Sensitivity analysis for digital pathology

Abstract

The ability to obtain and digitize pathology samples (such as in histopathology or cytopathology) has opened up the possibility of accurate and automated computer-operated analysis or computer-assisted analysis when diagnosing a wide variety of medical conditions. Typical digital pathology image analysis pipelines involve a large number of image preprocessing and processing operations. Small changes in the parameters used in these algorithms can lead to significant changes in the final pathological findings. Therefore, the present application proposes to perform sensitivity analysis on a digital pathology image analysis pipeline to evaluate the sensitivity of a given pathology result to changes in parameters in the digital pathology image analysis pipeline.

Description

Technical Field

[0001] The present invention relates to an apparatus for processing digital pathology image data to generate pathology scoring data and sensitivity analysis of the pathology scoring data, as well as an associated method, a digital pathology system, a computer program element and a computer readable medium. Background Art

[0002] In the field of pathology, there is a continuing trend toward automation of aspects of the analysis of digital pathology images. Digital image processing techniques are well suited to analyzing large numbers of input images in a standardized manner. However, such methods are highly dependent on variations in the initial preparation of the samples used to obtain the pathology images (such as the exact techniques and conditions present when, for example, preparing, storing, and examining histopathology slides).

[0003] Here, reference is made to sensitivity analysis, which can be found, for example, in the article on sensitivity analysis https: / / en.wikipedia.org / wiki / Sensitivity_analysis Further background details can be found in . Sensitivity analyses are Mathematical model or the output of a system (numerical or otherwise) Uncertainty The study of how different sources of uncertainty in their inputs can be partitioned and allocated.

[0004] In Otsu methods in computer vision, see e.g., https: / / en.wikipedia.org / wiki / Otsu% 27s_method (by Nobuyuki Otsu (named after Ōtsu Nobuyuki) is used to automatically perform clustering-based image Thresholding , or, reduce the grayscale image to a binary image. The algorithm assumes that the image contains two classes of pixels following a bimodal histogram (foreground pixels and background pixels), it then calculates the optimal threshold to separate the two classes such that their combined diffusion (intra-class variance ) is minimized, or equivalently (since the sum of pairwise squared distances is a constant), such that its between-class variance is maximized. Thus, Otsu’s method is roughly Fisher's Discriminant Analysis The Otsu method also directly involves Jenks Optimization Method .

[0005] US2013 / 0030305A1 describes a method for determining the stage of fibrosis in the liver. The method comprises the following steps: obtaining input data related to the liver, the input data being generated using an imaging system based on second harmonic generation; identifying a plurality of morphological features of the liver from the input data related to the liver; generating a plurality of measurement results based on the identified plurality of morphological features; and determining the stage of fibrosis in the liver based on the generated plurality of measurement results.

[0006] WO2016 / 138041A2 describes methods, systems and devices for evaluating the status of cells in a sample, involving imaging of cells, transforming cell images into biophysical metrics, and transforming biophysical metrics into prognostic indicators at the cell and subject levels. Automated devices, processes and analyses are also described.

[0007] The publication “A Method for Normalizing Histology Slides for Quantitative Analysis” by Macenko et al. (published in ISBI 2009 Proceedings, pp. 1107-1110, IEEE 2009, 978-1-4244-3932-4 / 09) discusses digital processing techniques aimed at overcoming known inconsistencies in histopathology staining procedures.

[0008] However, such techniques can be further improved. Summary of the invention

[0009] Therefore, there is a need for improved digital pathology imaging methods.

[0010] The objects of the invention are solved by the subject-matter of the attached independent claims, wherein further embodiments are included in the dependent claims.

[0011] According to a first aspect, there is provided an apparatus for digitally processing digital pathology image data to generate pathology scoring data and sensitivity analysis of the pathology scoring data. The apparatus comprises:

[0012] - input unit; and

[0013] -Processing unit.

[0014] The input unit is configured to obtain digital pathology image data, including an image of a pathology sample.

[0015] The processing unit is configured to classify objects in the digital pathology image data into multiple candidate objects, assign a first pathological state to at least one candidate object among the multiple candidate objects according to one or more detection thresholds and / or detection probabilities, obtain initial pathological score data of the pathological sample based on the candidate objects among the multiple candidate objects to which the first pathological state has been assigned, perturb the one or more detection thresholds and / or the detection probabilities according to a perturbation function to generate perturbation detection thresholds and / or perturbation detection probabilities, reallocate the first pathological state to at least one candidate object among the multiple candidate objects according to the one or more perturbation detection thresholds and / or perturbation detection probabilities, obtain updated pathological score data of the pathological sample based on the candidate objects to which the first pathological state has been reallocated according to the one or more perturbation detection thresholds and / or the perturbation detection probabilities, and compare the initial pathology score data with the updated pathology score data to obtain the sensitivity of the pathology score data.

[0016] Thus, it is possible to generate and provide feedback about the sensitivity of the pathology score data to actual changes in one or more thresholds used in the analysis that has been used to generate the score. The thresholds may, for example, be related to the probability of pixel or region classification of an object in a digital pathology image. Alternatively or in combination, the thresholds may be related to the intensity and / or wavelength representing the expression of a biomarker in a pathology sample.

[0017] In other words, as well as automatically generating a pathology score that enables diagnosis of a condition to be performed based on a digital pathology image, a medical expert may be provided with an assessment of how sensitive the pathology score is to minor changes in the analysis. This may be particularly important when the output pathology score is at or near a boundary value, where a pathology score slightly above or slightly below the boundary value may result in an alternative treatment option. The medical expert may use the sensitivity of the pathology technique in decision making and reporting, and may optionally perform a manual inspection of the indicated problematic areas. Thus, making the pathology process faster and more reliable.

[0018] Optionally, the one or more detection thresholds and / or detection probabilities characterize changes in automatic morphological feature detection when determining the presence of a candidate object in the digital pathology image data.

[0019] Thus, the sensitivity of the pathology scoring method when detecting features such as cells, nuclei, or membranes can be tested. A morphological feature detection algorithm such as "Morphology Close" can have several input parameters that change its effectiveness. By generating several versions of the morphological feature detection using initial input parameters and randomly perturbed input parameters, the sensitivity of the pathology scoring algorithm to the morphological feature detection algorithm when considered with a specific digital pathology image can be considered.

[0020] Optionally, the one or more detection thresholds represent probabilities of automatically classifying an object in the digital pathology image data into the plurality of candidate objects.

[0021] Optionally, the detection threshold and / or the detection probability characterizes an intensity level and / or a wavelength range of light emitted from a pathology sample and represented in the digital pathology image data.

[0022] Typically, a pathology sample is processed with a biomarker that colors the sample a first color when the biomarker is in contact with a material target and a second color (or no coloration) when the biomarker is not in contact with a material target. Such coloration can be observed, for example, using a light microscope and / or a fluorescence microscope (in the case of a fluorescently labeled biomarker). Changes in the biomarker application protocol, sample preparation, storage, and / or viewing all result in sensitivity in the provision of a pathology score. Therefore, in this optional embodiment, it is possible to evaluate the impact of changes in the intensity level and / or wavelength range of light emitted from a pathology sample on the final pathology score associated with the sample.

[0023] Optionally, the intensity level and / or wavelength range emitted from the pathology sample and represented in the digital pathology image data is indicative of a relative level of expression of a biomarker.

[0024] Optionally, the processing unit is further configured to obtain a plurality of disturbance detection thresholds and / or disturbance detection probabilities forming a sensitivity function of a pathology score of the digital pathology image data, and calculate a relative detection threshold change required to provide a change in the pathology score.

[0025] Thus, generating a sensitivity function based on multiple perturbation detection thresholds enables the rate of change of the sensitivity function to be evaluated as the rate of change of one or more of the detection thresholds and / or probabilities rather than at discrete thresholding points. From this, optimal thresholds for specific sample types and other analysis settings can be provided that are less likely to cause unpredictable changes in, for example, the pathology score.

[0026] Optionally, the processing unit is further configured to obtain the initial pathology score data and / or the updated pathology score data by calculating the number of candidate objects in the field of interest of the digital pathology data to which the first pathological state has been assigned before and / or after the disturbance, or by calculating a percentage score of the number of candidate objects in the field of interest of the digital pathology data to which the first pathological state has been assigned before and / or after the disturbance.

[0027] Optionally, the processing unit is further configured to: assign a second pathological state to at least one candidate object showing positive expression of a first biomarker of interest, assign a third pathological state to at least one candidate object showing negative expression of the first biomarker of interest, obtain the initial pathology score data by performing calculations using the candidate objects in the field of interest of the digital pathology data in the second and third pathological states, perturb a detection threshold used to assign the second and third pathological states to the candidate objects and reassign the second and third pathological states to the candidate objects after the perturbation, and obtain updated pathology score data by performing calculations using the candidate objects in the second and third pathological states after the perturbation.

[0028] Histopathology protocols involving the measurement of negative and positive marker expression are complex to perform, and therefore the effects of variable measurement thresholds in sample preparation and digital image processing of the digital pathology images may result in a more accurate assessment of the results of such complex protocols.

[0029] Optionally, the device further comprises:

[0030] - Output unit.

[0031] The output unit is configured to optionally combine the initial and / or updated pathology score data to output the sensitivity of the pathology score data to a user and / or output the sensitivity of the pathology score relative to a clinical threshold.

[0032] Thus, the medical professional is quickly and conveniently provided with an assessment of the sensitivity of the pathology parameter to internal test variations.Hitherto, the impact of threshold variations on pathology tests has been difficult to detect.

[0033] Optionally, the processing unit is further configured to spatially subsample the digital pathology image data into a plurality of sectors, and obtain initial pathology score data and an updated pathology score for each sector, and generate a spatial sensitivity mask of the digital pathology image data using the initial pathology score data and the updated pathology score for each sector. The output unit is configured to display the spatial sensitivity mask to a user, optionally as a semi-transparent overlay in spatial alignment superimposed on the digital pathology image data.

[0034] Thus, an intuitive and user-friendly assessment of the impact of the sensitivity of the pathology analysis pipeline to threshold changes is provided to the medical expert when generation of pathology scoring data from digital pathology images may be provided.

[0035] Optionally, the pathological sample is a histopathological sample or a cytopathological sample.

[0036] According to a second aspect, a digital pathology system is provided. The digital pathology system comprises:

[0037] -Digital pathology image acquisition equipment; and

[0038] - an apparatus according to the first aspect or one of its embodiments,

[0039] Therein, the digital pathology scanner is configured to receive a pathology sample, automatically analyze the pathology sample and thereby obtain digital pathology image data, and transmit the digital pathology image data to the device.

[0040] According to a third aspect, a method for digitally processing digital pathology image data to generate pathology scoring data and sensitivity analysis of the pathology scoring data is provided. The method comprises:

[0041] a) obtaining digital pathology image data, including images of pathology samples;

[0042] b) classifying objects in the digital pathology image data into a plurality of candidate objects;

[0043] c) assigning a first pathological state to at least one candidate object among the plurality of candidate objects according to one or more detection thresholds and / or detection probabilities;

[0044] d) obtaining initial pathology score data of the pathology sample based on a candidate object among a plurality of candidate objects to which the first pathology state has been assigned;

[0045] e) perturbing the one or more detection thresholds and / or the detection probability according to a perturbation function to generate a perturbed detection threshold and / or a perturbated detection probability;

[0046] f) reallocating the first pathological state to at least one candidate object among the plurality of candidate objects, or reallocating the first pathological state to at least one candidate object among the plurality of candidate objects according to the one or more disturbance detection thresholds and / or disturbance detection probabilities; and

[0047] g) obtaining updated pathology score data for the pathology sample based on the candidate objects to which the first pathological state has been reallocated according to the one or more disturbance detection thresholds and / or the disturbance detection probability, and comparing the initial pathology score data with the updated pathology score data to obtain the sensitivity of the pathology score data.

[0048] Optionally, the method is a method for processing digital histopathology image data to generate histopathology scoring data and sensitivity analysis of the histopathology scoring data.

[0049] Optionally, the method is a method for processing digital cytopathology image data to generate cytopathology scoring data and sensitivity analysis of the cytopathology scoring data.

[0050] According to a fourth aspect, there is provided a computer program element comprising instructions which, when the program is executed by a computer or a processing unit, cause the processor to perform the method of the third aspect.

[0051] According to a fifth aspect, there is provided a computer readable medium having stored thereon the computer program element of the fourth aspect.

[0052] In the following applications, the term "digital pathology image data" means a digital representation of a pathology slide obtained using, for example, a digital light microscope or a digital fluorescence microscope. Many scanners can support resolutions between 10x and 40x magnification. An image file associated with a 20x scan of a 15mm x 20mm sample is as large as 3.6GB. The image can be compressed to a more manageable size using, for example, JPEG-2000 compression. The image is typically accompanied by metadata containing scan parameters, scan plans, etc. In a digital histopathology image, identifiers defining the portion, block, and strain type of the portion may be included in the metadata.

[0053] In the following applications, the term "pathology" refers to the diagnosis of medical conditions by examination of samples of tissue, cells and body fluids. In particular, "histopathology" refers to the preparation of biopsy tissue samples, such as in resin blocks that are then adhered to serial sections prepared on slides. "Cytopathology" refers to the examination of free cells or tissue microfragments.

[0054] In the following applications, the term "pathology score data" means the examination of pathology (histopathology, cytopathology) to provide the likelihood of the presence of certain medical conditions. The pathology score can be derived by examining a histopathology slide and counting the number of objects of interest, such as cells or cell nuclei that meet the criteria for the presence of a certain disease, as opposed to the number of cells that do not meet the criteria. Thus, the pathology score can be provided as, for example, a cell count per unit area or percentage of a digital pathology image.

[0055] In the following application, the term "clinical threshold" means that the pathology score indicates that a medical condition is present based on the count of objects of interest identified in the digital pathology image data. In general, the implementation of a clinical threshold can be used to decide on different forms of medical treatment.

[0056] In the following applications, the term "sensitivity analysis" means an assessment of how a pathology score changes, and / or the rate at which a pathology score changes, when an internal threshold in an algorithm used to determine or generate the pathology score is changed (or perturbed). Such internal thresholds may not be visible to an end user of a digital pathology system. Additionally, in complex protocols, there may be so many internal thresholds that the complexity of changes to those thresholds and their impact on the digital pathology image data processing pipeline quickly overwhelms even an experienced user. Sensitivity analysis thus provides a realistic assessment of how changes (perturbations) to the parameters used in the analysis pipeline may affect the pathology score.

[0057] Therefore, in the following application, the term "sensitivity function" means the relationship between the level of an internal threshold in the analysis pipeline and its impact on the pathology score as applied to the same input digital pathology image data.

[0058] In the following application, the term "perturbation" means a change made to one or more of the internal thresholds of the digital pathology image analysis pipeline. The skilled reader will understand that the number and type of internal thresholds of the digital pathology image analysis pipeline are highly dependent on the protocol executed by the analysis pipeline. In a simple case, the image analysis pipeline may receive as input a digital histopathology image stained with hematoxylin conventionally thresholded at an absorption level of 0.1. A simple example of a perturbation would be to change the threshold used to threshold the digital histopathology image at a level of 0.095 or 0.15. This simple perturbation will cause more or fewer features in the input digital histopathology image to be present, and therefore any subsequent classification or morphology algorithm will detect a greater or lesser number of cells, for example, which results in a change in the final pathology score.

[0059] Of course, a complex analysis pipeline may have 10 or 100 individual thresholds or decision probabilities that can vary, which results in a highly sized possible disturbance space. A method of providing the disturbance is to change a decision threshold or decision probability on many given uncertainty ranges and record the sensitivity analysis method of the impact on the pathology score. The process can be repeated for other thresholds or decision capabilities. Although the technology is simple, it does not analyze the entire input space because it does not evaluate the simultaneous variation of thresholds. Of course, other more complex sensitivity analysis techniques may also be applied, such as regression analysis or partial derivative analysis across the entire disturbance space.

[0060] In the following applications, the term "automatic morphological feature detection" means an algorithm that can identify candidate objects in digital pathology images, such as cells, cell nuclei, cell membranes (whether intact or partial), etc. Techniques such as morphological opening or morphological closing can be used to perform automatic morphological feature detection. Of course, these algorithms can have many input parameters that define their performance, and all or some of these input parameters can be perturbed as part of the sensitivity analysis.

[0061] In the following application, the term "biomarker" in the context of the present application means an object or molecule detected in a pathology (histopathology, cytopathology) that enables the specific detection (isolation) of a specific expressed protein or other biological signal. The object or molecule can be detected, for example, by means of binding a fluorescently active protein to the biomarker target or by staining the biomarker target with a stain targeted to the biomarker.

[0062] In the following application, the term "digital pathology image acquisition device" (or digital pathology image scanner) means a specialized and automated digital optical analysis system capable of providing high-resolution and high-quality digital images of pathology sections (e.g., histopathology preparation sections or alternatively cytopathology samples). The digital pathology image acquisition device can be connected to a PACS system via a computer network, for example to enable further analysis of the images obtained on a remote computer in the computer network.

[0063] Therefore, the basic idea of ​​the present application is to provide a feedback mechanism to alert the user in cases where the automatic assessment of digital pathology images is at high risk of being inaccurate due to internal thresholds of the image processing algorithm used. It is proposed to represent the sensitivity of the pathology score relative to the selected threshold of the algorithm used in the analysis pipeline so that the pathologist can take this into account when establishing the diagnosis.

[0064] These and other aspects of the invention will be apparent from and elucidated with reference to the embodiments described hereinafter. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] Exemplary embodiments of the present invention will be described with reference to the following drawings:

[0066] Figure 1 shows a schematic illustration of a method according to the third aspect;

[0067] Figure 2 a) to d) schematically illustrate the impact of changes in algorithm parameters on the digital image analysis pipeline;

[0068] Figure 3 a) and b) schematically demonstrate the effect of adjusting the detection threshold on detected candidate objects in digital pathology image data;

[0069] Figure 4 A simple digital pathology image analysis pipeline is schematically illustrated;

[0070] Figure 5 An example graph showing tumor proportion score as a function of threshold on DAP uptake is shown;

[0071] Figure 6 Schematic illustration of a complex digital pathology image analysis pipeline;

[0072] Figure 7 a) and b) schematically illustrate examples of output displays of sensitivity information of digital pathology image data;

[0073] Figure 8 schematically illustrates an apparatus according to the first aspect; and

[0074] Fig. 9 A digital pathology system according to the second aspect is schematically illustrated. DETAILED DESCRIPTION

[0075] Immunohistochemical staining is widely used in the diagnosis of abnormal cells such as those found in cancerous tumors. Specific molecular markers are the characteristics of specific cell events, such as proliferation or cell death (apoptosis). Antibody-antigen interaction visualization can be completed in many ways. In the most common example, antibodies can be conjugated to enzymes, such as peroxidase, which can catalyze color development reactions (immunoperoxidase staining), alternatively, antibodies can also be labeled to fluorophores such as fluorescein or rhodamine (immunofluorescence).

[0076] The article “PD-L1 Immunohistochemistry Assays for Lung Cancer: Results from Phase 1 of the Blueprint PD-L1 IHC Assay Comparison Project” by Hirsch et al. in Journal of Thoracic Oncology, Volume 12, Number 2, Pages 208-222 discusses the analytical and clinical comparability of four PD-L1 IHC assays used in clinical trials.

[0077] Specific staining protocol benchmarks enable pathology scores to be obtained based on pathology samples. For example, the public "Interpretation Manual for PD-L1IHC 22C3 pharmDx is CE-IVD-Marked" by Dako (TM) (Agilent Technology Solutions (TM)) uses a quantitative immunohistochemistry assay of monoclonal mouse anti-PD-L1 Clone 22C3 for detection of PD-L1 in formalin-fixed, paraffin-embedded non-small cell lung cancer and melanoma tissues, and uses this test to derive a "tumor proportion score" as an indicator for treatment with "KEYTRUDA (TM)". However, the skilled person will appreciate that many pathology protocols involving staining are available and benefit from the application of the techniques discussed in this application. The PD-L1 test discussed in this application is provided as an example.

[0078] For example, reference "Manual interpretation for PD-L1 IHC 22C3 pharmDx is CE-IVD-Marked" shows an example of PD-L1 stained tissue. Brown indicates overexpression of the PD-L1 marker at the cell membrane. Arrows indicate local changes in the intensity of the expression of the marker. Blue cells are stained with a counterstain.

[0079] Typically, the scoring of immunohistochemical markers includes two steps. First, the segmentation of cells or regions of interest is performed (e.g., based on their appearance). Second, the quantification of the staining intensity in the region of interest is performed. Typically, distinction is made between two or more intensity levels. As an example, in the case of "Her2" staining, four levels (0, 1+, 2+, and 3+) are distinguished based on the color intensity of the membrane of the tumor cell, and the area fraction staining of its membrane. Similar principles apply to nuclear staining protocols similar to ER and PR.

[0080] A typical immunohistochemical assessment has many key steps. For example, viable tumor cells may need to be distinguished from other cell types. Alternatively or in combination, cells expressing a certain biomarker need to be distinguished from cells that do not express a certain biomarker. These assessments can be performed digitally by analyzing digital pathology images using image processing algorithms. However, the image processing algorithm may have one or more decision thresholds.

[0081] For example, it may be necessary to decide how to classify a nucleus as belonging to a viable tumor cell. The resulting probability may be thresholded at 0.5, but other levels may be selected, such as a lower (or more conservative) or higher (more aggressive) threshold. Alternatively, the detection threshold and / or detection probability applied in the digital imaging algorithm may be set according to a benchmark using expert judgment. For example, a machine learning algorithm (deep learning algorithm) trained on a corpus of example digital pathology images of the target condition may be provided, where annotation has been performed by an expert.

[0082] In most cases, the exact value of the threshold used in the algorithm will not significantly change the results of the pathology score. However, when the pathology score is closer to the clinical decision level, or if a given test is highly sensitive to one of the thresholds or parameters used in the image processing algorithm, inaccuracy in the assessment of the pathology score may become a problem. Solutions to these problems are now discussed.

[0083] The present invention will first be described broadly in terms of the steps of the method of the third aspect and alternative embodiments thereof.

[0084] A third aspect provides a method for digitally processing digital pathology image data to generate pathology scoring data and sensitivity analysis of the pathology scoring data, comprising:

[0085] a) obtaining 10 digital pathology image data, including images of pathology samples;

[0086] b) classifying 11 objects in the digital pathology image data into a plurality of candidate objects;

[0087] c) assigning 12 a first pathological state to at least one candidate subject of a plurality of candidate subjects according to one or more detection thresholds and / or detection probabilities;

[0088] d) obtaining 13 initial pathology scoring data of the pathology sample based on a candidate object among a plurality of candidate objects to which the first pathology status has been assigned;

[0089] e) perturbing 14 one or more detection thresholds and / or detection probabilities according to the perturbation function to generate a perturbation detection threshold and / or a perturbation detection probability;

[0090] f) assigning 15 the first pathological state to at least one candidate object among the plurality of candidate objects according to one or more disturbance detection thresholds and / or disturbance detection probabilities;

[0091] g) obtaining 16 updated pathology score data for the pathology sample based on the candidate object to which the first pathology state has been assigned according to one or more perturbation detection thresholds and / or perturbation detection probabilities, wherein a comparison of the initial pathology score data and the updated pathology score data defines a sensitivity of the pathology score data.

[0092] The skilled person will appreciate that the following algorithm is used as a digital image processing algorithm as implemented by, for example, a computer processor.

[0093] In step a), initial image data required for generating pathology scoring data and associated sensitivity analysis is obtained. In the case where sensitivity analysis of a digital pathology image is required close to the time of its acquisition, for example, the digital pathology image can be obtained from a digital light microscope or a digital fluorescence microscope as part of a digital pathology system. In the case where the sensitivity analysis is to be performed on archived images, the original digital pathology images stored in a PACS archive can be obtained from the PACS system via the computer network of the medical facility and transmitted to, for example, a PC terminal, a handheld computer, or another computer processing device. Therefore, the technology is applicable to archived digital pathology images and the time of image acquisition.

[0094] Unprocessed digital pathology images contain visual representations of many objects, such as cell nuclei, portions of cell nuclei, cell membranes, cancer cells, fibroblasts, immune cells, etc. For a certain immunohistochemical protocol, it is likely that a subset of these objects will be stained and analyzed. In the specific example of the "PD-L1 IHC22C3" test mentioned previously, it is often desirable to identify viable tumor cells in digital pathology images.

[0095] Optionally, a pre-processing step may be provided, including smoothing the raw digital pathology image to suppress noise. In the case of hematoxylin stained data, typically the pixels of the raw digital pathology image obtained in step a) are also subject to color characterization to remove pixels having a color component less than, for example, 10% of the full-scale intensity. Of course, the color characterization algorithm has configuration parameters that can be thresholded according to the present method. Optionally, a morphological closing algorithm is subsequently applied to the smoothed and color characterized digital pathology image to correct bright spots in the chromatin pattern.

[0096] In step b), there is therefore a classification of the objects in the digital pathology image data into candidate objects (which will be required in the subsequent analysis) and other objects that are not required in the subsequent analysis. In the case of the "PD-L1 IHC 22C3" test, the candidate objects are viable tumor cells. Thus, cytoplasmic staining artifacts, immune cells, normal cells, necrotic cells and debris are excluded from the analysis.

[0097] Optionally, the classification process can be performed in part or in whole according to a deep learning algorithm. In such an approach, following preprocessing of all slide images (digital pathology image data), various types of deep learning methods that can be applied to the present technology are broadly defined as supervised learning, unsupervised learning, semi-supervised learning, and multi-instance learning.

[0098] A deep learning algorithm comprises a cascade of layers that treat the output of the previous layer as the input of the subsequent layer. The components in each layer are typically nonlinear functions that perform tasks such as feature extraction. A deep learning algorithm may include unsupervised and / or supervised learning phases. In a deep learning algorithm, a critical parameter is typically a threshold on a probability (the output of a softmax layer) or a threshold on a derived uncertainty / (one or more) confidence measure in the output of the network.

[0099] Alternatively, the classification process can be performed in part or in whole using a supervised learning algorithm. Supervised learning techniques infer a function that represents the mapping between items in the digital input image data to their appropriate labels (such as "T-cells"). Examples of supervised learning algorithms are convolutional neural networks (CNNs) or support vector machines (SVMs).

[0100] Alternatively, the classification process can be performed partially or entirely using an unsupervised learning algorithm. Unsupervised learning techniques infer functions that describe hidden structures in unlabeled images. Examples of unsupervised learning methods are principal component analysis and k-means analysis.

[0101] Optionally, the classification process may be performed partially or entirely using a semi-supervised learning algorithm.

[0102] Optionally, the classification process may be performed partially or entirely using a multi-instance algorithm.

[0103] The skilled artisan will appreciate that other objects or combinations of objects may be identified depending on the specific immunohistochemical protocol to be applied. In the specific case of the "PD-L1 IHC 22C3" protocol, at least 100 candidate objects should be classified before generating a test score, although for other pathology test protocols, more or less than 100 candidate objects may need to be identified.

[0104] Optionally, the technique for finding viable tumor cells and assigning their status as candidate objects will apply a morphological opening algorithm on the digital pathology image to remove small cells and fibroblasts. Then, a morphological closing algorithm is applied to enlarge adjacent nuclei in the digital pathology image to form an object. Subsequently, a hole filling algorithm is applied in the tumor region to remove background areas. Finally, further iterations of the morphological opening algorithm are applied to smooth the boundaries of the tumor region, thereby compensating for artifacts resulting from previous morphological algorithm operations.

[0105] Optionally, a variety of image processing techniques may be used to perform classification of objects in digital pathology image data into candidate objects, such as deep learning, segmentation, feature extraction, unsupervised learning, clustering, K-means, principal component analysis, or supervised learning methods such as support vector machines or convolutional neural networks.

[0106] It is important to note that the image processing algorithms discussed in the preceding paragraphs are configurable using a wide variety of parameters. As a simple example, an optional pre-processing step may threshold the pixels of the digital pathology image at 10% intensity values, but instead a value of 5% or 15% may be selected. There is a complex interaction between the thresholding value selected and which objects in the digital pathology image may ultimately be classified as, for example, candidate objects in step b). The morphological algorithms and classification algorithms used to classify objects in step b) also have a full range of configuration parameter sets that affect the classification of candidate objects.

[0107] Figure 2 a) illustrates a digital pathology image 20 in which a solid mask 22 has been used to indicate areas belonging to connective tissue or background areas which are therefore not viable tumor cells 24 .

[0108] In step c), the first pathological state (histopathological state, cytopathological state) is assigned to one or more candidate objects in a plurality of candidate objects according to one or more detection thresholds and / or detection probabilities. Given the wide range of digital pathology tests to which the present invention can be applied, assigning the first pathological state to the candidate objects can be considered a small step in determining the subsequent diagnostic results. As a specific example, part of the "PD-L1 IHC 22C3" protocol is to calculate the number of viable tumor cells in a given area showing positive cell membrane staining for the PD-L1 biomarker. In this specific example, the candidate objects are viable tumor cells, and the assignment of the first pathological state is that the white object considered is "positive for the PD-L1 biomarker", or its equivalent digital representation. For example, on a given area of ​​a digital pathology image, the candidate objects can be represented in a data structure by a column vector, wherein a logical "zero" indicates no PD-L1 biomarker expression and a logical "one" indicates PD-L1 biomarker expression. Alternatively, the data structure may include a column vector representing the relative proportion of PD-L1 biomarker expression.

[0109] In step d), an initial pathology score (or initial histopathology score, initial cytopathology score) on the pathology sample is generated. In the case of the specific "PD-L1 IHC 22C3" protocol currently considered, this step includes calculating a "tumor proportion score" as the initial pathology score, which is the percentage of viable tumor cells (candidates) showing partial or complete membrane staining relative to all viable tumor cells (positive and negative) present in the sample. However, the skilled person will understand that for different protocols, different types of cells may be identified as candidates and different calculations may be made to obtain the initial pathology score.

[0110] Optionally, the initial pathology score may be displayed to a user via a graphical user interface (GUI) using one of the techniques discussed below with respect to output methods.

[0111] Typically, the user will be satisfied with the initial pathology score so obtained.However, as discussed with respect to steps b) and c), the image processing and classification techniques are parameterizable, and small changes in how the image processing and classification algorithms are parameterized may result in changes in the pathology score.

[0112] Figure 3a) schematically illustrates a digital pathology image 30 to which steps a), b) and c) have been applied. Hollow circles 32i represent non-viable tumor cells (with respect to a particular example) and filled circles 34i illustrate viable tumor cells. In the digital pathology image 30, a pre-processing step of thresholding the raw digital pathology image to pixels with an intensity greater than IT1=0.1 has been applied, which results in approximately 50% of the identified objects in the image being classified as candidate objects (such as viable tumor cells).

[0113] Figure 3 b) schematically illustrates a digital pathology image 36 to which a pre-processing step has been applied which thresholds the raw digital pathology image to pixels having an intensity greater than IT2=0.15. In this case, it appears that many viable tumor cells have been erroneously removed from the digital pathology image by the thresholding step, which results in approximately 12.5% ​​of the identified objects in the image being classified as candidate objects (such as viable tumor cells). This is an example of how a small change in an algorithm parameter may result in a significant change in the pathology results.

[0114] In step e), the perturbation is applied to the detection threshold and / or the detection probability according to a perturbation function. In other words, after the initial pathology score data has been calculated using a cutoff pixel intensity of IT1=0.1 Figure 3 In the simple example considered in a) and 3b), a disturbance detection threshold of IT2 = 0.15 has been used as disturbance detection threshold. In this simple example, the disturbance function is therefore a positive step function of 0.05.

[0115] Figure 4 Illustrated is the generation of a simple perturbation function for one variable (initial pixel thresholding) of a digital pathology image analysis pipeline.

[0116] The perturbation function generator 44 provides a column vector IT3, including three values ​​of the image level threshold. An initial digital pathology image 40 including objects that may or may not be classified as candidate objects 42 is input, and a thresholding function 46 is applied three times to generate three intermediate images generated at three different threshold levels generated by the perturbation function. The three intermediate images are input to an image classifier 48 to generate digital pathology output images 50a, 50b, and 50c including different numbers of candidate objects depending on the level of the perturbation function applied to the initial threshold.

[0117] Of course, the perturbation function generator 44 can generate the perturbation function in many different ways. The skilled person will appreciate that the output of the perturbation function will be scaled to a range of values ​​suitable for the stage in the analysis pipeline and the specific function to which the perturbation is applicable. Optionally, the perturbation function generator 44 is configured to randomly select the detection threshold and / or the detection probability from a probability distribution, such as a Gaussian distribution, a uniform distribution, a chi-square distribution, etc. The perturbation function generator 44 can be configured to select the detection threshold from a pre-provided range of values, an algebraic function, a random number generator operating within a predetermined scale, a chaotic attractor function, etc.

[0118] The perturbation function may be based on historically acquired calibration and training data. For example, a population of pathologists may be used to classify test slides, and this would enable a realistic range of perturbation values ​​to be derived from the observed variation in results from the population of pathologists. The derived realistic range may then be provided as a perturbation function.

[0119] In step f), the following steps are provided: reassigning the first pathological state to one or more candidate objects of the plurality of candidate objects, or at least one candidate object of the plurality of candidate objects of the disturbance according to one or more disturbance detection thresholds and / or disturbance detection probabilities.

[0120] Using a perturbation detection threshold, the first pathological state can be used to detect a smaller or larger number of candidate objects among multiple candidate objects, depending on the amplitude of the perturbation and the sensitivity of the analysis pipeline to the perturbation. Optionally, the parameter perturbations applied in the preprocessing and / or classification steps, which means generating multiple candidates for the perturbations. In other words, with reference to the specific example of PD-L1 biomarker detection, the perturbations of the algorithm applied in steps a) and b) may result in some objects initially included in the multiple candidate objects to be excluded from the multiple candidate objects to be perturbed, and vice versa. Subsequently, the algorithm used to assign the first pathological state to one or more viable tumor cells (candidate objects) may also be sensitive to changes in the detection threshold and / or detection capability.

[0121] In step g), an updated pathology score for the pathology (histopathology, cytopathology) sample is provided based on the candidate objects that have been assigned the changed number of the first pathological state. Since the updated pathology score has been calculated based on the perturbed portion threshold and / or probability, it is likely that it will be slightly (or significantly) different from the initial pathology score. Therefore, the comparison between the initial pathology score data and the updated pathology score data defines the sensitivity of the pathology score data obtained using a particular analysis pipeline.

[0122] Alternatively, a detection threshold and / or a detection probability may be considered as a detection condition.

[0123] Figure 2b-2d) illustrate the results of a specific example of the use of the PD-L1 biomarker (negative and positive expression) in determining the tumor proportion score (TPS) present in the digital pathology image slice 20. The grey shaded areas indicated as 26b, 26c and 26d respectively show areas with overexpression of the PD-L1 biomarker for three different intensity thresholds.

[0124] exist Figure 2 In b), a pixel intensity threshold at 0.1 is applied, resulting in an overall TPS of 33%.

[0125] exist Figure 2 In c), a pixel intensity threshold of 0.15 is applied, resulting in an overall TPS of 5%.

[0126] exist Figure 2 d), a pixel intensity threshold of 0.2 was applied, resulting in an overall TPS of 1%. Significantly, a TPS of 1% is borderline for conditions where no PD-L1 expression occurs, which represents a critical clinical decision boundary for treatment with certain compounds. Figure 2 The graphical representation of sensitivity analysis and variable TPS scores can automatically provide results that enable medical experts to more accurately and confidently assess the sensitivity of digital pathology scores in a clinical context.

[0127] Although many references have been made to a specific test for the Dako(TM) PD-L1 biomarker, the skilled artisan will appreciate that this specific test involves a complex assessment of positive and negative biomarker expression. Sensitivity analysis of digital pathology image data to perturbations of internal analysis pipeline parameters can also be performed for simpler pathology, histopathology, or cytopathology methods.

[0128] Optionally, one or more detection thresholds are associated with the sensitivity of the test "PD-L1" pathology score.

[0129] For example, the "CD3" test is often used as a marker for T cells in classifying lymphomas.

[0130] When using, for example, a digital light microscope to obtain digital pathology image data of a sample according to the CD3 protocol to collect a stained portion package, either a software imaging package is used to obtain digital pathology image data of the sample at different magnifications (4x, 10x, 20x, 40x), or multiple digital pathology images at different magnifications can also be collected by a digital light microscope. Then, the image is segmented into different tissue regions of interest containing tumor cells and objects of interest by one or more critical detection thresholds (parameters). Following this segmentation process, a distinction can be made between different staining amounts (expression levels) in tumor tissue (including tumor cells and objects of interest). This difference between different regions of the same tumor is called tumor heterogeneity. Therefore, a specific expression type or a first pathological state is assigned to certain regions of tumor tissue, while another expression type and a second pathological state can be assigned to another region of tumor tissue and even another expression type and at least a pathological state can be assigned to another region.

[0131] In this example, CD3 staining illustrates i) intratumoral lymphocytes of lymphocytes located in the stroma of tumor lesions or within tumor cell nests and ii) peritumoral lymphocytes of lymphocytes surrounding tumor lesions. For lymphocytes, distribution and density scores are applied. Scores range from 0 to 6. Lymphocyte distribution scores ranging from 0 to 3 are defined as follows: 0 = lack of lymphocytes within the tissue, 1 = presence of lymphocytes occupying <25% of the tissue, 2 = presence of lymphocytes occupying 25% to 50% of the tissue, and 3 = presence of lymphocytes occupying >50% of the tissue. Lymphocyte density ranging from 0 to 3 is defined as follows: 0 = lack, 1 = moderate, 2 = moderate, and 3 = severe.

[0132] The data obtained can be used for sensitivity analysis, the parameters of the algorithm for segmenting and / or identifying lymphocytes are a detection threshold (parameter) adapted to the perturbation according to the described algorithm. Moreover, the detection threshold can also be manually adjusted to define. This definition assists the pathologist during the evaluation of the tissue to prove whether the evaluation of the tissue is reasonable and within the objectively defined range.

[0133] Optionally, one or more detection thresholds are associated with the sensitivity of the test "CD3" pathology score.

[0134] For example, the "CD8" protocol can be used to identify effector T cells in tumor tissue.

[0135] When using, for example, a digital light microscope to obtain digital pathology image data of a sample according to a CD8 protocol to acquire a stained portion package, either a software imaging package is used to obtain digital pathology image data of the sample at different magnifications (4x, 10x, 20x, 40x), or multiple digital pathology images at different magnifications can also be acquired by the digital light microscope.

[0136] Then, the image is digitally segmented into different tissue regions of interest containing tumor cells and objects of interest by one or more critical thresholds. Following the segmentation process, a digital distinction is made between different staining concentrations (expression levels) in tumor tissue (including tumor cells and objects of interest). This difference between different regions of the same tumor is called tumor heterogeneity. Therefore, a specific expression type or a first pathological state is assigned to certain regions of tumor tissue, while another expression type and a second pathological state can be assigned to another region of tumor tissue and even another expression type and at least a pathological state can be assigned to another region.

[0137] In this example, CD8 staining illustrates i) intratumoral lymphocytes of lymphocytes in the stroma of tumor lesions or tumor cell nests and ii) peritumoral lymphocytes of lymphocytes surrounding tumor lesions. Factors affecting lymphocyte distribution and density scores are how accurately the stroma region is segmented, for example, and / or the inspection threshold for CD8 positive calls and / or the threshold for nuclear detection. Scoring varies from 0 to 6 ranges. Lymphocyte distribution scores ranging from 0 to 3 are defined as follows: 0 = lack of lymphocytes in tissues, 1 = presence of lymphocytes occupying <25% of tissues, 2 = presence of lymphocytes occupying 25% to 50% of tissues, and 3 = presence of lymphocytes occupying >50% of tissues. Lymphocyte density ranging from 0 to 3 is defined as follows: 0 = lack, 1 = moderate, 2 = moderate, and 3 = severe.

[0138] The data obtained can be used for a sensitivity analysis according to the algorithm described in the application, wherein the assignment to the first and second pathological states is an example of a threshold value (parameter) suitable for perturbation. A threshold value is defined by comparing different areas of the tumor with its surrounding tissue, thereby obtaining different threshold values. Moreover, the detection threshold value can also be manually adjusted. The definition assists the pathologist during the evaluation of the tissue to prove whether the evaluation of the tissue is reasonable and within the objectively defined range.

[0139] Optionally, one or more detection thresholds are associated with the sensitivity of the test "CD8" pathology score.

[0140] For example, the "ER" protocol (estrogen receptor) is often used in connection with breast screening programs. In acquiring digital pathology image data of a sample according to the ER protocol using, for example, a digital light microscope to acquire a stained section package, either digital pathology image data of the sample at different magnifications (4x, 10x, 20x, 40x) is acquired using a software imaging package, or multiple digital pathology images at different magnifications can also be acquired by the digital light microscope.

[0141] Then, the image is digitally segmented into different tissue regions of interest containing tumor cells and objects of interest by one or more critical thresholds. Following the segmentation process, a digital distinction is made between different expression levels in tumor tissue (including tumor cells and objects of interest). This difference between different regions of the same tumor is called tumor heterogeneity. Therefore, a specific expression type or a first pathological state is assigned to certain regions of tumor tissue, while another expression type and a second pathological state can be assigned to another region of tumor tissue and even another expression type and at least a pathological state can be assigned to another region.

[0142] In this example, the total score (TS) is the sum of the ratio score (PS) and the intensity score (IS), ranging from 0; 2-8. Positive results for ER and PR are defined as TS≥3. A ratio score (PS) representing the proportion of tumor cells with positive nuclear staining is assigned. PS ranges from 0 to 5, where 0=0, 1=0-1 / 100, 2=>1 / 100-1 / 10, 3=>1 / 10 to 1 / 3, 4=>1 / 3 to 2 / 3 and 5=>2 / 3 to 1. An intensity score (IS) representing the average staining intensity of all positive tumor cells is assigned. IS ranges from 0 to 3, where 0=negative, 1=weak, 2=intermediate and 3=strong. Often, ER and PR analyses are combined in one diagnostic test.

[0143] The obtained data can be used for sensitivity analysis, the ratio score is an example of a threshold (parameter) suitable for perturbation according to the algorithm described herein and the intensity score is an example of another threshold (parameter) suitable for perturbation. By comparing different areas of the tumor, thus obtaining different thresholds, a threshold is defined. Moreover, the detection threshold can also be manually adjusted. This definition assists the pathologist during the evaluation of the tissue to prove whether the evaluation of the tissue is reasonable and within the objectively defined range.

[0144] Optionally, one or more detection thresholds are associated with the sensitivity of the test "ER" pathology score.

[0145] When using, for example, a digital light microscope to obtain digital pathology image data of a sample according to a PR protocol to acquire a stained portion package, either a software imaging package is used to obtain digital pathology image data of the sample at different magnifications (4x, 10x, 20x, 40x), or multiple digital pathology images at different magnifications can also be acquired by the digital light microscope.

[0146] Then, the image is digitally segmented into different tissue regions of interest containing tumor cells and objects of interest by one or more critical thresholds. Following the segmentation process, a distinction is made between different expression levels in tumor tissue (including tumor cells and objects of interest). This difference between different regions of the same tumor is called tumor heterogeneity. Therefore, a specific expression type or a first pathological state is assigned to certain regions of tumor tissue, while another expression type and a second pathological state can be assigned to another region of tumor tissue and even another expression type and at least a pathological state can be assigned to another region.

[0147] In this example, the total score (TS) is the sum of the ratio score (PS) and the intensity score (IS), ranging from 0; 2-8. Positive results for ER and PR are defined as TS≥3. A ratio score (PS) representing the proportion of tumor cells with positive nuclear staining is assigned. PS ranges from 0 to 5, where 0=0, 1=0-1 / 100, 2=>1 / 100-1 / 10, 3=>1 / 10 to 1 / 3, 4=>1 / 3 to 2 / 3 and 5=>2 / 3 to 1. An intensity score (IS) representing the average staining intensity of all positive tumor cells is assigned. IS ranges from 0 to 3, where 0=negative, 1=weak, 2=intermediate and 3=strong. Often, ER and PR analyses are combined in one diagnostic test.

[0148] The data obtained can be used for sensitivity analysis according to the algorithm described in the present application. For example, the ratio score is an example of a threshold (parameter) suitable for perturbation, and the intensity score is an example of another parameter suitable for perturbation. By comparing different areas of the tumor, thus obtaining different thresholds, a threshold is defined. Moreover, the detection threshold can also be manually adjusted. The definition assists the pathologist during the evaluation of the tissue to prove whether the evaluation of the tissue is reasonable and within the objectively defined range.

[0149] Optionally, one or more detection thresholds are associated with the sensitivity of the test "PR" pathology score.

[0150] For example, the "HER2" protocol (Herceptin) is used to detect the presence of abnormal levels of Herceptin in breast tissue, a predictive biomarker for breast cancer. In obtaining digital pathology image data of a sample according to the HER2 protocol using, for example, a digital light microscope to acquire a stained portion package, either a software imaging package is used to obtain digital pathology image data of the sample at different magnifications (4x, 10x, 20x, 40x), or multiple digital pathology images at different magnifications can also be acquired by the digital light microscope.

[0151] Then, the image is digitally segmented into different tissue regions of interest containing tumor cells and objects of interest by one or more critical thresholds. Following the segmentation process, different expression levels in tumor tissue (including tumor cells and objects of interest) are distinguished. This difference between different regions of the same tumor is called tumor heterogeneity. Therefore, a specific expression type or a first pathological state is assigned to certain regions of tumor tissue, while another expression type and a second pathological state can be assigned to another region of tumor tissue and even another expression type and at least a pathological state can be assigned to another region.

[0152] In this example, HER2-negative tumors (score 0) retain membrane HER2 expression of tumor cells <10%. A score of +1 corresponds to membrane HER2 expression in >10% of tumor cells, while a score of +2 corresponds to tumor cells with weak to moderate HER2 expression in the membrane of >10% of tumor cells. Strong complete membrane staining in >10% of tumor cells corresponds to a score of +3. The obtained data can be used for sensitivity analysis according to the algorithm of the application, the membrane, i.e., spatial information is an example of a threshold (parameter) for perturbation and the intensity score is an example of another parameter suitable for perturbation. By comparing different areas of the tumor, different thresholds are thus obtained, and a threshold is defined. Moreover, the defined detection threshold can also be adjusted manually. This definition assists the pathologist during the evaluation of the tissue to prove whether the evaluation of the tissue is reasonable and within the objectively defined range.

[0153] Optionally, one or more detection thresholds are associated with the sensitivity of the test "HER2" pathology score.

[0154] For example, the "EGFR" protocol (epidermal growth factor receptor) test can be used to test for cell mutations that lead to overexpression of EGFR, which has been associated with many cancers.

[0155] When using, for example, a digital light microscope to obtain digital pathology image data of a sample according to an EGFR protocol to acquire a stained portion package, either a software imaging package is used to obtain digital pathology image data of the sample at different magnifications (4x, 10x, 20x, 40x), or multiple digital pathology images at different magnifications can also be acquired by the digital light microscope.

[0156] Then, the image is digitally segmented into different tissue regions of interest containing tumor cells and objects of interest by one or more critical thresholds. Following the segmentation process, different expression levels in tumor tissue (including tumor cells and objects of interest) are distinguished. This difference between different regions of the same tumor is called tumor heterogeneity. Therefore, a specific expression type or a first pathological state is assigned to certain regions of tumor tissue, while another expression type and a second pathological state can be assigned to another region of tumor tissue and even another expression type and at least a pathological state can be assigned to another region.

[0157] In this example, EGFR is preferably calculated for colon tumors as follows: EGFR-negative tumors do not have membrane staining above background in all tumor cells. Conversely, EGFR-positive staining, and therefore, expression is defined as any IHC staining of the tumor cell membrane above background levels; whether it is complete or incomplete circumferential staining. The staining intensity and therefore expression is a score of +1, +2 or +3, where more than 0% of the tumor cells are stained and therefore, positive for EGFR. The data obtained can be used for sensitivity analysis according to the algorithm described in this application, and the intensity score is a threshold (parameter) for perturbation. By comparing different areas of the tumor, different thresholds are obtained, and a threshold is defined. Moreover, the detection threshold can also be manually adjusted. This definition assists the pathologist during the evaluation of the tissue to prove whether the evaluation of the tissue is reasonable and within the objectively defined range.

[0158] Optionally, one or more detection thresholds are associated with the sensitivity of the test "EGFR" pathology score.

[0159] For example, the "Ki67" protocol detects the Ki7 antigen. Following the standard procedure, a digital pathology of a sample prepared according to "Ki67" is obtained.

[0160] When using, for example, a digital light microscope to obtain digital pathology image data of a sample according to a Ki67 protocol to acquire a stained portion package, either a software imaging package is used to obtain digital pathology image data of the sample at different magnifications (4x, 10x, 20x, 40x), or multiple digital pathology images at different magnifications can also be acquired by the digital light microscope.

[0161] Then, the image is digitally segmented into different tissue regions of interest containing tumor cells and objects of interest by one or more critical thresholds. Following the segmentation process, different expression levels in tumor tissue (including tumor cells and objects of interest) are distinguished. This difference between different regions of the same tumor is called tumor heterogeneity. Therefore, a specific expression type or a first pathological state is assigned to certain regions of tumor tissue, while another expression type and a second pathological state can be assigned to another region of tumor tissue and even another expression type and at least a pathological state can be assigned to another region. There are different methods to obtain Ki67 scores. One method involves selecting five different regions of a tumor, wherein 100 tumor cells in each region are evaluated. The percentage of positive Ki67 cells (stained nuclei) in 100 cells is considered and the results of the five regions are added up.

[0162] The data obtained can be used for sensitivity analysis, spatial information, i.e., nuclear staining is a threshold (parameter) that undergoes perturbations. By comparing different areas of the tumor, thus obtaining different thresholds, a threshold is defined. Moreover, the detection threshold can also be manually adjusted. This definition assists the pathologist during the evaluation of the tissue to prove whether the evaluation of the tissue is reasonable and within the objectively defined range.

[0163] Optionally, one or more detection thresholds are associated with the sensitivity of the test "Ki67" pathology score.

[0164] Optionally, one or more detection thresholds and / or detection probabilities represent changes in automatic morphological feature detection when determining the presence of a candidate object in the digital pathology image data.

[0165] Optionally, the one or more detection thresholds represent probabilities of automatically classifying an object in the digital pathology image data into a plurality of candidate objects.

[0166] Optionally, the detection threshold and / or the detection probability characterizes the intensity level and / or wavelength range of light emitted from the pathology sample and represented in the digital pathology image data.

[0167] Optionally, the intensity level and / or wavelength range emitted from the pathology sample and represented in the digital pathology image data is indicative of the relative level of expression of the biomarker.

[0168] Optionally, the method further comprises:

[0169] f1) repeating steps c) to g) to obtain a plurality of disturbance detection thresholds and / or disturbance detection probabilities forming a sensitivity function of a pathology score of the digital pathology image data;

[0170] f2) optionally, calculating the relative detection threshold change required to provide a change in the pathology score. Optionally, the method further comprises:

[0171] d1) obtaining initial pathology score data and / or updating pathology score data by calculating the number of candidate objects in the field of interest of the digital pathology data that have been assigned the first pathology state before and / or after the disturbance, or by calculating the percentage score of the number of candidate objects in the field of interest of the digital pathology data that have been assigned the first pathology state before and / or after the disturbance.

[0172] Figure 5 A diagram 52 showing a specific example of a sensitivity function 54 for a tumor proportion score (TPS) obtained using "PD-L1 IHC 22C3" is illustrated. The X-axis represents the threshold for the assessment of the presence of PD-L1 staining. In other words, the X-axis 58 represents the range of potential perturbation values. The Y-axis 60 represents the percentage value of the TPS. For a particular clinical protocol, an optimal value for the pathology score data may be provided (in this case, the dashed line 56 represents the X-axis TPS intercept at 5%). In this case, a threshold of PD-L1 intensity at 0.15 provides an updated pathology score close to that clinical threshold.

[0173] Many of the simple samples discussed above involve detecting positive expression of a single biomarker. However, "PD-L1 IHC" requires detecting both positive and negative expression of the PD-L1 biomarker. The following example addresses this situation:

[0174] Optionally, the method further comprises:

[0175] c1) assigning a second pathological state to at least one candidate subject showing positive expression of the first biomarker of interest;

[0176] c2) assigning a third pathological state to at least one candidate subject showing negative expression of the first biomarker of interest, and

[0177] d2) obtaining initial pathology score data by performing calculations using the candidate object in the field of interest of the digital pathology data in the second and third pathological states;

[0178] e1) perturbing the detection thresholds used to assign the second and third pathological states to the candidate objects;

[0179] e2) reassigning the second and third pathological states to the candidate subject after the perturbation; and

[0180] g1) obtaining updated pathology score data by performing calculations using the candidate objects in the second and third pathology states after the perturbation.

[0181] Optionally, multi-biomarker imaging can be supported by the technology. In other words, in step c3), the second pathological state can be assigned to at least one candidate showing positive expression of the second biomarker of interest. In step c4), the third pathological state can be assigned to at least one candidate showing positive expression of the third biomarker of interest. Optionally, the first, second and third biomarkers are immunofluorescent biomarkers.

[0182] An immunohistochemistry example of a more complex analysis pipeline related to the derivation of TPS (Tumor Proportion Score) in PD-L1 stained tissue includes the following optional steps:

[0183] 1) Obtain IHC stained histopathology images at relatively high magnification (usually at 10x, 20x, or 40x).

[0184] 2) Apply an algorithm that (semi-)automatically segments the image into objects of interest using one or more critical threshold parameters.

[0185] 3) Applying algorithms that (semi-)automatically distinguish between normal expression and overexpression in the subject of interest using one or more critical thresholds or parameter settings.

[0186] 4) Perform sensitivity analysis as described in this application.

[0187] 5) Provide visual feedback or an alert message if the sensitivity analysis indicates that the score may change from one decision level to another as an actual change in the critical threshold occurs.

[0188] 6) Providing visual feedback of the areas of the digital pathology image data that are most sensitive to threshold changes.

[0189] 7) Provide visual feedback of areas that do not meet the criteria for scoring.

[0190] In one or more of steps 2), 3) and / or 4), deep learning or machine learning methods may optionally be applied.

[0191] Therefore, given an initial RGB image of PD-L1 stained tissue, color deconvolution can be performed to separate the absorption of hematoxylin from the absorption of the DAB label using methods as discussed in the paper "A Method for Normalizing Histology Slides for Quantitative Analysis by Macenko et al." cited in the "Background" section above.

[0192] The location of the kernel in the deconvolved image is obtained by the following steps:

[0193] 8) Smoothing of the hematoxylin staining data (to suppress noise) and thresholding at an absorbance level of 0.1.

[0194] 9) Apply morphological closing to correct for bright spots in the chromatin pattern.

[0195] 10) Morphological opening was applied to remove small cells and fibroblasts.

[0196] 11) Apply morphological closing to grow adjacent nuclei together to form an object.

[0197] 12) Apply hole filling to the tumor area and remove the background area.

[0198] 13) A final morphological opening is performed to smooth the boundaries of the tumor region to compensate for artifacts resulting from previous morphological operations.

[0199] In this case, the denominator of TPS is obtained after uptake of DAB (PD-L1) staining data. DAB uptake above a certain threshold (in regions with viable tumor cells) indicates regions with overexpression, in other words, PD-L1 tumor cells. Figure 2 As mentioned previously in b) to 2d), these images are sensitive to perturbations of the threshold parameters in the analysis pipeline.

[0200] Figure 5 Plotted are TPS as a function of perturbations in the detection threshold on the DAB absorption (x-axis) (y-axis). Figure 5 The "optimal threshold" for the assessment of PD-L1 staining in the case of is indicated by the vertical dashed line. A low detection threshold of 0.05 would have included more tumor cells, which resulted in a significantly higher estimate of TPS. A high detection threshold of 0.05 would have included fewer tumor cells, causing the TPS to drop to the critical clinical decision level of 1%.

[0201] Figure 6 An example of a complex analysis pipeline 61 is illustrated for implementation of the specific example given above.

[0202] The digital pathology image data input unit 62 receives and pre-processes the digital pathology image data. The user provides perturbation configuration data 63 (e.g., via an input graphical user interface on the digital pathology system and / or software analysis software). The analysis pipeline in this particular example is divided into nuclear position 63 determinants 63 and tumor cell identifiers 64.

[0203] In the nuclear position determinant 63, the preparation unit 64a smoothes the hematoxylin staining data (to suppress noise) and thresholds at a given absorption level. The first morphological closing unit 66a corrects bright spots in the chromatin pattern.

[0204] In the tumor cell identifier 64, the morphological opening unit 67a is configured to remove the images of small cells and fibroblasts in the digital pathology image. The morphological closing unit 68a is configured to grow adjacent nuclei together to form one object. The hole filling unit 69a is configured to apply hole filling to the tumor region and remove the background region identified by the low absorption value. The morphological opening unit 70a is applied to smooth the border of the tumor region to thereby compensate for the artifacts resulting from the previous morphological operation.

[0205] In the analysis pipeline, each processing subunit 65a to 70a is connected to a corresponding disturbance determination unit 65b to 70b. In a specific example, the disturbance unit 70b selects the disturbance value of the morphological closing algorithm based on a Gaussian distribution. The rest of the disturbance unit provides the disturbance value based on a step function. The disturbance control unit 71 receives the disturbance configuration data 63 input by the user and possibly determines the disturbance specification to provide sufficient coverage on the total search space of the given disturbance option. Having determined the disturbance specification, the disturbance setting for each of the disturbance units is calculated and transmitted to the disturbance units 65b to 70b. Of course, not all steps of the analysis pipeline need to be characterized by adjustable inspection thresholds and / or probabilities, and one or more settings of the analysis pipeline can be "hard-coded" or fixed without being restricted to disturbances.

[0206] For reasons of computational simplicity, the disturbance control unit 71 may operate a “one at a time” sensitivity analysis protocol, wherein one or more of the disturbance units 65 b to 70 b are adjusted at a time, wherein the other disturbance units are held at a constant disturbance value.

[0207] Optionally, the perturbation control unit 17 may apply a "random sampling" sensitivity analysis protocol, in which one or more of the perturbation units 65b to 70b are addressed with randomly generated perturbation values ​​(optionally within fixed boundaries). In this way, a large sample space of perturbation values ​​may be sampled in a computationally efficient manner, and the cross-correlations between the perturbation settings of the perturbation units 65b to 70b may be evaluated in a manner utilizing a "one at a time" sensitivity analysis protocol.

[0208] Alternatively, the perturbation control unit 71 may apply a "brute force" sensitivity analysis protocol, in which one or more of the perturbation units 65b to 70b are addressed with all combinations of their respective perturbation thresholds. This has the advantage that the sensitivity analysis is exhaustive, although it may take so much computation time that the technique may only be applicable to powerful computers or computational analysis pipelines.

[0209] Alternatively, the perturbation control unit 71 may apply a "gradient descent" algorithm, or another optimization algorithm, to calculate the optimal sensitivity. The result collection unit 72 obtains and saves the pathology scores generated by the application of various combinations of perturbation settings to the analysis pipeline for the same input digital pathology image.

[0210] In an example, the tumor proportion score can be calculated as the number of PD-L1 positive tumor cells present in the plurality of candidates divided by the total number of PD-L1 positive and PD-L1 negative tumor cells present in the plurality of candidates (viable tumor cells).

[0211] Tumor proportion scores can be used to differentiate between no PD-L1 expression (partial or complete cell membrane staining in less than 1% of viable tumor cells), PD-L1 expression (partial or complete cell membrane staining in between 1% and 49% of viable tumor cells), and high PD-L1 expression (partial or complete cell membrane staining in greater than or equal to 50% of viable tumor cells).

[0212] Thus, perturbations in the detection threshold and / or detection capabilities used in the analysis pipeline may provide an alert to the medical professional that the pathology protocol is too sensitive such that it may affect the overall treatment indication.

[0213] Optionally, the method further comprises:

[0214] h1) optionally combining initial and / or updated pathology score data to output the sensitivity of the pathology score data to the user and / or outputting the sensitivity of the pathology score relative to a clinical threshold.

[0215] Thus, the sensitivity of the pathology score data can be displayed on a graphical user interface (GUI) of a digital pathology system, or digital pathology software used on a personal computer (PC) or other digital display device. The pathology score data can be reported to the user as a string of numbers, or as, for example, a color map or heat map. This provides direct and intuitive feedback to the user of the digital pathology system and / or analysis software regarding the sensitivity of a given result in the context of the original digital pathology slide image.

[0216] Optionally, the method further comprises:

[0217] b2) subsampling the digital pathology image data into a plurality of sectors;

[0218] g2) obtaining initial pathology score data and updating the pathology score for each sector;

[0219] h2) generating a spatial sensitivity mask of the digital pathology image data; and

[0220] h3) Displaying the spatial sensitivity mask to the user, optionally as a semi-transparent overlay aligned with the digital pathology image data.

[0221] Figure 7 a) illustrates a possible display format in which digital pathology image data 80 has been divided into a plurality of sub-portions 80a, 80b, 80c, ... In this display embodiment, an initial pathology score, an updated pathology score, and an overall sensitivity for each sub-portion have been calculated. A display legend 82 provides guidance for the interpretation of the GUI. Saturated sub-portions represent low sensitivity to perturbations in the analysis pipeline, and non-saturated sub-portions represent high sensitivity to perturbations in the analysis pipeline.

[0222] Figure 7 b) illustrates another possible display format in which the digital pathology image data 84 has been divided into a plurality of sub-portions 84a, 84b, ... In this display embodiment, the initial pathology score, updated pathology score, and overall sensitivity for each sub-portion have been calculated. In this GUI format, the user can move a mouse cursor 86 around the displayed digital pathology image data 84. An optional dialog box 88 reports the current spatial position of the cursor relative to the origin 90 of the displayed digital pathology image data 84. The sensitivity feedback dialog box displays the calculated sensitivity of the pathology score at the position of the mouse cursor.

[0223] Optionally, a message may be provided to the user in the output step indicating the sensitivity of the analysis. As an example, a calculation of a relative threshold change required for a relative change in score may be provided. Optionally, a required threshold change that will result in a different diagnostic result may be provided.

[0224] Optionally, the pathology sample is a histopathology sample or a cytopathology sample.

[0225] Figure 8 An apparatus 100 according to the first aspect is illustrated.

[0226] According to a first aspect, there is provided an apparatus 100 for digitally processing digital pathology image data to generate pathology scoring data and sensitivity analysis of the pathology scoring data. The apparatus 100 comprises:

[0227] - input unit 102; and

[0228] -Processing unit 104.

[0229] The input unit 102 is configured to obtain digital pathology image data, including images of pathology samples.

[0230] The processing unit 104 is configured to classify objects in the digital pathology image data into multiple candidate objects, assign a first pathological state to at least one candidate object among the multiple candidate objects according to one or more detection thresholds and / or detection probabilities, obtain initial pathological score data of the pathological sample based on the candidate object among the multiple candidate objects to which the first pathological state has been assigned, perturb one or more detection thresholds and / or detection probabilities according to a perturbation function to generate perturbed detection thresholds and / or perturbation detection probabilities, reassign the first pathological state to at least one candidate object among the multiple candidate objects, or reassign the first pathological state to at least one candidate object among the perturbed multiple candidate objects according to one or more perturbation detection thresholds and / or perturbation detection probabilities, obtain updated pathological score data of the pathological sample based on the candidate object to which the first pathological state has been reassigned according to one or more perturbation detection thresholds and / or perturbation detection probabilities, and compare the initial pathological score data with the updated pathological score data to obtain the sensitivity of the pathological score data.

[0231] The device may further include an output unit 106 .

[0232] The input unit 102 may include a data communication modem capable of transmitting digital pathology image data, such as a USB (TM) connection, a FireWire (TM) connection, a DICOM connection, etc. The input unit 102 may include a hard drive and / or a removable hard drive, a USB drive, a DVD drive, or another device for transmitting stored data. The data may be received via a secure communication network, such as a LAN or WAN, or a secure wireless device.

[0233] It will be appreciated that the processing unit 102 may be implemented as virtually any data processor capable of processing image data. For example, the apparatus 100 may be implemented on a personal computer, a smartphone processor, an embedded processor, a digital signal processor (DSP), or a processing unit instantiated on a field programmable gate array (FPGA). Optionally, part or all of the functions performed by the processing unit 100 may be performed using acceleration capabilities available using a graphics processing unit GPU.

[0234] Where the data output is raw data to be interpreted by another component, the output unit 106 may be provided with a similar range of modalities as discussed with respect to the input unit 102. The output unit 16 may also include a graphical interface that displays the results of the sensitivity analysis on a graphical user interface (GUI).

[0235] Fig. 9 A digital pathology system 110 is illustrated.

[0236] According to a second aspect, a digital pathology system 110 is provided. The digital pathology system 110 comprises:

[0237] - A digital pathology image acquisition device 112; and

[0238] - an apparatus 114 according to the first aspect or one of its embodiments,

[0239] The digital pathology image acquisition device 112 is configured to obtain digital pathology image data and transmit the digital pathology image data to the apparatus 114 .

[0240] The computer program element can be stored on a computer unit, which can also be an embodiment of the present invention. The computing unit can be suitable for executing or causing the performance of the steps of the method described above. Moreover, it can be suitable for operating the components of the device described above. The computing unit can be suitable for automatically operating and / or running the user's command. The computer program can be loaded into the working memory of a data processor. The data processor can therefore be equipped to perform the method of the present invention.

[0241] This exemplary embodiment of the invention covers both a computer program with intervention from the beginning being installed and a computer program which by means of an update turns an existing program into a program using the invention.

[0242] Optionally, the slide image data can be uploaded to a "PACS" system or local hospital server via a local area network. Of course, the slide image can be saved on a physical medium, such as a digital versatile disc (DVD), a tape drive, or a USB disk and physically sent to the location of the hosting server, where the physical medium can be loaded onto the server.

[0243] The slide image data is then processed according to the second aspect or an alternative embodiment thereof. The multi-view data of the field of view of the biological sample is then transmitted to a client device for use, for example, by a graphical user interface capable of interpreting the multi-view data. Optionally, the multi-view data is interpreted into a display format (such as .JPG, .GIF or another imaging format) on the server, and a GUI display of the multi-view data is transmitted to the client (an example of a "web-based application").

[0244] The computer program may be stored and / or distributed on a suitable medium (such as an optical storage medium or solid state medium supplied together with or as part of other hardware), and may be distributed in other forms (such as via the Internet or other wired or wireless telecommunications systems). However, the computer program may also be presented on a network like the World Wide Web, and may also be downloaded from such a network into the working memory of a data processor. The image processing method according to the second aspect will then be executed on.

[0245] According to a further exemplary embodiment of the present invention, a medium for producing a computer program element available for downloading is provided, the computer program element being arranged to perform a method according to one of the previously described embodiments of the present invention.

[0246] It should be noted that embodiments of the present invention are described with reference to different subject matters. In particular, some embodiments are described with reference to method type claims, whereas other embodiments are described with reference to device type claims. However, those skilled in the art will understand from the above and the following description that, unless otherwise notified, in addition to any combination of features belonging to one type of subject matter, other combinations between features related to different subject matters are also considered to be disclosed by the present application.

[0247] All features may be combined to provide synergistic effects that exceed the simple sum of the features.

[0248] While the invention has been illustrated and described in detail in the drawings and foregoing description, such illustration and description are to be considered illustrative or exemplary and not restrictive. The invention is not limited to the disclosed embodiments.

[0249] Other variations to the disclosed embodiments can be understood and effected by those skilled in the art in practicing the claimed invention, from a study of the drawings, the disclosure, and the appended claims.

[0250] In the claims, the word "comprising" does not exclude other elements or steps, and the quantifier "a" or "an" does not exclude a plurality. A single processor or other unit may fulfill the functions of several items recited in the claims. The mere fact that certain measures are recited in mutually different dependent claims does not indicate that a combination of these measures cannot be used to advantage. Any reference signs in the claims should not be construed as limiting the scope.

Claims

1. An apparatus (100) for digitally processing digital pathology image data to generate pathology scoring data and performing sensitivity analysis on the pathology scoring data, wherein: The sensitivity analysis means an evaluation of how the pathology score data is affected by a perturbation of a parameter used in determining or generating the pathology score data, the apparatus comprising: - an input unit (102); and - a processing unit (104); wherein the input unit (102) is configured to obtain digital pathology image data, the digital pathology image data comprising an image of a pathology sample; and Wherein, the processing unit (104) is configured to: classifying objects in the digital pathology image data into a plurality of candidate objects, assigning a first pathological state to at least one candidate object of the plurality of candidate objects based on one or more detection thresholds and / or detection probabilities, wherein automatic morphological feature detection is used to identify the candidate objects; obtaining initial pathology score data of the pathology sample based on the candidate objects among the plurality of candidate objects to which the first pathological state has been assigned; perturbing the one or more detection thresholds and / or the detection probabilities according to a perturbation function to generate perturbation detection thresholds and / or perturbation detection probabilities, wherein the perturbation function is configured to generate the perturbation detection thresholds and / or the perturbation detection probabilities, including selection using a step function, random selection using a probability distribution, selection using a pre-provided range of values, selection using an algebraic function, selection using a random number generator, or selection using a chaotic attractor function; reassigning the first pathological state to at least one candidate object among the plurality of candidate objects or reassigning the first pathological state to at least one candidate object among the plurality of candidate objects for disturbance according to one or more disturbance detection thresholds and / or disturbance detection probabilities; and Obtaining updated pathology score data of the pathology sample based on the candidate object to which the first pathological state has been reassigned according to the one or more perturbation detection thresholds and / or the perturbation detection probability, and comparing the initial pathology score data with the updated pathology score data to obtain the sensitivity of the pathology score data.

2. The device (100) according to claim 1, in, The one or more detection thresholds and / or detection probabilities characterize changes in the automatic morphological feature detection when determining the presence of a candidate object in the digital pathology image data.

3. The device (100) according to one of claims 1 or 2, in, The one or more detection thresholds represent a probability of automatically classifying the object in the digital pathology image data into the plurality of candidate objects.

4. The device (100) according to one of claims 1 or 2, in, The detection threshold and / or the detection probability characterizes an intensity level and / or a wavelength range of light emitted from a pathology sample and represented in the digital pathology image data.

5. The device (100) according to claim 4, in, The intensity level and / or the wavelength range emitted from the pathology sample and represented in the digital pathology image data is indicative of a relative level of expression of a biomarker.

6. The device (100) according to one of claims 1 or 2, in, The processing unit (104) is further configured to: A plurality of perturbation detection thresholds and / or perturbation detection probabilities are obtained and a sensitivity function of the pathology score for the digital pathology image data is generated, and optionally, a relative detection threshold change required to provide a change in the pathology score is calculated.

7. The device (100) according to one of claims 1 or 2, in, The processing unit (104) is configured to: The initial pathology score data and / or the updated pathology score data are obtained by counting the number of candidate objects in the field of interest of the digital pathology data to which the first pathology state has been assigned before and / or after perturbation or by calculating a percentage score of the number of candidate objects in the field of interest of the digital pathology data to which the first pathology state has been assigned before and / or after perturbation.

8. The device (100) according to claim 7, And among them, The processing unit (104) is further configured to: assigning a second pathological state to at least one candidate subject showing positive expression of the first biomarker of interest; assigning a third pathological state to at least one candidate subject showing negative expression of the first biomarker of interest, and obtaining the initial pathology score data by performing calculations using the candidate objects in the field of interest of the digital pathology data in the second pathology state and the third pathology state; perturbing a detection threshold for assigning the second pathological state and the third pathological state to the candidate object and reassigning the second pathological state and the third pathological state to the candidate object after the perturbation; and Updated pathology score data is obtained by performing calculations using the candidate objects in the second pathological state and the third pathological state after the perturbation.

9. The device (100) according to one of claims 1 or 2, further comprising: - output unit (106); Wherein, the output unit (106) is configured to optionally combine the initial pathology score data and / or the updated pathology score data and / or output the sensitivity of the pathology score relative to a clinical threshold, and output the sensitivity of the pathology score data to a user.

10. The device (100) according to claim 9, in, The processing unit (104) is further configured to spatially subsample the digital pathology image data into a plurality of sectors, and wherein the initial pathology score data and the updated pathology score are obtained for each sector, and wherein the processing unit is configured to generate a spatial sensitivity mask for the digital pathology image data using the initial pathology score data and the updated pathology score for each sector; and Therein, the output unit (106) is configured to display the spatial sensitivity mask to a user, optionally as a semi-transparent overlay in spatial alignment superimposed on the digital pathology image data.

11. The device according to claim 1 or 2, in, The pathological sample is a histopathological sample or a cytopathological sample.

12. A digital pathology system (110), comprising: -Digital pathology image acquisition device (112); as well as - a device (114) according to one of claims 1 to 11, Wherein, the digital pathology image acquisition device is configured to: receive a pathology sample, automatically analyze the pathology sample and thereby acquire digital pathology image data, and transmit the digital pathology image data to the apparatus.

13. A method for digitally processing digital pathology image data to generate pathology scoring data and performing sensitivity analysis on the pathology scoring data, wherein: The sensitivity analysis refers to an evaluation of how the pathology score data is affected by a perturbation of a parameter used in determining or generating the pathology score data, the method comprising: a) obtaining (10) digital pathology image data, wherein the digital pathology image data comprises an image of a pathology sample; b) classifying (11) objects in the digital pathology image data into a plurality of candidate objects; c) assigning (12) a first pathological state to at least one candidate object of the plurality of candidate objects based on one or more detection thresholds and / or detection probabilities, wherein automatic morphological feature detection is used to identify the candidate object; d) obtaining (13) initial pathology score data of the pathology sample based on the candidate objects among the plurality of candidate objects to which the first pathological state has been assigned; e) perturbing (14) the one or more detection thresholds and / or the detection probabilities according to a perturbation function to generate perturbation detection thresholds and / or perturbation detection probabilities, wherein the perturbation function is configured to generate the perturbation detection thresholds and / or the perturbation detection probabilities, including by selection using a step function, by random selection using a probability distribution, by selection using a pre-provided range of values, by selection using an algebraic function, by selection using a random number generator, or by selection using a chaotic attractor function; f) reallocating (15) the first pathological state to at least one candidate object among the plurality of candidate objects or reallocating the first pathological state to at least one candidate object among the plurality of disturbance candidate objects according to one or more disturbance detection thresholds and / or disturbance detection probabilities; and g) obtaining (16) updated pathology score data of the pathology sample based on the candidate object to which the first pathological state has been reassigned according to the one or more perturbation detection thresholds and / or the perturbation detection probability, and comparing the initial pathology score data with the updated pathology score data to obtain the sensitivity of the pathology score data.

14. A computer program element comprising instructions which, when said program is executed by a computer or a processing unit, cause said computer or processing unit to carry out the method according to claim 13.

15. A computer readable medium having stored thereon a computer program element according to claim 14.

Images