. Morphological closing is then applied, where the size of the structuring element also scales with the expected nucleus diameter. Lastly, a size filter removes all objects smaller than 1 / 10 of the largest object.

[0153] FIG. 36 depicts an initial map generated for an input image. FIG. 37 depicts three phases of a morphologic refinement. Section S6 ViCE’s Graphical User Interface

[0154] FIG.38 provides an exemplary depiction of a ViCE help screen, containing icon description. FIG.39 provides an exemplary depiction of an user interface of ViCE under usage.

[0155] Operating ViCE can be as simple as entering an image’s location and pressing RUN. File locations can either be entered within the large text field, capable of copying and pasting text, or through a dialog box, generated for point and click file selection by pressing OPEN FILE. Toggle batch processing ON by pressing the FAST FORWARD icon. Note that upon batch process completion, a dialogue box appears, requesting the user to select the save location of a CSV composed of the aggregate of batch-process-generated measurement CSV’s. Green icons represent settings toggled ON. Icons that do not change color represent an action. The icons of the left and right arrows tell ViCE to analyze the previous and next image respectively. Pressing CLOSE ALL FIGURES immediately closes all figures generated by ViCE. In the event of software error, pressing PROGRAM RESTART closes and reopens ViCE. Data is saved at the end of each image analysis. Section S7 Validation of ViCE

[0156] FIG. 40 shows the Area Under the Curve measured using the ROC curve. The area under the curve was 0.84. FIG. 41 shows the correlation between the ViCE and the manual measurement in the estimation of villus height (R2 = 0.88). FIG.42 shows the correlation between the ViCE and the manual measurement in the estimation of crypt depth (R2 = 0.49). FIG.43 shows the correlation between the ViCE and the manual measurement in the estimation of ratio of villus height to crypt depth (Vh / Cd) (R2 = 0.77) ILLUSTRATIVE COMPUTER SYSTEMS

[0157] Techniques operating according to the principles described herein may be implemented in any suitable manner. Included in the discussion above are a series of flow charts showing the steps and acts of various processes for automated tissue analysis. The processing and decision blocks of the flow charts above represent steps and acts that may be included in algorithms that carry out these various processes. Algorithms derived from these processes may be implemented as software integrated with and directing the operation of one or more single- or multi-purpose processors, may be implemented as functionally-equivalent circuits such as a Digital Signal Processing (DSP) circuit or an Application-Specific Integrated Circuit (ASIC), or may be implemented in any other suitable manner. It should be appreciated that the flow charts included herein do not depict the syntax or operation of any particular circuit or of any particular programming language or type of programming language. Rather, the flow charts illustrate the functional information one skilled in the art may use to fabricate circuits or to implement computer software algorithms to perform the processing of a particular apparatus carrying out the types of techniques described herein. It should also be appreciated that, unless otherwise indicated herein, the particular sequence of steps and / or acts described in each flow chart is merely illustrative ofthe algorithms that may be implemented and can be varied in implementations and embodiments of the principles described herein.

[0158] Accordingly, in some embodiments, the techniques described herein may be embodied in computer-executable instructions implemented as software, including as application software, system software, firmware, middleware, embedded code, or any other suitable type of computer code. Such computer-executable instructions may be written using any of a number of suitable programming languages and / or programming or scripting tools, and also may be compiled as executable machine language code or intermediate code that is executed on a framework or virtual machine.

[0159] When techniques described herein are embodied as computer-executable instructions, these computer-executable instructions may be implemented in any suitable manner, including as a number of functional facilities, each providing one or more operations to complete execution of algorithms operating according to these techniques. A “functional facility,” however instantiated, is a structural component of a computer system that, when integrated with and executed by one or more computers, causes the one or more computers to perform a specific operational role. A functional facility may be a portion of or an entire software element. For example, a functional facility may be implemented as a function of a process, or as a discrete process, or as any other suitable unit of processing. If techniques described herein are implemented as multiple functional facilities, each functional facility may be implemented in its own way; all need not be implemented the same way. Additionally, these functional facilities may be executed in parallel and / or serially, as appropriate, and may pass information between one another using a shared memory on the computer(s) on which they are executing, using a message passing protocol, or in any other suitable way.

[0160] Generally, functional facilities include routines, programs, objects, components, data structures, etc. that perform particular tasks or implement particular abstract data types. Typically, the functionality of the functional facilities may be combined or distributed as desired in the systems in which they operate. In some implementations, one or more functional facilities carrying out techniques herein may together form a complete software package. These functional facilities may, in alternative embodiments, be adapted to interact with other, unrelated functional facilities and / or processes, to implement a software program application.

[0161] Some exemplary functional facilities have been described herein for carrying out one or more tasks. It should be appreciated, though, that the functional facilities and division of tasks described is merely illustrative of the type of functional facilities that may implement the exemplary techniques described herein, and that embodiments are not limited to being implemented in any specific number, division, or type of functional facilities. In some implementations, all functionalities may be implemented in a single functional facility. It should also be appreciated that, in some implementations, some of the functional facilities described herein may be implemented together with or separately from others (i.e., as a single unit or separate units), or some of these functional facilities may not be implemented.

[0162] Computer-executable instructions implementing the techniques described herein (when implemented as one or more functional facilities or in any other manner) may, in some embodiments, be encoded on one or more computer-readable media to provide functionality to the media. Computer-readable media include magnetic media such as a hard disk drive, optical media such as a Compact Disk (CD) or a Digital Versatile Disk (DVD), a persistent or non- persistent solid-state memory (e.g., Flash memory, Magnetic RAM, etc.), or any other suitable storage media. Such a computer-readable medium may be implemented in any suitable manner,including as computer-readable storage media 906 of FIG. 9 described below (i.e., as a portion of a computing device 900) or as a stand-alone, separate storage medium. As used herein, “computer- readable media” (also called “computer-readable storage media”) refers to tangible storage media. Tangible storage media are non-transitory and have at least one physical, structural component. In a “computer-readable medium,” as used herein, at least one physical, structural component has at least one physical property that may be altered in some way during a process of creating the medium with embedded information, a process of recording information thereon, or any other process of encoding the medium with information. For example, a magnetization state of a portion of a physical structure of a computer-readable medium may be altered during a recording process.

[0163] In some, but not all, implementations in which the techniques may be embodied as computer-executable instructions, these instructions may be executed on one or more suitable computing device(s) operating in any suitable computer system, including the exemplary computer system of FIG. 1, or one or more computing devices (or one or more processors of one or more computing devices) may be programmed to execute the computer-executable instructions. A computing device or processor may be programmed to execute instructions when the instructions are stored in a manner accessible to the computing device or processor, such as in a data store (e.g., an on-chip cache or instruction register, a computer-readable storage medium accessible via a bus, a computer-readable storage medium accessible via one or more networks and accessible by the device / processor, etc.). Functional facilities comprising these computer-executable instructions may be integrated with and direct the operation of a single multi-purpose programmable digital computing device, a coordinated system of two or more multi-purpose computing device sharing processing power and jointly carrying out the techniques described herein, a single computing device or coordinated system of computing devices (co-located or geographically distributed)dedicated to executing the techniques described herein, one or more Field-Programmable Gate Arrays (FPGAs) for carrying out the techniques described herein, or any other suitable system.

[0164] FIG. 9 illustrates one exemplary implementation of a computing device in the form of a computing device 900 that may be used in a system implementing techniques described herein, although others are possible. It should be appreciated that FIG. 9 is intended neither to be a depiction of necessary components for a computing device to execute one or more of the facilities described in connection with FIG. 1 in accordance with the principles described herein, nor a comprehensive depiction.

[0165] Computing device 900 may comprise at least one processor 902, a network adapter 604, and a computer-readable storage medium 906. Computing device 900 may be, for example, a desktop or laptop personal computer, a personal digital assistant (PDA), a smart mobile phone, a server, a wireless access point or other networking element, or any other suitable computing device. Network adapter 904 may be any suitable hardware and / or software to enable the computing device 900 to communicate wired and / or wirelessly with any other suitable computing device over any suitable computing network. The computing network may include wireless access points, switches, routers, gateways, and / or other networking equipment as well as any suitable wired and / or wireless communication medium or media for exchanging data between two or more computers, including the Internet. Computer-readable media 906 may be adapted to store data to be processed and / or instructions to be executed by processor 902. Processor 902 enables processing of data and execution of instructions. The data and instructions may be stored on the computer-readable storage medium 906.

[0166] The data and instructions stored on the computer-readable storage media 906 may comprise computer-executable instructions implementing techniques which operateaccording to the principles described herein. In the example of FIG. 9, the computer-readable storage medium 906 stores computer-executable instructions implementing various facilities and storing various information as described above. The computer-readable storage medium 906 may store one or more of the facilities depicted in FIG. 1, which may implement one or more of the techniques described herein.

[0167] While not illustrated in FIG. 9, a computing device may additionally have one or more components and peripherals, including input and output devices. These devices can be used, among other things, to present a user interface. Examples of output devices that can be used to provide a user interface include printers or display screens for visual presentation of output and speakers or other sound generating devices for audible presentation of output. Examples of input devices that can be used for a user interface include keyboards, and pointing devices, such as mice, touch pads, and digitizing tablets. As another example, a computing device may receive input information through speech recognition or in other audible format. OTHER EMBODIMENTS

[0168] From the foregoing description, it will be apparent that variations and modifications may be made to the embodiments described herein to adapt it to various usages and conditions. Such embodiments are also within the scope of the following claims.

[0169] It will be understood by one skilled in the art that this disclosure is not limited in its application to the details of construction and the arrangement of components set forth in the above description or illustrated in the drawings. The embodiments herein are capable of other embodiments, and capable of being practiced or carried out in various ways. Also, it will be understood that the phraseology and terminology used herein is for the purpose of description and should not be regarded as limiting. The use of “including,” “comprising,” or “having” andvariations thereof herein is meant to encompass the items listed thereafter and equivalents thereof as well as additional items. Unless limited otherwise, the terms “connected,” “coupled,” and “mounted,” and variations thereof herein are used broadly and encompass direct and indirect connections, couplings, and mountings. In addition, the terms “connected” and “coupled” and variations thereof are not restricted to physical or mechanical connections or couplings. Further, terms such as up, down, bottom, and top are relative, and are employed to aid illustration, but are not limiting. Use of ordinal terms such as “first,” “second,” “third,” etc., in the claims to modify a claim element does not by itself connote any priority, precedence, or order of one claim element over another or the temporal order in which acts of a method are performed, but are used merely as labels to distinguish one claim element having a certain name from another element having a same name (but for use of the ordinal term) to distinguish the claim elements. The word “exemplary” is used herein to mean serving as an example, instance, or illustration. Any embodiment, implementation, process, feature, etc. described herein as exemplary should therefore be understood to be an illustrative example and should not be understood to be a preferred or advantageous example unless otherwise indicated.

[0170] The above-presented description and figures are intended by way of example only and are not intended to limit the illustrative embodiments in any way except as set forth in the following claims. It is particularly noted that persons skilled in the art can readily combine the various technical aspects of the various elements of the various illustrative embodiments that have been described above in numerous other ways, all of which are considered to be within the scope of the claims.

[0171] The recitation of a listing of elements in any definition of a variable herein includes definitions of that variable as any single element or combination (or subcombination) oflisted elements. The recitation of an embodiment herein includes that embodiment as any single embodiment or in combination with any other embodiments or portions thereof.

[0172] While some embodiments of the present disclosure have been shown and described herein, it will be obvious to those skilled in the art that such embodiments are provided by way of example only. Numerous variations, changes, and substitutions will now occur to those skilled in the art without departing from the present disclosure. It should be understood that various alternatives to the embodiments described herein, or combinations of one or more of these embodiments or aspects described therein may be employed in practicing the present disclosure. It is intended that the following claims define the scope of the present disclosure and that methods and structures within the scope of these claims and their equivalents be covered thereby.

Claims

AMENDED CLAIMS received by the International Bureau on 18 June 2025 (18.06.2025)What is claimed is:

1. [amended] A computer-implemented method of stain normalization image processing for digitized biological tissue images, the method comprising: identifying, from an input image of a biological tissue section, a plurality of stain channels, each indicating a different stain of a plurality of stains applied to the biological tissue section and comprising a range of pixel intensity values corresponding to a range of normalized concentrations for that stain; generating, for each stain channel within the plurality of the stain channels, a plurality of images based on pixel intensity relative stain concentration, wherein each image within the plurality of images covers the same area of the input image and captures pixel intensity values within a different segment of the stain channel’s range of pixel intensity values; and outputting one or more stain-normalized images based on one or more of the plurality of images generated for each stain channel.

2. The computer- implemented method of claim 1 , wherein the range of pixel intensity values for each stain channel comprises a pixel intensity value for each pixel of the input image that represents a stain intensity, in the pixel, corresponding to concentration of the stain indicated by the stain channel.

3. The computer-implemented method of claim 2, wherein the range of pixel intensity values within each stain channel comprises a range of grayscale intensity values.

4. The computer-implemented method of claim 1, wherein identifying the plurality of stain channels from the input image of the biological tissue section comprises deconvolving color values for the input image of the biological tissue section into the plurality of stain channels.

5. The computer- implemented method of claim 1, wherein at least one image, within the plurality of images generated for each stain channel, captures one or more dynamics of pixel intensity levels across the pixel intensity values of the at least one image.

6. The computer-implemented method of claim 5, wherein the one or more dynamics of pixel intensity levels, captured within the image, comprise at least one of: spatial dynamics indicating variations in pixel intensity across two or more regions of the image; contrast dynamics indicating a variation between one or more of the relatively highest pixel intensity values within the image and one or more of the relatively lowest pixel intensity values within the image; brightness and saturation dynamics corresponding to relative quantities of pixel intensity values designated as a high pixel intensity value and pixel intensity values designated as low pixel intensity values; and distribution dynamics indicating a distribution of the pixel intensity values captured by the image.

7. The computer-implemented method of claim 1 , wherein: the plurality of images generated for a particular stain channel within the plurality of stain channels comprises a stain-background image, a stain-high image, and a stain-low image; the stain-background image captures a segment of pixel intensity values corresponding to a background level of a stain indicated by the particular stain channel; the stain-high image captures a segment of pixel intensity values that are higher than the pixel intensity values in each of the other segments of the particular stain channel’s range of pixel intensity values; and the stain-low image captures a segment of pixel intensity values that are above the pixel intensity values of the segment captured in the stain-background image and below the pixel intensity values of the segment captured in the stain-high image.

8. The computer-implemented method of claim 7, wherein: the plurality of images generated for the particular stain channel further comprises a stain-moderate image; and the stain-moderate image captures a segment of pixel intensity values that are above the pixel intensity values of the segment captured in the stain-low image and below the pixel intensity values of the segment captured in the stain-high image.

9. The computer-implemented method of claim 7, wherein the particular stain channel comprises a hematoxylin channel corresponding to a hematoxylin stain.

10. The computer-implemented method of claim 7, wherein the particular stain channel comprises an eosin channel corresponding to an eosin stain.

11. The computer- implemented method of claim 1, wherein the input image is captured via a camera of a microscope.

12. A computer-implemented method for constructing scale invariant images of biological tissue sections, the method comprising: identifying a plurality of nuclei of cells within an image of a section of biological tissue; setting a length scale based on a size of one or more of the plurality of nuclei; and scaling, for a morphological structure other than a nuclei that is represented in the image, a parameter related to length using the length scale that is based on the size of the one or more nuclei.

13. The computer- implemented method of claim 12, wherein setting the length scale based on the size of one or more of the plurality of nuclei comprises setting the length scale based on an average size of the plurality of nuclei.

14. The computer- implemented method of claim 12, wherein the image of the section of biological tissue is a stain-normalized image in which pixel intensity values corresponding to stain concentrations of a stain have been normalized, wherein the stain is a stain, applied to the section of biological tissue, that stains for cell nuclei.

15. The computer- implemented method of claim 14, wherein the stain comprises at least one of: a hematoxylin stain; a fluorescent stain that binds to DNA; a Feulgen stain; a methyl green stain; or a Giemsa stain.

16. The computer- implemented method of claim 14, wherein the image is a stain- high normalized image comprising pixel intensity values that are higher than pixel intensity values, corresponding to stain concentrations of the stain, of one or more additional stain-normalized images.

17. The computer-implemented method of claim 12, wherein: the section of biological tissue comprises a section of intestinal tissue; the morphological structure comprises a villus-crypt pair; and the parameter related to length comprises a villus-height to crypt-depth ratio for the villus-crypt pair.

18. The computer- implemented method of claim 12, wherein the image is captured via a camera of a microscope at an unknown level of magnification.

19. At least one computer-readable storage medium having encoded thereon executable instructions that, when executed by at least one processor, cause the at least one processor to carry out a method, the method comprising or one a combination of the methods of claims 1-18.

20. An apparatus comprising: at least one processor; and at least one computer-readable storage medium having encoded thereon executable instructions that, when executed by the at least one processor, cause the at least one processor to carry out a method, the method comprising or one a combination of the methods of claims 1-18.

21. A computer-implemented method comprising: generating, from an input image of a biological tissue section, a stain channel indicating a stain of interest; determining, from the stain channel, a lower bound; creating a segmentation map based on the lower bound; and identifying a morphological structure of the biological tissue section from the segmentation map.

22. The computer-implemented method of claim 21, wherein: the stain channel comprises a range of pixel intensity values corresponding to a range of stain concentrations;the method further comprises generating a plurality of images for the stain channel, each of which captures a different subrange in the stain channel’s range of pixel intensity values; the plurality of images comprises a stain-high image that captures a subrange of pixel intensity values that are higher than the pixel intensity values of any of the other subranges in the range of pixel intensity values; and the lower bound comprises pixel intensity values from the stain-high image.

23. The computer-implemented method of claim 22, wherein: creating the segmentation map from the lower bound comprises applying a correction term to the lower bound and creating the segmentation map from the corrected lower bound; and the method further comprises: generating, from the input image, a plurality of stain channels, comprising the stain channel indicating the stain that stains for nuclei, wherein each stain channel within the plurality of stain channels corresponds to a different stain applied to the biological tissue section; generating, from the plurality of stain channels, an upper bound; and generating the correction term based on the upper bound.

24. The computer-implemented method of claim 23, wherein: each stain channel within the plurality of stain channels comprises a different range of pixel intensity values corresponding to a different range of stain concentrations; the method further comprises generating a plurality of images for each stain channel within the plurality of stain channels;each of the plurality of images comprises a stain-background image, a stain-high image, and a stain-low image, each of which captures pixel intensity values within a different subrange of the corresponding stain channel’s range of pixel intensity values; the stain-background image captures a subrange of pixel intensity values corresponding to a background level of a stain indicated by the corresponding stain channel; the stain-high image captures a subrange of pixel intensity values that are higher than the pixel intensity values in the subranges of the stain-background image and the stain-low image; and the stain-low image captures a subrange of pixel intensity values that are higher than the pixel intensity values in the subrange of the background image and lower than the pixel intensity values in the subrange of the stain-high image.

25. The computer-implemented method of claim 24, wherein: each of the plurality of images further comprises one or more stain-medium images comprising a subrange of pixel intensity values that are between the pixel intensity values in the subranges of the stain-high images and the stain-low images.

26. The computer-implemented method of claim 24, wherein: the upper bound comprises pixels based on one or more of the images generated for each of the plurality of stain channels except for the stain-background images.

27. The computer-implemented method of claim 21, wherein: the stain channel indicates a stain that stains for nuclei; andthe stain comprises at least one of: a hematoxylin stain; or a fluorescent stain that binds to DNA.

28. A computer-implemented method comprising: identifying, from an input image of a biological tissue section, a plurality of stain channels, each indicating a different stain applied to the biological tissue section; normalizing, based on the plurality of stain channels, each pixel intensity value for the input image of the biological tissue section to one or more normalized values to yield one or more stain-normalized images of the biological tissue section; and identifying a morphological structure of the biological tissue section from the one or more stain-normalized images of the biological tissue section.

29. The computer-implemented method of claim 28, wherein: the biological tissue section is a section of intestinal tissue; and the morphologic structure is a villus-crypt pair.

30. The computer-implemented method of claim 29, wherein the plurality of stain channels comprises a hematoxylin channel, corresponding to a hematoxylin stain, and an eosin channel, corresponding to an eosin stain.

31. The computer- implemented method of claim 28, further comprising: identifying, from at least one of the input image or a stain-normalized image generated from the input image, a plurality of nuclei; and setting a length scale based on a size of one or more of the plurality of nuclei.

32. The computer-implemented method of claim 31 , wherein: the method further comprises determining a quantitative histological measure of the morphologic structure using the length scale; and the quantitative histological measure comprises a villus-height to crypt-depth ratio of a villus-crypt pair.

33. The computer-implemented method of claim 28, wherein: identifying the morphologic structure comprises determining that the morphological structure satisfies a quality metric; and the method further comprises, in response to determining that the morphological structure satisfies the quality metric, generating a user interface, comprising a display of the morphological structure, that at least one of: visually draws attention to the morphological structure; visually indicates that the morphological satisfies the quality metric; or is prioritized in a queue of user interfaces, wherein each user interface in the queue comprises an image of a different section of the biological tissue.

34. The computer- implemented method of claim 33, further comprising: determining that an image of an additional section of the biological tissue either doesn’t include any instance of the morphological structure or includes an additional instance of the morphological structure that doesn’t satisfy the quality metric; in response to determining that the morphological structure in the section of the biological tissue satisfies the quality metric, adding the input image to a digital queue of usable images for a pathologist to review; and in response to determining that the image of the additional section of the biological tissue either doesn’t include any instance of the morphological structure or includes an additional instance of the morphologic structure that doesn’t satisfy the quality metric, precluding the image of the additional section from the digital queue of usable images.

35. The computer-implemented method of claim 28, wherein identifying the morphological structure from the one or more stain-normalized images comprises generating a segmentation map of the one or more stain-normalized images and identifying the morphologic structure from the segmentation map.

36. The computer-implemented method of claim 35, wherein: the method further comprises identifying a lower bound of the stain-normalized image; and generating the segmentation map comprises at least one of: segmenting the lower bound to generate the segmentation map; orapplying a correction term to the lower bound and segmenting the corrected lower bound to generate the segmentation map.

37. A computer-implemented method comprising: determining, from an input image of an intestinal tissue section, a plurality of stain channels comprising a hematoxylin channel, corresponding to a hematoxylin stain applied to the intestinal tissue section, and an eosin channel, corresponding to an eosin stain applied to the intestinal tissue section; normalizing, based on the plurality of stain channels, each pixel intensity value for the input image of the intestinal tissue section to one or more normalized values, to yield one or more stain-normalized images of the intestinal tissue section; and identifying one or more villus-crypt pairs of the intestinal tissue section from the one or more stain-normalized images of the intestinal tissue section.

38. The computer-implemented method of claim 37, wherein: identifying the one or more villus-crypt pairs comprises identifying a villus-crypt pair by identifying a central line, corresponding to a portion of the intestinal tissue, from which both a first line, corresponding to a villus, is projected in a first direction and a second line, corresponding to a crypt, is projected in a second direction; and the central line corresponds to a center of intensity identified for pixels of at least one of a lower bound or a corrected lower bound generated from the one or more stain-normalized images.

39. At least one computer-readable storage medium having encoded thereon executable instructions that, when executed by at least one processor, cause the at least one processor to carry out a method, the method comprising or one a combination of the methods of claims 1-38.

40. An apparatus comprising: at least one processor; and at least one computer-readable storage medium having encoded thereon executable instructions that, when executed by the at least one processor, cause the at least one processor to carry out a method, the method comprising or one a combination of the methods of claims 1-38.

Images