Method for automatically detecting abnormal cells in urine exfoliated cells

Abstract

The invention provides a method for automatically detecting abnormal cells in urine cast-off cells, which comprises the following steps of: training an abnormal cell frame selection recognition algorithm, and generating an abnormal cell frame selection recognition model; according to the abnormal cell frame selection recognition model, abnormal cell frame selection recognition of a to-be-tested sample is achieved. The method is suitable for early preliminary diagnosis screening and postoperative reexamination of bladder urothelium carcinoma, has the advantages of early noninvasive detection, convenience in analysis and report, simplicity and standardization in operation, suitability for automatic analysis of instruments and the like, and not only can eliminate non-cell impurities in a complex background, but also can effectively separate overlapped cells, so that the detection accuracy is improved. And a foundation is laid for subsequent single cell analysis.

Description

Technical Field

[0001] The present invention relates to the field of medical testing technology, and in particular to a method for automatically detecting abnormal cells in urine exfoliated cells. Background Art

[0002] Bladder cancer is a common malignancy of the urinary system, accounting for over 90% of bladder cancers. The five-year recurrence rate after surgery is as high as 30%-80%. Therefore, early screening and detection of recurrence of bladder urothelial cancer are particularly important. Currently, the main methods for initial screening and postoperative follow-up of bladder urothelial cancer include cystoscopic biopsy, urine cytology, and fluorescence in situ hybridization (FISH) of urine exfoliated cells.

[0003] Cystoscopic biopsy is the gold standard for diagnosing bladder urothelial carcinoma with good specificity. However, biopsy is invasive and complicated to perform, requiring patients to endure great pain. It is also difficult to diagnose early because the tumor must reach a certain volume for detection. It has poor sensitivity for bladder carcinoma in situ, limiting its use in early diagnosis and postoperative testing.

[0004] Urine exfoliative cytology has the advantages of being non-invasive, simple, repeatable, and highly specific. It is a widely used non-invasive test. However, it also has the disadvantages of poor sensitivity, inability to diagnose low-grade cancers, lack of attention from hospitals, and lack of professional cell morphology technicians to provide reliable test results. At the same time, manual microscopic examination of urine exfoliative cytology is highly subjective and relies on the personal experience of doctors. The diagnostic criteria and result reporting format rely on the personal experience of testing technicians and are difficult to quantify and unify.

[0005] FISH testing of urine exfoliated cells has high sensitivity and comparable specificity to urine exfoliated cells, but the test is expensive, technically complex, and requires microscopic examination for diagnosis. Given this, a new method for detecting abnormal urine exfoliated cells with high sensitivity, high standardization, low dependency, and the ability to automatically process them in batches is urgently needed to address the shortcomings of existing technologies. Summary of the Invention

[0006] In response to the technical issues mentioned in the background art, a method for automatically detecting abnormal cells in urine exfoliated cells is provided. Through standardized image acquisition, intelligent cell recognition, and abnormality screening, this invention improves the sensitivity, efficiency, and standardization of urine exfoliated cytology testing, making it suitable for the early diagnosis and postoperative follow-up of urinary system tumors such as bladder cancer.

[0007] The technical means adopted in the present invention are as follows:

[0008] A method for automatically detecting abnormal cells in urine exfoliated cells comprises the following steps: training an abnormal cell selection and recognition algorithm to generate an abnormal cell selection and recognition model; and implementing abnormal cell selection and recognition of a sample to be tested based on the abnormal cell selection and recognition model.

[0009] Furthermore, the training of the abnormal cell selection and recognition algorithm to generate the abnormal cell selection and recognition model includes the following steps:

[0010] SA1. Collect urine samples from patients, centrifuge them, and use a liquid-based thin-layer cytology slide preparation and staining machine to form urine cytology slides. Stain the slides using the Papanicolaou (Pap) staining method.

[0011] SA2, photographing the slide to obtain 40× magnified microscopic images of all specimen positions on the slide, and then stitching them together to generate a complete digital urine cytology slide image;

[0012] SA3. Medical experts perform instance segmentation and annotation of cells in the digital image, clearly labeling three types of target cells and non-target objects, and generating a labeling model based on the three types of target cells and non-target objects;

[0013] SA4: Train the tagging model generated in SA3 to ensure that the positioning recall rate is no less than 98.86% and the precision rate is no less than 95.33%;

[0014] SA5: Automatically crop the cell region identified by the segmentation model in step SA4 into a cell slice of fixed size and input it into a coarse classification convolutional neural network model; the coarse classification network is used to eliminate non-target objects and retain only target cells;

[0015] SA6: Input the target cell image selected by coarse classification in SA5 into the trained semantic segmentation model to further perform pixel-level segmentation on the internal structure of the cell.

[0016] SA7. After obtaining the segmentation results, structural and color statistical features are extracted. These features include: the ratio of the area of ​​the cell nucleus to the cytoplasm, i.e., the nuclear-cytoplasmic ratio; the color mean and standard deviation of the cell nucleus region under the RGB channel; and the roundness, aspect ratio, and perimeter of the cell nucleus outline, thereby reflecting the morphological characteristics of the cell. The statistical results are then input into the rule engine, and the specific type of each cell is determined through a comprehensive analysis of abnormal patterns. The classification labels include normal cells, atypical cells, atypical urothelial cells, and degenerative cells.

[0017] Furthermore, in the SA2, a full-slide digital slide scanning system with a 40× objective lens is used to photograph the glass slide.

[0018] Furthermore, the target cells include normal urothelial cells, atypical urothelial cells and atypia urothelial cells; the non-target objects include degenerative cells, squamous cells, impurities, white blood cells and red blood cells.

[0019] Furthermore, the abnormal cell selection and identification of the sample to be tested according to the abnormal cell selection and identification model includes the following steps:

[0020] SB1. Collect urine samples from patients, centrifuge them, and use a liquid-based thin-layer cytology slide preparation and staining machine to form urine cytology slides. Stain the slides using the Papanicolaou (Pap) staining method.

[0021] SB2, photographing the slide using a full-slide digital slide scanning system with a 40× objective lens, obtaining 40× magnified microscopic images of all specimen positions on the slide, and then stitching them together to generate a complete digital urine cytology slide image;

[0022] SB3. Input the digital image into the trained cell localization model. The model automatically analyzes the cell boundaries and morphological features in the image, identifies and selects all suspected target cell regions, and crops the original image based on the selection results to generate a fixed-size cell slice, which serves as input data for the subsequent recognition model.

[0023] SB4: The generated cell slices are fed into the trained cell recognition model. First, the coarse classification network removes non-target cell types, retaining the urothelial cell images. The fine classification network then further identifies the selected target cell images and analyzes whether there are abnormal morphological features.

[0024] For images identified as target cells, the semantic segmentation model is used to perform fine segmentation of their internal structures, extract the mask images of the cell nucleus and cytoplasm regions, and calculate the key morphological and color feature parameters based on the mask images;

[0025] SB5: Determine whether abnormal cells exist in the urine cytology slide image based on the nuclear-cytoplasmic ratio, RGB pixel mean, standard deviation, nuclear circularity, and aspect ratio according to the preset abnormality judgment rules;

[0026] SB6. The system automatically highlights the locations of all cells identified as abnormal on the original digital slice image, and outputs the corresponding image position coordinates and identification results to facilitate subsequent doctor review and diagnostic assistance.

[0027] Furthermore, the fixed size is 256 pixels×256 pixels.

[0028] Furthermore, in step SB4, each target cell region identified by the positioning model is cropped from the original digital slice image into an independent cell slice image; for each cell slice image, a semantic segmentation model based on a convolutional neural network is used to finely segment its internal structure; low-level and high-level image features are extracted through multi-layer convolution, and a semantic mask image is generated through pooling and upsampling operations to distinguish between the cell nucleus and cytoplasm; and the nuclear-cytoplasm ratio can be obtained by calculating the pixel area of ​​the mask area.

[0029] Compared with the prior art, the present invention has the following advantages:

[0030] The method of the present invention is suitable for early initial diagnosis screening and postoperative review of bladder urothelial carcinoma. It has the advantages of early non-invasive detection, convenient analysis and reporting, simple and standardized operation, and suitability for automatic instrument analysis. The algorithm can not only exclude non-cellular impurities in complex backgrounds, but also effectively separate overlapping cells, laying the foundation for subsequent single-cell analysis. As one of the important cell biological parameters, the nuclear-cytoplasmic ratio can effectively reflect abnormal changes in cells, such as increased nuclear volume or decreased cytoplasm, and is a key indicator for the initial screening of abnormal cells. The quantification of judgment indicators reduces dependence on the doctor's personal experience and reduces the interference of human subjective factors, and has broad prospects for clinical application. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.

[0032] Figure 1 Schematic diagram of the workflow of the present invention. DETAILED DESCRIPTION

[0033] In order to enable those skilled in the art to better understand the solutions of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts should fall within the scope of protection of the present invention.

[0034] It should be noted that the terms "first", "second", etc. in the description and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that the numbers used in this way can be interchanged where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0035] like Figure 1 As shown, the present invention provides a method for automatically detecting abnormal cells in urine exfoliated cells, comprising the following steps: training an abnormal cell selection and recognition algorithm to generate an abnormal cell selection and recognition model; and implementing abnormal cell selection and recognition of a sample to be tested based on the abnormal cell selection and recognition model.

[0036] As a preferred embodiment, in this application, training an abnormal cell selection and recognition algorithm and generating an abnormal cell selection and recognition model include the following steps:

[0037] SA1. Collect urine samples from patients, centrifuge them, and use a liquid-based thin-layer cytology slide preparation and staining machine to form urine cytology slides. Stain the slides using the Papanicolaou (Pap) staining method.

[0038] SA2, photographing the slide using a full-slide digital slide scanning system with a 40× objective lens, obtaining 40× magnified microscopic images of all specimen positions on the slide, and then stitching them together to generate a complete digital urine cytology slide image;

[0039] SA3. Medical experts perform instance segmentation and annotation of cells in digital images, clearly marking three types of target cells and non-target objects, and generate a labeling model based on the three types of target cells and non-target objects; the target cells include: normal urothelial cells, atypical urothelial cells, and atypical urothelial cells; the non-target objects include: degenerate cells, squamous cells, impurities, white blood cells, and red blood cells.

[0040] SA4: Train the tagging model generated in SA3 to ensure that the positioning recall rate is no less than 98.86% and the precision rate is no less than 95.33%;

[0041] SA5: Automatically crop the cell region identified by the segmentation model in step SA4 into a cell slice of fixed size and input it into a coarse classification convolutional neural network model; the coarse classification network is used to eliminate non-target objects and retain only target cells;

[0042] SA6: Input the target cell image selected by coarse classification in SA5 into the trained semantic segmentation model to further perform pixel-level segmentation on the internal structure of the cell. The semantic segmentation model adopts a deep convolutional neural network architecture and generates corresponding semantic masks (Masks) for the cell nucleus and cytoplasm regions in the image through a multi-scale feature extraction mechanism, thereby realizing automatic distinction and positioning of the nuclear and cytoplasmic regions. This stage adopts a pixel-level supervised learning method, using the nuclear and cytoplasmic partition labels annotated by experts as training data, and optimizes the model performance through the cross entropy loss function and the Dice coefficient indicator;

[0043] SA7. After obtaining the segmentation results, structural and color statistical features are extracted. These features include the ratio of the area of ​​the nucleus to the cytoplasm (nucleocytoplasm ratio), the color mean and standard deviation of the nuclear region in the RGB channels, and the roundness, aspect ratio, and perimeter of the nuclear outline, reflecting the morphological characteristics of the cells. The statistical results are then input into the rule engine, which comprehensively analyzes abnormal patterns to determine the specific type of each cell. Classification labels include normal cells, atypical cells, atypical urothelial cells, and degenerate cells. The training goal at this stage is to ensure that the model has a recognition accuracy of at least 95% for abnormal cells. If the actual performance does not meet the target, iterative training will be required by adding samples, adjusting the network structure, or optimizing parameters until the performance requirements are met.

[0044] In the present application, the abnormal cell selection and identification of the sample to be tested according to the abnormal cell selection and identification model includes the following steps:

[0045] SB1. Collect urine samples from patients, centrifuge them, and use a liquid-based thin-layer cytology slide preparation and staining machine to form urine cytology slides. Stain the slides using the Papanicolaou (Pap) staining method.

[0046] SB2, photographing the slide using a full-slide digital slide scanning system with a 40× objective lens, obtaining 40× magnified microscopic images of all specimen positions on the slide, and then stitching them together to generate a complete digital urine cytology slide image;

[0047] SB3. Input the digital image into the trained cell localization model. The model automatically analyzes the cell boundaries and morphological features in the image, identifies and selects all suspected target cell regions, and crops the original image based on the selection results to generate a fixed-size cell slice, which serves as input data for the subsequent recognition model.

[0048] SB4: The generated cell slices are fed into the trained cell recognition model. First, the coarse classification network removes non-target cell types, retaining the urothelial cell images. The fine classification network then further identifies the selected target cell images and analyzes whether there are abnormal morphological features.

[0049] For images identified as target cells, the semantic segmentation model is called to perform fine segmentation of their internal structure, extract the mask images of the cell nucleus and cytoplasm regions, and calculate the key morphological and color feature parameters based on the mask images; in step SB4, each target cell region identified by the positioning model is cropped from the original digital slice image into an independent cell slice image; for each cell slice image, a semantic segmentation model based on a convolutional neural network is used to perform fine segmentation of its internal structure; low-level and high-level image features are extracted through multi-layer convolution, and a semantic mask image is generated through pooling and upsampling operations to distinguish the two regions of the cell nucleus and cytoplasm; the nuclear-cytoplasm ratio can be obtained by calculating the pixel area of ​​the mask area.

[0050] SB5: Determine whether abnormal cells exist in the urine cytology slide image based on the nuclear-cytoplasmic ratio, RGB pixel mean, standard deviation, nuclear circularity, and aspect ratio according to the preset abnormality judgment rules;

[0051] SB6. The system automatically highlights the locations of all cells identified as abnormal on the original digital slice image, and outputs the corresponding image position coordinates and identification results to facilitate subsequent doctor review and diagnostic assistance.

[0052] In this application, as a preferred embodiment, the fixed size is 256 pixels×256 pixels.

[0053] Abnormal cells in urine exfoliated cells of the present invention are identified and classified by an algorithm developed based on a neural network system. The classification of urine exfoliated cells is detailed in Table 1:

[0054] Table 1 Types of urine exfoliated cells

[0055]

[0056]

[0057] The urine exfoliated cells test results are issued based on the type and number of abnormal cells in the urine exfoliated cells, which can achieve early diagnosis and detection of bladder urothelial cancer, reduce patients' pain, and improve diagnostic efficiency.

[0058] The serial numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments. In the above embodiments of the present invention, the description of each embodiment has its own focus. For parts not described in detail in a particular embodiment, please refer to the relevant description of other embodiments. In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented by other means.

[0059] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for automatically detecting abnormal cells in urine exfoliated cells, characterized in that: The following steps are involved: An abnormal cell selection and recognition algorithm is trained to generate an abnormal cell selection and recognition model; and abnormal cell selection and recognition of the sample to be tested is implemented based on the abnormal cell selection and recognition model.

2. The method for automatically detecting abnormal cells in urine exfoliated cells according to claim 1, characterized in that: The training of the abnormal cell selection and recognition algorithm and the generation of the abnormal cell selection and recognition model include the following steps: SA1. Collect urine samples from patients, centrifuge them, and use a liquid-based thin-layer cytology slide preparation and staining machine to form urine cytology slides. Stain the slides using the Papanicolaou (Pap) staining method. SA2, photographing the slide to obtain 40× magnified microscopic images of all specimen positions on the slide, and then stitching them together to generate a complete digital urine cytology slide image; SA3. Medical experts perform instance segmentation and annotation of cells in the digital image, clearly labeling three types of target cells and non-target objects, and generating a labeling model based on the three types of target cells and non-target objects; SA4: Train the tagging model generated in SA3 to ensure that the positioning recall rate is no less than 98.86% and the precision rate is no less than 95.33%; SA5: Automatically crop the cell region identified by the segmentation model in step SA4 into a cell slice of fixed size and input it into a coarse classification convolutional neural network model; the coarse classification network is used to eliminate non-target objects and retain only target cells; SA6: Input the target cell image selected by the coarse classification in SA5 into the trained semantic segmentation model to further perform pixel-level segmentation on the internal structure of the cell; SA7. After obtaining the segmentation results, structural and color statistical features are extracted. These features include: the ratio of the area of ​​the cell nucleus to the cytoplasm, i.e., the nuclear-cytoplasmic ratio; the color mean and standard deviation of the cell nucleus region under the RGB channel; and the roundness, aspect ratio, and perimeter of the cell nucleus outline, thereby reflecting the morphological characteristics of the cell. The statistical results are then input into the rule engine, and the specific type of each cell is determined through a comprehensive analysis of abnormal patterns. The classification labels include normal cells, atypical cells, atypical urothelial cells, and degenerative cells.

3. The method for automatically detecting abnormal cells in urine exfoliated cells according to claim 1, characterized in that: In the SA2, a full-slide digital slide scanning system with a 40× objective lens is used to photograph the glass slide.

4. The method for automatically detecting abnormal cells in urine exfoliated cells according to claim 1, characterized in that: The target cells include normal urothelial cells, atypical urothelial cells and atypia urothelial cells; the non-target objects include degenerative cells, squamous cells, impurities, white blood cells and red blood cells.

5. The method for automatically detecting abnormal cells in urine exfoliated cells according to claim 1, characterized in that: The abnormal cell selection and identification of the sample to be tested according to the abnormal cell selection and identification model includes the following steps: SB1. Collect urine samples from patients, centrifuge them, and use a liquid-based thin-layer cytology slide preparation and staining machine to form urine cytology slides. Stain the slides using the Papanicolaou (Pap) staining method. SB2, photographing the slide using a full-slide digital slide scanning system with a 40× objective lens, obtaining 40× magnified microscopic images of all specimen positions on the slide, and then stitching them together to generate a complete digital urine cytology slide image; SB3. Input the digital image into the trained cell localization model. The model automatically analyzes the cell boundaries and morphological features in the image, identifies and selects all suspected target cell regions, and crops the original image based on the selection results to generate a fixed-size cell slice, which serves as input data for the subsequent recognition model. SB4: The generated cell slices are fed into the trained cell recognition model. First, the coarse classification network removes non-target cell types, retaining the urothelial cell images. The fine classification network then further identifies the selected target cell images and analyzes whether there are abnormal morphological features. For images identified as target cells, the semantic segmentation model is used to perform fine segmentation of their internal structures, extract the mask images of the cell nucleus and cytoplasm regions, and calculate the key morphological and color feature parameters based on the mask images; SB5: Determine whether abnormal cells exist in the urine cytology slide image based on the nuclear-cytoplasmic ratio, RGB pixel mean, standard deviation, nuclear circularity, and aspect ratio according to the preset abnormality judgment rules; SB6. The system automatically highlights the locations of all cells identified as abnormal on the original digital slice image, and outputs the corresponding image position coordinates and identification results to facilitate subsequent doctor review and diagnostic assistance.

6. A method for automatically detecting abnormal cells in urine exfoliated cells according to claim 1 or 5, characterized in that: The fixed size is 256 pixels×256 pixels.

7. The method for automatically detecting abnormal cells in urine exfoliated cells according to claim 5, characterized in that: In step SB4, each target cell region identified by the positioning model is cropped from the original digital slice image into an independent cell slice image; for each cell slice image, a semantic segmentation model based on a convolutional neural network is used to finely segment its internal structure; low-level and high-level image features are extracted through multi-layer convolution, and a semantic mask image is generated through pooling and upsampling operations to distinguish the cell nucleus and cytoplasm. By calculating the pixel area of ​​the mask area, the nuclear-cytoplasm ratio can be obtained.

Images