TW202229862A - Automatic interpretation system of anti-cellular antibodies and automatic interpretation method thereof capable of automatically analyzing all anti-cellular antibody immunostaining images - Google Patents

Abstract

The present invention provides an automatic interpretation system of anti-cellular antibodies and an automatic interpretation method thereof, which can automatically analyze all anti-cellular antibody immunostaining images and effectively and correctly interpret cell morphology classifications corresponding to anti-cellular antibody immune images, including a competent level, expert level and mixed patterns of the International Consensus on ANA Patterns (ICAP). Accordingly, the present invention can shorten the time and labor costs of interpretation and increase consistencies of interpretation results and correctness of diagnosis.

Description

Automatic interpretation system and automatic interpretation method of anti-cell antibody

The present invention relates to an auxiliary interpretation method for autoimmune diseases, in particular to an automatic interpretation system for anti-cell antibodies and an automatic interpretation method thereof.

Press, anti-cellular antibody (ACA) indirect immunofluorescence method, also known as anti-nuclear antibody (anti-nuclear antibody, ANA) indirect immunofluorescence assay (indirect immunofluorescence assay, IIFA) is a self- The gold standard for the screening of immune diseases, which is mainly based on the treatment of serum samples, and the determination of whether the cells are fluorescently stained under a fluorescence microscope. Fluorescent staining results allow clinicians to analyze which autoantibodies and which autoimmune diseases the patient may have.

The clinical judgment and analysis of ANA morphology are mostly carried out by manual interpretation, and must rely on the interpretation of experienced medical examiners to obtain reliable results. However, the ANA indirect immunofluorescence staining method requires a large amount of testing, not every method. All medical institutions have medical examiners with experience in interpreting ANA patterns, and manual interpretation takes a lot of time, which will cause damage to the medical examiners’ eyes. In order to improve the lack of manual interpretation of ANA patterns, there are currently The automatic interpretation machine for identification, however, the automatic interpretation machine can only interpret 6 kinds of ANA patterns, that is, it cannot correctly interpret the mixed pattern, and does not conform to the ICAP (International Consensus on ANA Patterns) international classification method. Therefore, clinically It still has to be done manually.

The main purpose of the present invention is to provide an automatic interpretation system for anti-cellular antibodies and an automatic interpretation method thereof, which can automatically analyze all anti-cellular antibody immunostaining images, and effectively and correctly interpret the cells corresponding to the anti-cellular antibody immune images. Morphological classification, including the competency level (competent level), expert level (expert level) and mixed morphology of the International Consensus on ANA Patterns (ICAP) classification standard.

Another object of the present invention is to provide an automatic interpretation system for anti-cellular antibodies and an automatic interpretation method thereof, which can effectively shorten the time and labor cost of interpretation, and increase the consistency of interpretation results and the accuracy of diagnosis.

The reason is that, in order to achieve the above-mentioned object, the present invention discloses an automatic interpretation system for anti-cellular antibodies, which comprises a data module, an input module and a processing module, wherein the data module stores a plurality of sample immunizations. The cell morphology classification and classification characteristics of the stained image can be updated regularly or irregularly, and all or at least a part of the data can be used as the input source for machine learning; the input module is obtained from a sample to be tested and at least An immunostaining image to be analyzed; the processing module processes the data from the data module through the algorithm model of machine learning, and calculates a set of classification criteria, and after analyzing the characteristics of the immunostaining image to be analyzed, and then through The classification standard determines the cell classification morphology of the immunostaining image to be analyzed.

More specifically, the processing module has an operation unit, which receives the cell morphology classification and classification characteristics of the immunostaining images of the samples from the data module, and constructs a set of classification standards through operation to identify a set of classification criteria. a unit that receives the immunostaining image to be analyzed and performs identification processing to obtain an identification result, including the presence or absence, appearance, brightness, and distribution of cytoplasm, nucleus and/or metaphase cells in the immunostained image to be analyzed, and the relationship between them An analysis unit analyzes and processes the identification result according to the classification standard, and obtains an interpretation result of the cell morphology classification corresponding to the immunostaining image to be analyzed.

In addition, the analysis unit further includes a preliminary judgment part and a final judgment part; wherein, the preliminary judgment part analyzes the identification result according to the classification standard to obtain a preliminary judgment result; the final judgment part is based on a predetermined threshold or It is the majority decision algorithm that analyzes the initial judgment result and produces the judgment result.

Among them, the classification features include the presence or absence of cytoplasm, nucleus and metaphase cells, number, proportion, brightness, distribution, appearance, and their relationship with each other.

Wherein, the input module includes an image imaging unit for obtaining a predetermined number of immunostaining images to be analyzed from the sample to be tested.

Another embodiment of the present invention discloses an automatic interpreting method for anti-cellular antibodies, which provides a cell morphology classification standard, analyzes the sample to be analyzed and its parameters to be analyzed through the cell morphology classification standard, and then judges the target to be analyzed. The morphological classification of the cells corresponding to the samples was analyzed.

The sample to be analyzed is an image of immune cells stained with a predetermined antibody.

In an embodiment of the present invention, the step of interpreting the sample to be analyzed further includes:

Determine the staining results of the sample to be analyzed. If the nucleus, cytoplasm, and metaphase cells are not stained, the sample to be analyzed is judged to be Non-reactive (AC-0). If the nucleus, cytoplasm, and metaphase cells have is stained, a classification candidate result is obtained by comparing the cell morphology classification criteria, wherein, when the sample to be analyzed contains a variety of immune cell staining morphology, the cell morphology classification in the classification candidate result is classified as having a probability greater than a predetermined value. , such as 0.5.

Analyze the classification candidate result, when the classification candidate result contains only a single cell classification form, interpret the classification result of the sample to be analyzed as the classification candidate result; and when the classification candidate result includes multiple cell classification forms, analyze the classification first. Whether the classification candidate results include the cell morphology classification of AC-1, AC-2 and/or AC-4, if AC-2, and AC-1 and/or AC-4 are present, AC-1 and/or AC-1 should be excluded AC-4, include AC-2 in the initial classification result, if there is no AC-2, but there is AC-1, include AC-1 in the initial classification result of the sample to be analyzed, and according to a predetermined threshold, from other cells Screen out the cell classification patterns that exceed the predetermined threshold from the classification patterns, as the initial classification result of the sample to be analyzed. If there are no AC-2 and AC-1, according to the predetermined threshold, screen out the cell classification patterns from the cell classification patterns. The cell classification pattern that exceeds the predetermined threshold value is used as the initial classification result of the sample to be analyzed; wherein, the predetermined threshold value is the probability of the appearance of the cell classification pattern. For example, the predetermined threshold value can be set to be greater than or equal to 0.5.

Analyze the preliminary classification result, wherein, if the preliminary classification result only includes a single cell classification form, then interpret the classification result of the sample to be analyzed as the preliminary classification result; and when the preliminary classification result contains multiple cells in different images When classifying morphology, analyze whether it conforms to the mixed morphology in the classification of cell morphology, which means that the number of occurrences of cell classification morphology is greater than or equal to 2 is included in the classification result. The classification results of the samples are mixed patterns.

The present invention discloses an automatic interpretation system of anti-cellular antibodies and an automatic interpretation method thereof. By inputting a large number of anti-cellular antibody immune images, their classification characteristics and cell classification patterns into a processing module, the processing module can be processed through machine learning or Deep learning establishes a set of cell morphology classification standards, and the analysis module has to interpret the cell morphology classification corresponding to the sample to be analyzed according to the cell morphology classification standard, so as to achieve the effect of automatically interpreting the cell morphology classification of the sample to be analyzed. To achieve the effect of shortening the interpretation time, improving the efficiency and accuracy of the interpretation.

"Machine learning" disclosed in the present invention refers to an algorithm that recognizes a large number of samples, data, experience and other information through a machine, and obtains features from it, thereby training the machine to establish an operation mode or an algorithm, thereby achieving the purpose of the machine having artificial intelligence .

The "deep learning" disclosed in the present invention is a part of machine learning. However, if it is applied to data such as image types, the difference from machine learning is that deep learning omits the step of extracting features and directly extracts features through convolution. After the features are classified (meaning to establish the algorithmic rules), the most representative deep learning algorithm is the Convolutional Neural Network (CNN), which contains many different architectures. As is known to those of ordinary skill in the art to which the present invention pertains, convolutional neural networks include many different architectures, such as VGG19, ResNet50V2, DenseNet121, MobileNetV2, Xception, InceptionResNetV2.

The "indirect immunofluorescence assay (IIFA)" disclosed in the present invention is a major diagnostic tool known in the art for screening autoantibodies from human epithelial cells, and can be used as a systemic autoantibody The gold standard for systemic autoimmune rheumatic diseases (SARD), such as systemic lupus erythematosus (SLE), Sjogren's syndrome (SjS), and systemic sclerosis (SSc). The images obtained by indirect immunofluorescence measurement of HEp-2 cells (human epithelial cells) provide clinically relevant information, which can be used as the basis for subsequent antigen-specific detection.

The "International Consensus on ANA Patterns (ICAP)" disclosed in the present invention is an international classification method for ANA patterns. At present, 29 representative patterns have been identified. Among them, cell patterns are divided into three types. Major categories: nuclear morphology, cytoplasmic morphology, mitotic morphology, and each morphology has been defined and described in detail, as shown in Tables 1 to 3 below; and cell morphology is further divided into competency level (competent level) and expert level ( expert level), in terms of clinical manual interpretation, medical examiners need to identify at least the cell morphology of all competency levels.

Table 1: Morphology and characteristics of nuclear staining Classification Feature description AC-1 The nuclear homogeneous staining of interphase cells (Nuclear homogeneous), the chromatin of metaphase cells is also homogeneous staining. AC-2 The nucleoplasm of interphase cells is densely stained with small granules (some areas are dense and some areas are loose), and the chromatin of metaphase cells is also densely stained with small granules. AC-3 There are multiple (40-80/cell) discrete bright spots in the nucleoplasm of interphase cells, and multiple (40-80/cell) discrete bright spots in the chromatin of metaphase cells. AC-4 The nucleoplasm of interphase cells is finely and densely stained (some nucleoli are stained and some are not stained), and the chromatin of metaphase cells is not stained. AC-5 The large rough granules in the nucleoplasm of interphase cells are stained (some nucleoli are stained and some are not stained), and the chromatin of metaphase cells is not stained. AC-6 Multiple (6-20/cell) countable discrete bright spots in the nucleoplasm of interphase cells. AC-7 A few (1-6/cell) countable discrete bright spots in the nucleoplasm of interphase cells. AC-8 The nucleolus of interphase cells is uniformly brightly stained, and the chromatin of metaphase cells is not stained. AC-9 The nucleoli of interphase cells are irregularly brightly stained, like loose clods (Cajal bodies), and the chromatin of metaphase cells will be brighter. AC-10 The nucleoli of interphase cells are punctately bright, and some small bright spots (NOR90) can be seen in the chromatin of metaphase cells. AC-11 The nucleoplasm of interphase cells is uniformly stained, the periphery is particularly bright, and the cell junction is also particularly bright, and the chromatin of metaphase cells is not stained. AC-12 The nuclei of interphase cells were punctately brightly stained, and the cell junction was also particularly bright. The chromatin of metaphase cells was not stained. AC-13 Interphase cells were brightly stained according to different cell stages, and the brightness, size and granules of different nucleoplasmic pleomorphisms were different, and the chromatin of metaphase cells was not stained. AC-14 According to different cell stages, there are different brightness staining in the nucleoplasm, and the chromatin of metaphase cells is centromere. AC-29 Meet the following five characteristics 1. Interphase cells appear AC4 (Nucleoplasm fine speckled) and its characteristics. 2. The chromatin of mitotic cells is brightly stained (AC1 characteristic). 3. NORs (small bright spots in the chromatin of mitotic cells) can be seen inside the chromatin. 4. Cytoplasm is stained. 5. The nucleolus is stained.

Table 2: Morphology and characteristics of cytoplasmic staining Classification Feature description AC-15 Nuclear fiber linear staining (fibrillar linear). AC-16 The cytoplasmic fibers are filamentous, and the edge of the nucleus will be bright locally. AC-17 Cytoplasmic fibers were stained with short segments. AC-18 Many discrete bright spots in the cytoplasm. AC-19 The cytoplasm was cloudy and uniformly stained. AC-20 The cytoplasm is finely granular. AC-21 The cytoplasm was stained with coarse granular fibers. AC-22 At one end of the cytoplasm, around the nucleus, the band-like discontinuous dots or granules are brightly stained. AC-23 Cytoplasm rod-shaped + circular staining.

Table 3: Morphology and characteristics of dividing cell staining Classification Feature description AC-24 Bright spots on both sides of the metaphase cells or at the centrosome of the cytoplasm. AC-25 Separate bright staining of spindle fibers in metaphase cells, not necessarily nucleocytoplasmic staining (nucleocytoplasmic staining can be classified as AC26 NuMa-like pattern). AC-26 Umbrella-like separate bright stains in metaphase cells are very similar to AC24, but nucleoplasmic staining is required (speckled AC4). AC-27 Bright staining between cells near the end of metaphase. AC-28 Only the chromosomes of metaphase cells were stained, and no interphase cells were stained.

The excitation wavelengths of the automatic image capture device (NOVA View) used in the present invention to observe cell images are 400 nm (DAPI) and 490 nm (FITC). NOVA View software uses DAPI fluorescence to locate and focus HEp-2 cells, and perform image analysis based on the FITC signal; at least 3 cell images are acquired for each well in each plate, and each cell image is Must contain at least 25 interphase cells (interphase) and 2 mitotic (metaphase) cells; and the software measures the average intensity of the FITC image in light intensity units (LIU), which is used to distinguish positive samples from For negative samples, the measured LIU can be converted into an estimated endpoint titer by means of a mode-specific dilution curve. According to previous studies, cell staining images with NOVA View automatic image acquisition equipment at a 1:80 screening dilution of indirect immunofluorescence (IIF) have achieved high consistency with manual interpretation with fluorescence microscopy.

Hereinafter, in order to further illustrate the present invention, several embodiments are given and described in detail with the drawings as follows.

Referring to FIG. 1, the automatic interpretation system (10) of anti-cellular antibodies disclosed in the first embodiment of the present invention mainly includes a data module (20), an input module (30) and a processing module (40) ,in:

The data module (20) stores the cell morphology classification and classification features of the immunostaining images of the plurality of samples, wherein the classification features include the presence or absence of cytoplasm, nuclei and metaphase cells, number, proportion, brightness, distribution, appearance, and their relationship with each other, and the cell morphology classification of the immunostaining images of these samples may be the result of interpretation by automatic image capture equipment, the result of manual interpretation, or the result of combined machine and manual interpretation. .

The input module (30) includes an image imaging unit (31) for obtaining at least three immunostaining images to be analyzed from the sample to be tested; further, the input module (30) is more It includes an image adjustment unit (32) for adjusting the pixel, intensity, brightness, size, tone or other parameter values related to image display of the immunostained image to be analyzed, such as removing unnecessary colors in the immunostained image to be analyzed , so that only a single color of the immunostaining result appears in the immunostained image to be analyzed after processing.

The processing module (40) has an arithmetic unit (41), which uses the cell morphology classification and the classification characteristics of the immunostaining images of the samples in the data module (20) as input data, and a predetermined calculation model performs processing operation, outputting a classification standard, an identification unit (42), receiving each immunostaining image to be analyzed and performing identification processing, and outputting an identification result, wherein the identification result includes cytoplasm and nucleus in the immunostaining image to be analyzed and/or parameters such as presence or absence, appearance, brightness, distribution, and relationship between metaphase cells, an analysis unit (43), analyzes and processes the identification result according to the classification standard, and obtains the corresponding immunostaining image to be analyzed. Interpretation results of one of the cell morphological classifications;

Wherein, the analysis unit (43) further includes a preliminary judgment part (431), which analyzes the identification result according to the classification standard to obtain a preliminary judgment result, and a final judgment part (432), which filters out the identification result according to a predetermined threshold The preliminary judgment result, and the judgment result (also called the final judgment result) must be produced by the majority decision algorithm.

Through the composition of the above-mentioned elements, the automatic interpretation system (10) of the anti-cellular antibody disclosed in the first embodiment of the present invention uses the data module to store a large number of anti-cellular antibody immune images and cell morphology classification results, characteristics and other data, and , the data of the data module (20) can be updated in real time, and then the data in the data module (20) is used as input data, so that the operation unit (41) can pass a predetermined machine learning algorithm or model A classification standard related to ANA cell morphology is constructed, and the classification standard includes the corresponding characteristics of each cell morphology classification, the steps and processes of interpreting the image to be analyzed, etc.; therefore, when the input unit (30) obtains these After the immunostaining images to be analyzed, the identification unit (42) will first perform identification processing, and then the analyzing unit (43) will analyze the identification results according to the classification criteria to generate interpretation results of the immunostained images to be analyzed, It can also correspond to a predetermined cell morphology classification, which can be provided to clinical medical treatment as a follow-up treatment method or a basis for disease judgment, so as to shorten the cost of interpreting ANA IIFA images, and achieve automatic assistance in judging autoimmune diseases. effect.

Referring to FIG. 2, a second embodiment of the present invention discloses an automatic interpretation method for anti-cellular antibodies, which includes the following steps:

Step 101: Input a sample to be analyzed.

Wherein, the sample to be analyzed is at least one immune cell staining image.

Step 102: Process the sample to be analyzed to obtain a parameter to be analyzed.

Wherein, the parameters to be analyzed include staining results, appearance, brightness, distribution of cytoplasm, nuclei and/or metaphase cells, and their correlation with each other.

Step 103: Provide a cell morphology classification standard.

Wherein, the cell morphology classification standard is obtained from the cell morphology classification and its classification characteristics of a plurality of known immunostaining images through operation processing, and these known immunostaining images are obtained by performing image amplification before operation processing. Processing or/and image adjustment processing to improve the accuracy of the cell morphology classification criteria.

Step 104 : according to the cell shape classification standard, analyze and process the parameters to be analyzed to obtain the cell shape classification corresponding to the sample to be analyzed.

Further, as shown in FIG. 3 , step 104 further includes the following steps:

Step 104-1: Sequentially determine the staining status of the nucleus, cytoplasm and metaphase cells in the sample to be analyzed.

If the nucleus, cytoplasm and metaphase cells are not stained, the sample to be analyzed is judged to be non-reactive, and the interpretation result of AC-0 is output;

If at least any one of the nucleus, cytoplasm and metaphase cells is stained, then according to the cell morphology classification standard, the parameters to be analyzed are further analyzed based on the staining results displayed in different parts, and a classification candidate result is obtained, wherein the The classification candidate results will include at least one cell morphology classification.

For example, if the nucleus is not stained, but the cytoplasm is stained, it is necessary to analyze the pattern characteristics of the cytoplasm, and then analyze whether there are special characteristics of metaphase cells, such as: AC-24 has the characteristics of bright centrosome staining, AC -25 has the characteristics of umbrella-like separate bright dyes in metaphase cells;

If the nucleus is stained, but the cytoplasm is not stained, you need to find the metaphase cells first, analyze the pattern of the nucleus, and then analyze whether there are special features in the metaphase cells and whether the cells overlap or whether there are special features at the junction, such as AC-11 It has the characteristics of bright staining at the periphery of smooth cells and particularly bright at the cell junction, AC-12 has the characteristics of punctate bright staining at the periphery of the nucleus and particularly bright at the cell junction, and AC-27 has the characteristics of intercellular bridge;

If both the nucleus and cytoplasm are stained, first identify the metaphase cells, and then analyze the pattern of the nucleus and cytoplasm, whether there are special features in the metaphase cells, and whether there are special features in the overlapping cells or at the junction.

Step 104-2: Analyze the classification candidate result.

When the classification candidate result contains only a single cell classification form, interpret the classification result of the sample to be analyzed as the classification candidate result;

When the classification candidate result contains multiple cell classification forms, first analyze whether the classification candidate result contains the cell shape classification of AC-2. If the classification candidate result contains AC-2, interpret the classification result of the sample to be analyzed. It is AC-2. If the candidate classification result does not contain AC-2, then analyze whether the candidate classification result of the sample to be analyzed contains the cell classification morphology of AC-1. If the candidate classification result contains AC-1 , including AC-1 in the initial classification result of the sample to be analyzed, and according to a predetermined threshold, screen out the cell classification forms exceeding the predetermined threshold from other cell classification forms, as the initial classification result of the sample to be analyzed, On the other hand, if the classification candidate result does not contain AC-1, according to the predetermined threshold, the cell classification forms exceeding the predetermined threshold are selected from the cell classification forms as the initial classification result of the sample to be analyzed;

The predetermined threshold is a numerical value, which can represent the number of occurrences or the probability of occurrence, that is, the cell classification forms whose repetition rate meets the predetermined threshold are screened out through the majority decision model in statistical analysis, and are included in the initial classification As a result, for example, the predetermined threshold value is 0.5, which means that the cell morphology in the classification candidate result can only be included in the preliminary classification result if the probability of occurrence is greater than or equal to 0.5.

Step 104-3: Analyze the preliminary classification result.

If the preliminary classification result only includes single cell classification form, the classification result of the sample to be analyzed is the preliminary classification result, that is, the final interpretation part adopts the result of the preliminary interpretation part, and the classification result is the single cell classification form;

If the initial classification result contains multiple cell classification forms in different images, only those whose occurrence times are greater than or equal to a predetermined value will be included in the classification result, which means that the final interpretation unit will screen out the corresponding cell classification according to the set value. If there are more than two types of cell classification forms, the classification result is a mixed form, for example, the number of occurrences can be set to be greater than or equal to 2.

The automatic interpretation method of the anti-cell antibody disclosed in the present invention can directly determine the corresponding cell classification form directly from the content of the immunostaining image, and overcome the inability to accurately determine the mixed morphology or the expert level in the prior art. The lack of cell classification and morphology, and the classification results provided by the automatic interpretation method of anti-cell antibodies can provide clinical reference for further antibody testing or disease treatment, without relying on a lot of manpower for image interpretation.

In addition, it must be emphasized that through the automatic interpretation system and automatic interpretation method of anti-cellular antibodies disclosed in the present invention, the operator only needs to provide the ANA IIFA image of the sample to be analyzed, rather than the antibody related to a specific autoimmune disease. The stained image is used as the input source for interpretation. In other words, the automatic interpretation system and automatic interpretation method of anti-cellular antibodies disclosed in the present invention greatly shorten the detection time and detection cost, and further improve the accuracy of automatic detection.

To further illustrate the technical features and effects of the present invention, the following examples and diagrams will be used for detailed description as follows.

Example 1: Collection of samples and their ANA IIFA images

90,109 images with ANA IIFA (each sample will produce 3-8 images) were collected from the NOVA View image automatic image capture equipment of Taichung Veterans General Hospital, Taiwan, and if the sample and its images conform to Any of the following conditions will be excluded: lack of formal reports, periodic pre-test samples, severely abnormally exposed images, with 34,982 remaining samples. As shown in Table 4 below, 18,380 IIFA images were judged to be negative (AC-0), and negative was defined as negative immunofluorescence staining for the nucleus, cytoplasm, and metaphase cells; the remaining 16,602 ANA IIFA images (from 10641 patients) were judged to be non-AC-0 and corresponded to more than one cell morphology classification; and subsequent related antibody tests were performed; among them, MSA in Table 4 is the abbreviation of myositis specific antibody panel.

Table 4: ANA pattern classification distribution ICAP Competency Level Classification Number of patients (%) Number of samples(%) Quantity for Machine Learning Relevant tests provided by the laboratory Clinical judgment Number of patients receiving follow-up tests/number of positives (%) train verify test Nuclear Homogenization (AC-1) 6293 (24.3) 8717 (24.9) 21596 5398 3106 anti-DsDNA antibody 2166/472 (21.8) anti-histone antibody 115/30 (26.1) Nuclei dense with fine spots (AC-2) 247 (1) 401 (1.1) 981 244 95 none none Centromere (AC-3) 545 (2.1) 910 (2.6) 2239 561 307 anti-centromeric antibody 304/293 (96.4) Nuclear Speck (AC-4/AC-5) 10632 (41) 15078 (43.1) 37577 9396 5813 Anti-SSA/SSB antibody 5403/1547 (28.6) Autoantibodies against Mi2, TIF NXP2, SAE1, Ku, Ro52 (MSA panel) 1039/442 (42.5) Anti-Sm/RNP antibody 245/144 (58.8) Discrete core points (AC-6/AC-7) 446 (1.7) 755 (2.2) 1882 470 213 Anti-NXP2 antibody 25/2 (8) Nucleolus (AC-8/AC-9/AC-10) 782 (3) 1022 (2.9) 2524 630 326 Anti-PM-Scl 75/100 (MSA panel) or Scl-70 antibody 187/9 (26.2) Cytoplasmic fibers (AC-15/AC-16/AC-17) 107 (0.4) 129 (0.4) 410 103 28 none none Cytoplasmic spots (AC-18/AC-19/AC-20) 2005 (7.7) 2534 (7.2) 6380 1595 1122 Anti-PL-7, PL-12, Jo-1, SRP, EJ, OJ, MDA5, Ro52 antibodies (MSA panel) 153/110 (71.9) Anti-Jo-1 antibody 90/13 (14.4) anti-ribosome-P antibody 26/10 (38.5) Cytoplasmic reticulum/AMA (AC-21) 264 (1) 380 (1.1) 980 245 105 anti-mitochondrial antibody 89/53 (59.6) Anti-mitochondrial-M2 antibody 10/9 (90) Pola/Golgi-like (AC-22) 37 (0.1) 55 (0.2) 463 116 15 none none Rod and Ring (AC-23) 44 (0.2) 54 (0.2) 635 159 18 none none AC-0 11373 (44.1) 16376 (47.1) Not AC-0 15299 (59.3) 18380 (52.9)

Example 2: Image Processing (1)

Due to the uneven number of morphological classifications of cells in Example 1, the results of machine learning may be poor. Then, some patients' sera with positive manifestations such as cytoplasmic fibers, cytoplasmic reticulum/AMA or cytoplasmic polar/Golgi-like patterns were collected. The ANA test was further carried out to obtain 121 cytoplasmic fiber-positive images, 74 cytoplasmic reticulum/AMA-positive images, and 406 cytoplasmic Pola/Golgi-like images, and through the brightness increase and decrease technology To amplify the patterns of rod-shaped and ring-shaped cells, 624 enhanced images were obtained; the above-mentioned patterns were still input to the input end as training data for machine learning, as shown in Table 4 above.

Example 3: Image Processing (2)

Since all the ANA IIFA image information of HEp-2 cells has only the green channel, the blue and red channels in the ANA IIFA images need to be removed to avoid noise. Perform image preprocessing through pixel adjustment method to appropriately enhance the intensity contrast of the image, so that the intensity value can be normalized to a ratio between 0 and 1, wherein the intensity adjustment method is as follows:

Input image increase value = (input image - input image minimum intensity value) / (input image maximum intensity value - input image minimum intensity value)

Example 4: Training rules for machine learning

According to the classification rules of ICAP, the morphological classification of ANA cells must take into account metaphase cells and interphase cells at the same time. Therefore, the input image is based on the image level of HEp-2 cell images (specimen-level or image-level) rather than cell-level based on HEp-2 cell images.

For this example, all patterns are trained and validated according to the 80:20 ratio and resized to 299 X 299 pixels. 30 epochs (epoch) of training were performed under the GPU machine learning environment (NVIDIA TITAN V, 12GB RAM), with a minimum batch size of 30.

Example 5: Interpretation rules of machine learning

Since the ANA test will have at least 3 immunofluorescence images for each sample, all the interpretation results need to be combined and the final interpretation result will be determined by a majority vote. Therefore, for each HEp-2 cell image, use The pre-trained model produces probability distributions on 11 possible cell shape classifications, and selects cell shape classifications with a probability greater than or equal to 0.5 as the initial interpretation image prediction result; in each sample image, at least two The cell morphology distribution supported by the image prediction results obtained from the HEp-2 cell image is used as the final interpretation result; however, the exceptions to the initial interpretation are as follows:

1. The main key to judging whether there is uniform nuclear and cytoplasmic staining (AC-1) lies in metaphase cells, but typical metaphase cells may only appear in one of all the images of each sample, so only one of them If the image prediction result contains nucleocytoplasmic uniform staining positive (ie AC-1), then the final interpretation will include nucleocytoplasmic uniform staining positive (AC-1) in the final interpretation result;

2. In theory, AC-2 will not coexist with AC-1 and/or AC-4, so when the image prediction result is AC-2, AC-1 and AC-4 are automatically ignored.

For example, please refer to FIG. 8. In this example, the threshold value of the initial judgment part is set to be greater than or equal to 0.5, and the threshold value of the final judgment part is set to be 2 or more; According to the set threshold, the classification candidate results of each immunofluorescence image are screened, and at the same time, it is analyzed whether there are the above-mentioned exceptions to the initial interpretation; since there are no exceptions to the initial interpretation in this case, the initial interpretation department leaves AC-4 (2). times), AC-19 (2 times), and AC-6 (1 time), the initial classification results, and then the final interpretation department filters out the classification results according to the set threshold, which is a mixture of AC-4 and AC-19 form.

Please refer to FIG. 9 again. In this example, the threshold settings of the initial interpretation part and the final interpretation part are the same as those in the previous example. However, since the classification candidate results of image 1 and image 2 contain the cell classification form of AC-2, this is This is an exception to the initial interpretation. Therefore, AC-2 is left in the initial classification results of Image 1 and Image 2, and AC-1 and AC-4 are left in the initial classification results because AC-2 does not appear in Image 3. ; Based on the number of occurrences of AC-2 in the initial classification result is equal to 2, so the final interpretation part only leaves AC-2 as the classification result.

Please refer to FIG. 10 again. In this example, the threshold settings of the initial judgment part and the final judgment part are the same as the above example; and although AC-4 appears 3 times and AC-1 appears 1 time in the initial classification result, based on the aforementioned As long as AC-1 appears in the initial classification result, AC-1 should be included in the classification result, regardless of the number of occurrences of AC-1, so the classification result in this example is AC-1. And the hybrid form of AC-4.

Example 6: Evaluation Method of Classification Interpretation Results

Through precision (the proportion of positive predictions that are correct), sensitivity (recall, the proportion of correctly classified positives), and F1 score (the weighted average performance of precision and recall), they are calculated as follows:

Precision = true positives / (true positives + false positives)

Sensitivity = true positives / (true positives + false negatives)

F1 Score = (2 X Accuracy X Sensitivity) / (Accuracy + Sensitivity)

In addition, the Kappa value (Cohen's kappa) can also be used as an indicator to evaluate whether the classification results are consistent, among which:

Kappa value = (P ₀ -P _C )/(1-P _C )

P ₀ is the observed agreement, the percentage of two measurement results that are consistent;

PC is the chance agreement, the percentage that two _measurements are expected to be the same.

Example 7: Comparing the Calculation Results of Different Machine Learning Models

The effects of different pre-training models are evaluated by test data 1, wherein the test data 1 (as shown in Figure 3) contains 1985 samples and 6195 images generated by them. The results are shown in Table 5.

From the results in Table 5, it can be seen that the highest F1 score (0.86) and the highest kappa value (0.82) can be obtained with the InceptionResNetV2 model. Therefore, the InceptionResNetV2 model will be used as the machine learning algorithm through the transfer learning technique. , the calculus architecture is shown in Figure 4.

The test data 1 is calculated by the InceptionResNetV2 model, and the accuracy, sensitivity, F1 score and Kappa value of different cell classification forms are shown in Table 6. Among them, the Kappa value shows that the prediction results of the InceptionResNetV2 model are different from those of experienced medical examiners. Consistency between interpretation results.

The results in Table 6 show that the accuracy is between 0.73 and 1.00, with an average of 0.93; the sensitivity is between 0.64 and 1.00, with an average of 0.81; the F1 score is between 0.72 and 1.00, with an average of 0.81. 0.86. The image features of the three cell types with the highest F1 scores are rods and rings, nuclear speckles and centromeres, while the image features of the four cell types with the lowest F1 scores are cytoplasmic speckles, nucleoli, and discrete nuclear dots. , cytoplasmic fibers. Furthermore, the Kappa value of each cell morphological classification ranges from 0.67 to 1.00, with an average of 0.82, indicating that the results of interpretation based on the classification criteria obtained by machine learning and the results of manual interpretation have achieved almost perfect consistency.

In other words, from the above results, it can be confirmed that the automatic interpretation system and automatic interpretation method of anti-cellular antibodies disclosed in the present invention can accurately interpret the cell morphology classification of ANA IIFA images, so as to reduce the cost of manual interpretation and improve the interpretation efficiency.

Table 5: F1 scores and Kappa values after calculation with different pretrained models pretrained model F1 Score Kappa value InceptionResNetV2 0.86 0.82 MobileNetV2 0.81 0.76 Xception 0.78 0.74 VGG19 0.78 0.73 ResNet50V2 0.73 0.68 DenseNet121 0.68 0.63

Table 6: Calculation results of each cell morphology classification in test data 1 using InceptionResNetV2 Accuracy Sensitivity F1 Score Kappa value nuclear homogeneity 0.92 0.87 0.89 0.79 Nuclei are dense with fine spots 0.88 0.79 0.83 0.83 Centromere 1.00 0.88 0.93 0.93 nuclear speck 0.98 0.99 0.99 0.74 discrete core points 0.94 0.70 0.80 0.79 Nucleolus 1.00 0.64 0.78 0.77 cytoplasmic fibers 0.73 0.89 0.80 0.80 cytoplasmic spots 0.82 0.65 0.72 0.67 Cytoplasmic reticulum/AMA 0.92 0.73 0.81 0.81 Polar/Golgi-like 1.00 0.80 0.89 0.89 rod and ring 1.00 1.00 1.00 1.00

Example 8: Conformance Test

544 images generated from 175 samples were randomly selected as test data 2, and the results of manual interpretation by 5 experienced medical examiners, the results of manual interpretation by 2 junior medical examiners, and the results of interpretation by the InceptionResNetV2 model were respectively used. The results were compared in pairs, and the consistency between them was analyzed. The results are shown in Figures 5 to 7.

The results from Figure 5 to Figure 7 show that the average consistency of the results obtained by 5 experienced medical examiners in the interpretation of 11 cell morphological classifications is 0.849, representing almost perfect consistency; The average agreement of the interpretation results obtained by the examiner and 2 junior medical examiners was 0.528, representing moderate agreement; and the automatic interpretation method of anti-cellular antibodies disclosed in the present invention was carried out through machine learning (such as the InceptionResNetV2 algorithm). The consistency between the interpretation results and the interpretation results obtained by 5 experienced medical examiners was 0.844, which represented an almost perfect consistency with each other.

The above results confirm that the accuracy of the automatic interpretation system and the automatic interpretation method of the anti-cellular antibody disclosed in the present invention in interpreting ANA IIFA images is significantly higher than that of the junior staff, which means that the automatic interpretation system of the anti-cellular antibody disclosed in the present invention and Its automatic interpretation method for ANA IIFA image interpretation can not only reduce the error rate caused by human factors in manual interpretation, but also can determine a variety of cell morphological classifications, including mixed morphological and expert-level cell morphological classification, which not only saves manpower and time Cost, more accurate and rapid interpretation of ANA IIFA images, in order to achieve the effect of assisting clinical judgment and treatment of autoimmune diseases.

(10): Automatic interpretation system of anti-cell antibodies (20):Data module (30): Input module (31): Image imaging unit (32): Image adjustment unit (40):processing module (41): operation unit (42): identification unit (43): Analysis Unit (431): Preliminary Interpretation Department (432): Final Interpretation Department (101): Input a sample to be analyzed (102): process the sample to be analyzed to obtain a parameter to be analyzed (103): provide a cell morphology classification standard (104): according to the cell morphology classification standard, analyze and process the parameters to be analyzed, and obtain the cell morphology classification corresponding to the sample to be analyzed (104-1): Sequentially determine the staining status of the nucleus, cytoplasm and metaphase cells in the sample to be analyzed (104-2): Analyze the classification candidate results (104-3): Analyze the preliminary classification results

FIG. 1 is a schematic diagram of a first embodiment of the present invention. FIG. 2 is a flowchart of a second embodiment of the present invention. FIG. 3 is an architecture diagram of the machine learning model described in the example of the present invention. FIG. 4 is a flow chart of the interpretation described in the example of the present invention. Figure 5 is an analysis of the consistency between the interpretation results obtained by different experienced medical examiners, in which, A1 to A5 represent 5 different experienced medical examiners respectively, and the numbers in the table represent the differences among these medical examiners. The results of the interpretation consistency of various morphologies, and the Kappa value less than 0 means poor consistency, 0-0.2 means very low consistency, 0.21-0.40 means little consistency, 0.41-0.60 means consistency, 0.61-0.80 means consistency Strong, 0.81-1 represents near-perfect consistency. Figure 6 is an analysis of the consistency between the interpretation results obtained by experienced medical examiners and junior medical examiners. Among them, A1 to A5 represent 5 different experienced medical examiners, and F1 to F2 represent 2 junior medical examiners respectively. Examiners, the numbers in the table represent the results of the consistency of interpretation of various forms between different experienced medical examiners and junior medical examiners. Figure 7 is an analysis of the consistency between the interpretation results obtained by experienced medical examiners and the automatic interpretation system. Among them, A1 to A5 represent 5 different experienced medical examiners, AI represents the automatic interpretation system, and the numbers in the table represent Interpretation consistency results for various modalities between different experienced medical examiners and automated interpretation systems. FIG. 8 is a schematic flowchart (1) of a machine learning interpretation rule disclosed in the present invention. FIG. 9 is a schematic flowchart (2) of a machine learning interpretation rule disclosed in the present invention. FIG. 10 is a schematic flowchart (3) of the machine learning interpretation rules disclosed in the present invention.

Claims (10)

An automatic interpretation system for anti-cell antibodies, comprising: 1. A data module, which stores the cell morphology classification and classification features of immunostaining images of multiple samples, wherein the classification features include the presence or absence, number, proportion, brightness, distribution, and appearance of cytoplasm, nuclei, and metaphase cells; and their relationship with each other; an input module for obtaining at least one immunostaining image to be analyzed from a sample to be tested; A processing module has an operation unit, which receives the cell morphology classification and classification characteristics of the immunostained images of the samples from the data module, and constructs a set of classification standards through operation, and an identification unit receives the to-be-to-be-stained images. Analyze the immunostaining image and perform identification processing to obtain an identification result, including the presence or absence, appearance, brightness, distribution of cytoplasm, nucleus and/or metaphase cells in the immunostained image to be analyzed, and their correlation with each other and other parameters , an analysis unit that analyzes and processes the identification result according to the classification standard, and obtains an interpretation result of the cell morphology classification corresponding to the immunostaining image to be analyzed. The automatic interpretation system for anti-cellular antibodies according to claim 1, wherein the input module includes an image imaging unit for obtaining a predetermined number of immunostaining images to be analyzed from the sample to be tested. The automatic interpretation system for anti-cellular antibodies according to claim 2, wherein the analysis unit further comprises a preliminary interpretation part and a final interpretation part; The preliminary judgment section analyzes the identification result according to the classification standard to obtain a preliminary judgment result; the final judgment section analyzes the preliminary judgment result according to a predetermined threshold or a majority decision algorithm, and outputs the judgment result. An automatic interpretation method for anti-cell antibodies, which comprises the following steps: inputting a sample to be analyzed to obtain at least one image of the sample to be analyzed; processing the image to obtain a parameter to be analyzed, wherein the parameter to be analyzed includes the staining result, appearance, brightness, distribution, and correlation between cytoplasm, nucleus and/or metaphase cells; Provide a cell morphological classification standard, which is obtained by calculating the cell morphological classification and classification characteristics of a plurality of known immunostaining images; According to the cell shape classification standard, the parameters to be analyzed obtained from the image of the sample to be analyzed are analyzed and processed to obtain the cell shape classification corresponding to the sample to be analyzed. The automatic interpretation method for anti-cell antibodies according to claim 4, wherein the image is an image of immune cells stained with a predetermined antibody. The method for automatic interpretation of anti-cell antibodies according to claim 4, wherein the known immune cell staining images are subjected to image amplification processing before the arithmetic processing. The automatic interpretation method for anti-cellular antibodies according to claim 4, wherein the number of the images is at least 3. The automatic interpretation method of the anti-cellular antibody according to claim 7, wherein, in step d, the following steps are further included: Step d1: Determine the presence or absence of nucleus, cytoplasm and metaphase cells in the parameters to be analyzed according to the staining results of each image, if not, then determine that the sample to be analyzed is non-reactive (Non-reactive, AC-0) , if yes, compare the cell morphology classification criteria to obtain a classification candidate result; Step d2: analyze the classification candidate results, wherein: When the classification candidate result contains only a single cell classification form, interpret the classification result of the sample to be analyzed as the classification candidate result; When the classification candidate result is that at least one image respectively contains a plurality of cell classification forms, first analyze whether the classification candidate results of the images include the cell shape classification of AC1, AC-2 and/or AC-4. If there is AC-2, and AC-1 and/or AC-4, then AC-1 and /AC-4 need to be excluded, and AC-2 should be included in the preliminary classification result. If there is no AC-2, but there is AC-1, AC-1 is included in the initial classification result of the sample to be analyzed, and according to a predetermined threshold, the cell classification forms exceeding the predetermined threshold are selected from other cell classification forms as the initial classification result of the sample to be analyzed. If Without AC-2 and AC-1, according to the predetermined threshold, screen out the cell classification forms that meet the predetermined threshold from the cell classification forms, as the initial classification result of the sample to be analyzed; Step d3: analyze the preliminary classification result, wherein, if the preliminary classification result only includes a single cell classification form, interpret the classification result of the sample to be analyzed as the preliminary classification result; and when the preliminary classification result includes a plurality of cell classification forms , it is necessary to analyze whether it conforms to the mixed morphology in the cell morphology classification according to a predetermined condition, wherein the predetermined condition includes that if the initial classification result contains AC-1, AC-1 should be included in the classification result, and other initial classification results. Cell classification patterns should appear a predetermined number of times before they can be included in the classification results. The method for automatically interpreting anti-cell antibodies according to claim 8, wherein the predetermined threshold is the occurrence probability of cell classification morphology. The method for automatic interpretation of cellular antibodies according to claim 8, wherein the predetermined threshold is set to be greater than or equal to 0.5, and the predetermined number of times is set to be greater than or equal to 2.

Images