Artificial intelligence model construction system and method for accurately identifying human circular sperm cells

Abstract

The invention relates to an artificial intelligence model construction system and method for accurately identifying human circular sperm cells. The system comprises an image collection unit, an image preprocessing unit, a data set segmentation unit, a pilot experiment model selection unit, a wrong sample enhanced model training unit and a model verification unit. The data set segmentation unit is used for segmenting the human circular sperm cell image into a training set, a verification set and a test set; the pilot experiment model selection unit is used for selecting the model with the optimal performance according to the performance of the series of to-be-selected models on the verification set; the error sample enhanced model training unit is used for constructing a false positive data set and a false negative data set, adding false positive data and false negative data into a currently trained picture by utilizing a random extraction method, and enhancing the ability of the model to cope with error examples; and the model verification unit is used for carrying out a double-blind experiment, detecting whether the screened cells express the specific protein of the human round sperm cells or not, and verifying the effectiveness and accuracy of the model.

Description

Technical Field

[0001] The present invention relates to the field of human round spermatid identification and artificial intelligence technology, and in particular to an artificial intelligence model construction system and method for accurately identifying human round spermatids. Background Art

[0002] About 50% of patients with nonobstructive azoospermia (NOA) can obtain sperm through testicular sperm aspiration or microspermia and then undergo intracytoplasmic sperm injection (ICSI) treatment (this technique was first reported in 1992) to produce offspring. About 50% of NOA patients do not have mature sperm during surgery and cannot produce offspring. Among them, 30% of NOA patients with spermatogenesis block do not have mature sperm during surgery but haploid human round spermatids (RS) can be seen. The roundspermatid injection (ROSI) technique, which appeared in 1996, injects human haploid RS into oocytes to combine sperm and egg to obtain in vitro cultured embryos, followed by ROSI-embryo transfer (ROSI-ET). This technique may enable NOA patients with spermatogenesis block to produce offspring. It has been nearly 30 years since the ROSI technique was used to treat offspring. The reason why this technique has not been widely used in clinical practice is firstly that the ROSI technique requires non-invasive identification of human round spermatids in clinical practice. The subjectivity, imprecision and inconsistency of the identification based on morphological experience by experienced embryologists make this technique irreproducible. Secondly, compared with the ICSI technique routinely used in the field of assisted reproduction, the clinical outcomes of ROSI are poor. Its embryo fertilization rate (35.99% vs. 63.11%), cumulative clinical pregnancy rate (39.47% vs. 60.19%) and cumulative live birth rate (15.79% vs. 53.90%) are all significantly lower than those of ICSI. The poor clinical outcomes of ROSI technique also come from the inaccuracy of the identification of human round spermatids.

[0003] Artificial Intelligence (AI) shows strong application prospects in the medical field, and the efficiency and accuracy of artificial intelligence have greatly promoted the progress of precision medicine. The AI ​​model constructed by the present invention can non-invasively and 100% accurately screen human round spermatids (RS). The AI ​​model can solve the non-repeatability of the round spermatidinjection (ROSI) technology caused by the inaccuracy and inconsistency of human identification of RS in the current clinic. The wide application and promotion of the model in the clinic will greatly promote the widespread development of ROSI technology in the clinic, so that all patients with non-obstructive azoospermia (NOA) and spermatogenesis blockage patients can benefit from this new technology and have the possibility of obtaining offspring, so it has important clinical application value. Summary of the invention

[0004] The present invention aims to provide an artificial intelligence model construction system and method for accurately identifying human round spermatids. The technical problems to be solved at least include how to accurately screen out human round spermatids non-invasively.

[0005] In order to achieve the above-mentioned object, the present invention provides an artificial intelligence model construction system for accurately identifying human round spermatids, including an image collection unit, an image preprocessing unit, a data set segmentation unit, a pilot experiment model selection unit, an error sample reinforcement model training unit and a model verification unit; The image collection unit is used to capture images of human round spermatids; The image preprocessing unit is used to preprocess the human round spermatid image according to the process of adjusting the image size, performing color normalization, scaling the resolution, segmenting the image, and converting the image into a tensor; The data set segmentation unit is used to segment the human round spermatid image into a training set (a data set for training an artificial intelligence model), a validation set (a data set for adjusting model parameters), and a test set (a data set for evaluating model performance); The pilot experiment model selection unit is used to select the model with the highest accuracy based on the performance of the series of models to be selected on the validation set; The error sample reinforcement model training unit is used to construct a false positive data set and a false negative data set, and selects false positive data from the false positive data set and false negative data from the false negative data set by random sampling method, and repeatedly supplies the selected model by the pilot experiment model selection unit for training, so as to enhance the model's ability to deal with false positive and false negative classes, and at the same time, the regional discarding technology is applied to randomly cut and discard the cell area image in blocks, so as to further enhance the model's robustness to deal with highly variable data; The model verification unit is used to conduct a double-blind experiment to detect whether the screened cells express human round spermatid-specific proteins, and then detect whether the screened cells are human round spermatids, thereby verifying the effectiveness and accuracy of the model.

[0006] Preferably, the image collection unit uses an inverted microscope to capture images of human round spermatids.

[0007] Preferably, the human round spermatids are haploid human round spermatids accurately screened out according to the DNA staining fluorescence intensity by flow cytometry after staining with the live cell staining solution Hoechst 33342.

[0008] Preferably, the human round spermatid-specific protein includes protamine 1 and peanut agglutinin.

[0009] Preferably, the series of models to be selected include R-CNN, Fast R-CNN, and Faster R-CNN models.

[0010] Preferably, the model verification unit uses the constructed AI model to accurately screen human round spermatids, and identifies and verifies these cells by immunofluorescence method. When the constructed AI model can non-invasively screen human round spermatids with an accuracy rate of 100% and does not require any staining, the AI ​​model is determined to be valid.

[0011] Preferably, the false positive data set includes a series of sample data in which the model to be selected mistakenly identifies non-target samples as target samples.

[0012] Preferably, the false negative data set includes a series of sample data in which the model to be selected mistakenly identifies the target samples as non-target samples.

[0013] Compared with the prior art, the beneficial effects of the present invention include: The artificial intelligence model for accurately identifying human round spermatids described in the present invention can non-invasively and 100% accurately screen out human round spermatids from testicular mixed single cells; the use of the AI ​​model does not rely on human experience, and the operator does not require any training; the use of the AI ​​model does not require the addition of any additional equipment, and the pixels of the image identified by the model are consistent with the pixels of the image captured under a conventional inverted microscope; the AI ​​model solves the inaccuracy of the current global clinical assisted reproductive field that relies on manual screening of human round spermatids, can screen RS with 100% accuracy, and will greatly improve the clinical outcomes of ROSI technology; the wide application of the AI ​​model in clinical practice can greatly promote the wide application of ROSI technology in clinical practice, so that all NOA spermatogenesis block patients can receive assisted reproductive treatment and have the possibility of obtaining offspring.

[0014] The evaluation results of the model constructed in this application on the RSs target test results in the test set show that the PR (precision-recall) after removing all background (and class confusion) is 0.921, and the precision-recall graph of the segmented graph shows that the PR after removing all background (and class confusion) is 0.921 (see Figure 5 ), the above results show that the accuracy of the model test is very high. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] The accompanying drawings are used to provide a further understanding of the technical solution of the present invention and constitute a part of the specification. Together with the specific implementation methods of the present application, they are used to explain the technical solution of the present invention and do not constitute a limitation on the technical solution of the present invention.

[0016] Figure 1 Schematic diagram of screening human round spermatids by flow cytometry after Hoechst 33342 staining.

[0017] Figure 2 Schematic diagram of selected human haploid round spermatids.

[0018] Figure 3 It is a flowchart for building an AI model.

[0019] Figure 4 Schematic diagram of the image data training process including confusion repetition and region placement.

[0020] Figure 5 It is a schematic diagram for evaluating the prediction accuracy of the AI ​​model in the human round spermatid test set. DETAILED DESCRIPTION

[0021] The present invention is described in more detail hereinafter to facilitate understanding of the present invention.

[0022] The inventive concept of this application is to first stain the single cells obtained from the testicles with the live cell staining solution Hoechst 33342 (Hoechst 33342), and then accurately screen haploid human round spermatids according to the fluorescence intensity of DNA staining through flow cytometry (this invasive method of identifying RS cannot be used clinically), and use immunofluorescence methods to stain these screened cells with protamine 1 (PRM1) and peanut lectin (PNA), which are specifically expressed molecules of human round spermatids, to prove that these screened cells are human round spermatids. Then, AI deep learning technology is used to automatically learn the unique feature representation of human round spermatids from these 1,134 human round spermatid image data, and further develop an AI model for specific non-invasive and accurate screening of human round spermatids. Finally, the AI ​​model was used to non-invasively and accurately screen haploid human round spermatids from testicular mixed single cells, and these screened cells were stained with human round spermatid-specific expression molecules PRM1 and PNA using immunofluorescence methods, thus proving that these screened cells were human round spermatids.

[0023] The present invention provides a method for constructing an artificial intelligence model for accurately identifying human round spermatids, comprising the following steps: S1. Screening and identification of human round spermatids using flow cytometry: After Hoechst 33342 staining, haploid round spermatids can be accurately distinguished by flow cytometry according to the fluorescence intensity of DNA staining. Using the above method, the present application screened human haploid round spermatids by flow cytometry in the early stage (such as Figure 1 and 2 As shown, Figure 1 Shown in the figure are haploid round spermatids in the screening box; Figure 2 The cell group indicated by the arrow in the figure is human round spermatids). The selected cells were stained with human round spermatid-specific expression molecules PRM1 and PNA using immunofluorescence method, thereby proving that the selected cells are human round spermatids.

[0024] Figure 1 and Figure 2 The results of flow cytometry screening of human round spermatids (RSs) are shown. Human testis single cell suspension was stained with Hoechst 33342 stain and then screened for human round spermatids using flow cytometry. The black rectangular box is a haploid human round spermatid.

[0025] Figure 2In the figure, human testicular single cell suspension was stained with Hoechst 33342 stain and then flow cytometry was used to select human round spermatids. The P5 cell population is RSs, and P6, P7, P8, and P9 are other types of cells. Note: RSs represents human round spermatids.

[0026] S2. Use AI model deep learning technology to develop a model for specific non-invasive and accurate screening of human round spermatids: Deep learning is a model based on artificial neural networks, whose main purpose is to automatically learn effective feature representations from data. In image processing, deep learning forms a convenient tool with the goal of computer vision tasks (including image classification, detection and segmentation). All image data is preprocessed for model development and incorporated into the final AI model. The specific workflow is divided into image collection, image preprocessing, segmentation data set, pilot experiment selection model, model training and verification model with error sample reinforcement ( Figure 3 ), after image preprocessing, this application finally includes 1134 human round spermatid images, and after preprocessing the image data, the target detection model is used to build an AI model using the training set image data and the R-CNN series (R-CNN, Fast R-CNN, Faster R-CNN) as the basic model (Baseline) for data analysis. Then, the validation set image data is used for training to obtain the most accurate model. The process is as follows Figure 4 shown.

[0027] Figure 3 The flowchart of building an AI model is shown. The process specifically includes image collection, image preprocessing, data set segmentation, model building, model optimization and model verification. Figure 3 Includes Hoechst 33342 (Hoechst 33342); Regions with Convolutional Neural Networks (R-CNN); ROC: Receiver Operating Characteristic (ROC); AUC: Area UnderCurve (AUC).

[0028] Figure 4Includes the image data training process of confusion duplication and region drop. Additional objects are randomly sampled in the false positive dataset (FalsePositive Zoo, FP Zoo; the model mistakenly identifies non-target samples as target samples) and the false negative dataset (FalseNegative Zoo, FN Zoo; the model mistakenly identifies target samples as non-target samples) to be added to a given image. False positives and false negatives are collected in real time during training, and FP Zoo and FN Zoo are continuously updated. Note: DropROI means region of interest drop.

[0029] The accuracy of the model in identifying human round spermatids was evaluated by using the test set image data through the receiver operating characteristic (ROC) curve. After removing all background and class confusion, the area under the ROC curve (AUC) was 0.921 ( Figure 5 ), which shows that the accuracy of the model test is very high.

[0030] Figure 5 This is the receiver operating characteristic (ROC) curve of the model for identifying human round spermatids. The evaluation results of the model on human round spermatids showed that the area under the ROC curve (AUC) after removing all background and class confusion was 0.921. The larger the AUC, the higher the accuracy of the model. This result shows that the accuracy of the model test is very high.

[0031] S3. Screening and identification of human round spermatids using AI models: This application further utilizes the constructed AI model to accurately screen human round spermatids. The identification and verification of these cells by immunofluorescence method shows that the accuracy of the model in non-invasive screening of human round spermatids has reached 100% and does not require any staining. In the future, the model will be further converted into an interactive application.

[0032] Based on the above construction method, the present invention also provides an artificial intelligence model construction system for accurately identifying human round spermatids, including an image collection unit, an image preprocessing unit, a data set segmentation unit, a pilot experiment model selection unit, an error sample reinforcement model training unit and a model verification unit; Build models, optimize models using error sample reinforcement model training unit technology, and verify models; The image collection unit is used to capture images of human round spermatids; The image preprocessing unit is used to preprocess the human round spermatid image according to the process of adjusting the image size, performing color normalization, scaling the resolution, segmenting the image, and converting the image into a tensor; The data set segmentation unit is used to randomly segment the human round spermatid images into a training set, a validation set and a test set, which are used to build a model, validate a model and test a model respectively; The pilot experiment model selection unit is used to select the model with the highest accuracy as the appropriate model according to the performance of the series of models to be selected on the validation set; Using data enhancement technology in model training using data in the validation set can enhance the robustness of the model (stability and accuracy of prediction ability). The wrong sample enhancement model training unit proposed in the present invention can enhance the model's ability to deal with wrong examples. This application compares the wrong sample enhancement model training unit with other widely recognized data enhancement technologies, including Mixup, Cutmix, Randaug, Autoaug, and Dropblock. The comparison results show that the wrong sample enhancement model training proposed in the present invention can make the model have the highest accuracy.

[0033] The error sample reinforcement model training unit is used to construct a false positive data set and a false negative data set, and selects false positive data from the false positive data set and false negative data from the false negative data set by random sampling method, and repeatedly supplies the selected model by the pilot experiment model selection unit for training, so as to enhance the model's ability to deal with false positive and false negative classes, and at the same time, the regional discarding technology is applied to randomly cut and discard the cell area image in blocks, so as to further enhance the model's robustness to deal with highly variable data; The model verification unit is used to conduct a double-blind experiment to detect whether the screened cells express human round spermatid-specific proteins, thereby verifying the effectiveness and accuracy of the model.

[0034] Preferably, the false positive data set includes a series of sample data in which the model to be selected mistakenly identifies non-target samples as target samples.

[0035] Preferably, the false negative data set includes a series of sample data in which the model to be selected mistakenly identifies the target samples as non-target samples.

[0036] Preferably, the image collection unit uses an inverted microscope to capture images of human round spermatids.

[0037] Preferably, the human round spermatids are haploid human round spermatids accurately screened out by flow cytometry based on the DNA staining fluorescence intensity after staining with the live cell staining solution Hoechst 33342 (Hoechst33342).

[0038] Preferably, the human round spermatid-specific proteins include protamine 1 (PRM1) and peanut lectin (PNA).

[0039] The above describes the preferred embodiments of the present invention, but it is not intended to limit the present invention. Those skilled in the art may make improvements and changes to the embodiments disclosed herein without departing from the scope and spirit of the present invention.

Claims

1. An artificial intelligence model construction system for accurately identifying human round spermatids, characterized in that: The artificial intelligence model construction system for accurately identifying human round spermatids comprises an image collection unit, an image preprocessing unit, a data set segmentation unit, a pilot experiment model selection unit, an error sample reinforcement model training unit and a model verification unit; The image collection unit is used to capture images of human round spermatids; The image preprocessing unit is used to preprocess the human round spermatid image according to the process of adjusting the image size, performing color normalization, scaling the resolution, segmenting the image, and converting the image into a tensor; The data set segmentation unit is used to randomly segment the human round spermatid images into a training set, a validation set and a test set, which are used to build a model, validate a model and test a model respectively; The pilot experiment model selection unit is used to select the model with the highest accuracy based on the performance of the series of models to be selected on the validation set; The error sample reinforcement model training unit is used to construct a false positive data set and a false negative data set, and selects false positive data from the false positive data set and false negative data from the false negative data set by a random sampling method, and repeatedly supplies the selected model of the pilot experiment model selection unit for training, so as to enhance the model's ability to deal with false positive and false negative classes; The model verification unit is used to conduct a double-blind experiment to detect whether the screened cells express human round spermatid-specific proteins, thereby verifying the effectiveness and accuracy of the model.

2. The artificial intelligence model construction system for accurately identifying human round spermatids according to claim 1, characterized in that: The image collection unit uses an inverted microscope to capture images of human round spermatids.

3. The artificial intelligence model construction system for accurately identifying human round spermatids according to claim 1, characterized in that: The human round spermatids are haploid human round spermatids that are accurately screened according to the DNA staining fluorescence intensity through flow cytometry after being stained with the live cell staining solution Hoechst 33342.

4. The artificial intelligence model construction system for accurately identifying human round spermatids according to claim 1, characterized in that: The human round spermatid-specific proteins include protamine 1 and peanut agglutinin.

5. The artificial intelligence model construction system for accurately identifying human round spermatids according to claim 1, characterized in that: The series of models to be selected include R-CNN, Fast R-CNN and Faster R-CNN models.

6. The artificial intelligence model construction system for accurately identifying human round spermatids according to claim 1, characterized in that: The model verification unit uses the constructed AI model to accurately screen human round spermatids, and identifies and verifies these cells through the immunofluorescence method. When the constructed AI model can non-invasively screen human round spermatids with an accuracy rate of 100% and does not require any staining, the AI ​​model is determined to be valid.

7. The artificial intelligence model construction system for accurately identifying human round spermatids according to claim 1, characterized in that: The training set is a data set used to train the artificial intelligence model; the validation set is a data set used to adjust the model parameters; and the test set is a data set used to evaluate the model performance.

8. The artificial intelligence model construction system for accurately identifying human round spermatids according to claim 1, characterized in that: The false positive data set includes a series of sample data in which the model to be selected mistakenly identifies non-target samples as target samples.

9. The artificial intelligence model construction system for accurately identifying human round spermatids according to claim 1, characterized in that: The false negative data set includes a series of sample data in which the model to be selected mistakenly identifies the target samples as non-target samples.

10. The artificial intelligence model construction system for accurately identifying human round spermatids according to claim 1, characterized in that: The error sample reinforcement model training unit is also used to apply regional discarding technology to randomly cut and discard cell area images in blocks, further enhancing the robustness of the model to cope with highly variable data.

Images