Automatic quantitative grading system for digital information of immature ovarian teratoma
By pre-processing and neural network model analysis of digital scanning sections of immature ovarian teratomas, a global characteristic grading model was constructed, which solved the subjectivity and repetition of ovarian immature teratoma grading, and achieved accurate automatic grading and clinical guidance.
Patent Information
- Application Number
- CN202510441776.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-09
- Publication Date
- 2025-07-25
AI Technical Summary
The prior art has high subjectivity and poor repetition in the tumor classification of immature ovarian teratomas, making it difficult to accurately reflect pathological characteristics, and affects the formulation of treatment plans.
The preprocessing method and neural network model based on digital scanning slices are adopted, and the global characteristic analysis is performed, and the slice grading model is automatically rated through global characteristics, combining immunohistochemistry characteristics, and the slice grading model based on key regions is constructed, and the Gaussian smooth denoising and global skeleton network are used for precise grading.
Accurate and automatic grading of immature ovarian teratomas is achieved, the accuracy of grading and correlation with clinical prognosis are improved, and clinical treatment is guided.
Smart Images
Figure CN120375054A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of medical devices, and particularly to the field of diagnosing ovarian immature teratoma using physiological parameters. Specifically, it is a method for automatically quantifying and grading digital information of ovarian immature teratoma. Background Art
[0002] Ovarian immature teratoma (IMT) is a rare malignant tumor of the female reproductive system, mostly occurring in children and young women. The treatment plan for IMT is often determined according to the tumor grade to decide whether postoperative chemotherapy is needed. Therefore, the tumor grading of IMT has certain guiding significance for treatment. However, the tumor grading of IMT is highly subjective and has poor repeatability. In order to explore high-risk pathological features related to clinical prognosis such as recurrence and metastasis, and improve the guiding role of different pathological grades for clinical prognosis, an automatic quantification and grading method for IMT based on digital information is proposed. By completing digital slide scanning, using the advantages of digital quantification, and combining the rapid development of intelligent technology in recent years, automatic identification and grading of pathological morphology and immunohistochemical features are carried out to realize and improve the guiding role of IMT grading for clinical practice.
[0003] In recent years, with the development of automated tissue sectioning and whole slide scanning technology, a large amount of digital pathological image data of cancer has been accumulated. At the same time, thanks to the rapid development of artificial intelligence technology, using artificial intelligence technology to predict cancer digital pathological images can greatly improve the prediction accuracy. Therefore, it is urgent to deeply study artificial intelligence methods for cancer digital pathological image prediction to achieve accurate prediction of cancer pathological images and assist doctors in diagnosis and treatment. Most current cancer digital pathological image prediction methods based on artificial intelligence technology use convolutional neural network models to achieve object-level prediction classification, such as cell segmentation and recognition, which do not consider the distribution characteristics between objects and the global characteristics of the slide, so they have great limitations. Summary of the Invention
[0004] To solve the problems, the present invention proposes a method for automatically quantifying and grading digital information of ovarian immature teratoma. Based on the digital scanned slides of ovarian immature teratoma, an artificial intelligence method is used to analyze the global characteristics of the slides, and automatic grading is achieved according to the pathological morphology and immunohistochemical features of the slides, so as to play a guiding role in clinical practice.
[0005] An automatic quantification and grading system for digital information of ovarian immature teratoma, characterized in that: it includes a processing unit that runs the following algorithm:
[0006] Step 1 Preprocessing of digital scanned slides:
[0007] S1.1 Calculate the equilibrium coordinates of each unit of the original slide data;
[0008] S1.2 Calculate the linear smoothing index graph;
[0009] S1.3 Perform Gaussian smoothing denoising on the original slice data according to the linear smoothing index σ(x, y), where the variance of the Gaussian function is determined according to the linear smoothing index;
[0010] Step 2: Construct a slice classification neural network model based on key regions:
[0011] The input of the model includes the denoised slice data I(x, y) obtained in Step 1, and several pre-labeled sub-regions S i (x′, y′),
[0012] The first layer of the model is:
[0013]
[0014] where α 1j is the convolution kernel acting on the key region, β 0j is the corresponding linear bias, P ij is the feature space of the same size as the key region, and the subscripts i, j mark its independent numbers. i is the sub-region label, and μ is the activation function of the neural network;
[0015] The second layer of the model is:
[0016]
[0017] ω 21 、ω 22 are the fully connected linear coefficients acting on the global, β1 is the corresponding linear bias, Q is the fully connected output vector of the second layer, and z is the vector label;
[0018] After passing through the output layer, the classification result is output.
[0019] It also includes a tumor slice, a scanning camera, a memory, and a display device.
[0020] The tumor slice is a slice of an ovarian immature teratoma sample from a clinical patient with ovarian immature teratoma.
[0021] The scanning camera is used to collect digital images of the ovarian immature teratoma sample slice.
[0022] The scanning camera includes a microscope
[0023] The microscope includes an optical microscope, an electron microscope, a fluorescence microscope, or a phase contrast microscope.
[0024] The processing unit is further configured to display the digital image of the training sample slice on a display device, so that an operator can mark the digital image according to the image displayed on the display device to form a training sample that can be utilized by a neural network.
[0025] A memory is configured to store the digital image of the ovarian immature teratoma sample slice collected by a scanning camera, and the digital image of the ovarian immature teratoma training sample slice collected by the scanning camera after being marked by the operator.
[0026] A display device is configured to display the digital image of the ovarian immature teratoma sample slice to facilitate observation and marking by an operator.
[0027] The cost function adopted during the training process of the neural network model is
[0028]
[0029] where S(n) is the actual output, is the marked output.
[0030] The inventive points and technical effects of the present invention:
[0031] 1. By completing digital slice scanning, taking advantage of the quantifiable nature of digitization and combining with the rapid development of intelligent technologies in recent years, automatic identification and grading of pathological morphology and immunohistochemical features are carried out to realize and improve the guiding role of IMT grading for clinical practice. In particular, through the slice preprocessing method, denoising processing specifically for ovarian immature teratoma sample slices is achieved, removing the noise that often appears in this tumor sample, ensuring the accuracy of the input neural network image, and thus improving the classification accuracy.
[0032] 2. A method for marking digital scanned slices and learning of a neural network model is proposed. Compared with classical methods, it can better reflect the pathological features of IMT. Further, through a global skeleton network, global modeling of local pathological features is realized, thereby achieving the goal of accurate slice grading. BRIEF DESCRIPTION OF THE DRAWINGS
[0033] Figure 1 It is an image of a sample slice with local markings. DETAILED DESCRIPTION OF THE EMBODIMENTS
[0034] To solve the above technical problems, the present invention proposes an automatic grading system and method for ovarian immature teratoma based on a neural network model.
[0035] The system includes: a tumor slice, a scanning camera, a processing unit, a memory, and a display device.
[0036] The tumor slice is an ovarian immature teratoma sample slice of a clinical ovarian immature teratoma patient.
[0037] A scanning camera for collecting digital images of sections of ovarian immature teratoma samples. The scanning camera may include a microscope for scanning magnified images of the samples. The microscope includes an optical microscope, an electron microscope, a fluorescence microscope, a phase contrast microscope, etc.
[0038] A processing unit for receiving digital images of sections of ovarian immature teratoma samples collected by the scanning camera and storing them in a memory; displaying digital images of training sample sections on a display device according to an algorithm so that an operator can mark the digital images based on the images displayed on the display device to form training sample images that can be utilized by a neural network; constructing a neural network model in the processing unit and training it using the training sample images to form a final neural network model, so that during actual use, the ovarian immature teratoma sample can be automatically quantified and graded according to the input digital image of the ovarian immature teratoma detection sample section. The specific processing method will be described in detail below.
[0039] A memory for storing digital images of sections of ovarian immature teratoma samples collected by the scanning camera, storing digital images of ovarian immature teratoma training sample sections marked by the operator, storing codes related to the neural network model, and storing the results of automatic quantification and grading of ovarian immature teratoma detection samples.
[0040] A display device for displaying digital images of sections of ovarian immature teratoma samples for the operator to observe and mark; for displaying ovarian immature teratoma detection samples and their results of automatic quantification and grading.
[0041] Specifically, it is implemented in three steps:
[0042] First, obtain digital scanned sections and store them in image form. After preprocessing, transfer to the next step.
[0043] Second, mark the preprocessed sections as reference learning samples and design a model to learn from the learning samples.
[0044] Finally, implement automatic quantification and grading of the sections according to the results of model learning.
[0045] The following algorithm is implemented in the processing unit:
[0046] Step 1: A preprocessing method for digital images of sections of ovarian immature teratoma samples (digital scanned sections).
[0047] Suppose the originally obtained digital scanned section is represented in matrix form as
[0048] D(x,y)
[0049] x and y are the coordinates of each quantization unit in the slice space.
[0050] After quantization, there is local noise in the slices obtained by using existing equipment, which affects the learning and automatic grading of subsequent steps. The present invention proposes an innovative preprocessing method, which can significantly improve the performance of automatic grading after processing the original slice data.
[0051] S1.1 Calculate the equilibrium coordinates of each unit of the original slice data.
[0052] Definition of equilibrium coordinates
[0053]
[0054] Where
[0055]
[0056] And
[0057]
[0058] The double vertical bars represent taking the absolute value. x1, y1…x8, y8 represent the coordinates of 8 surrounding units of x and y, and their relationship is as follows
[0059]
[0060] The equilibrium coordinates are regarded as the correction of the data under the assumption of linear smoothing of local changes. The degree of linear smoothing of local changes is evaluated by the correction amplitude.
[0061] S1.2 Calculate the linear smoothing index map
[0062]
[0063] Where (x′, y′) represents the coordinates of a unit in the neighborhood centered on x and y. And
[0064]
[0065] S1.3 According to the linear smoothing index σ(x, y), perform Gaussian smoothing denoising on the original slice data, and the variance of the Gaussian function is determined according to the linear smoothing index. Obtain the denoised slice data, that is
[0066]
[0067] Where
[0068]
[0069] e is the natural exponential function. u and v are the relative coordinates within the Gaussian smoothing window.
[0070] The above preprocessing method can significantly improve the denoising effect of preprocessing compared with the classical constant variance Gaussian filtering function, making the subsequent steps more robust to the noise of the original data.
[0071] Step 2: Marking and learning method for digital scanned slices.
[0072] Most of the current classical cancer digital pathology image prediction methods based on artificial intelligence technology use convolutional neural network models to achieve object-level prediction and classification, such as cell segmentation and recognition. They do not consider the distribution characteristics between objects and the global characteristics of slices, and have limitations.
[0073] An innovative slice grading model based on key regions is proposed, and a marking and learning method is further proposed. The key region refers to the local region in the IMT digital scanned slice that has been professionally demonstrated and pre-marked, such as Figure 1 shown in the example of local marking of sample slices. The convolutional layer adopted in the present invention extracts features only within this region, and can better reflect the IMT pathological features compared with the classical method. Further, through the global skeleton network, the global modeling of local pathological features is realized, so as to achieve the goal of slice grading.
[0074] The input of the above model includes the denoised slice data I(x, y) obtained in step 1, and several pre-marked sub-regions S i (x′, y′), where (x′, y′) are the coordinates in the sub-region, and each sub-region is a subset of the slice data.
[0075] The global model is further implemented as follows. Its first layer is defined as
[0076]
[0077] where α 1j is the convolution kernel acting on the key region, and β 0j is the corresponding linear bias, which is used to extract features in the key region and map them to the feature space P ij . P ij is a feature space of the same size as the key region, and the subscripts i, j mark its independent numbers. i is the sub-region mark, and in the present invention, 16 groups are taken. j is the convolution kernel mark, and in the present invention, 32 groups are taken. Therefore, the size of the feature space is w′×h′×512. w′ and h′ are the sizes of the sub-regions.
[0078] μ is the activation function of the neural network, defined as follows (the same below)
[0079]
[0080] e is the natural exponential function. Compared with the classical sigmoid function, min represents taking the smaller value of the two. The above activation function can improve the IMT grading performance.
[0081] The first convolutional layer extracts features within the marked key regions, which can better reflect the IMT pathological features and improve the grading performance compared with the classical method (extracting features from the entire image).
[0082] The second layer of the model is defined as
[0083]
[0084] ω 21 and ω 22 are fully connected linear coefficients acting globally, β1 is the corresponding linear bias. Q is the fully connected output of the second layer. A global skeleton network is established through the above fully connected coefficients to jointly model the global features and the local features of the key regions to achieve slice grading. z is the vector label.
[0085] μ is the activation function of the neural network.
[0086] It can be understood that if more accurate results are needed, convolutional and pooling layers can also be added.
[0087] The model outputs three independent grades as follows
[0088]
[0089] where
[0090]
[0091] ω3 is the linear coefficient of the output layer, β2 is the corresponding linear bias. μ is the activation function of the neural network. n represents the number of grading levels.
[0092] It can be understood that more refined classification at more levels can be achieved with the support of more samples, and the model fully supports this.
[0093] During marking, first mark the grades of the learning samples, and then mark the key regions. After all marking is completed, input the marked learning samples into the above model, and the marked output has the corresponding grading level of 1, and the remaining grades are 0.
[0094] Adopt the cost function
[0095]
[0096] Substitute the learning samples and train the above model.
[0097] Step 3: After training, use the model to evaluate the input samples, and take the maximum value of the output series as the evaluation grade.
[0098] The present invention provides a method for automatically quantifying and grading digital information of ovarian immature teratoma. Experimental results show that the method of the present invention realizes automatic grading according to the pathological morphology and immunohistochemical characteristics of slices, and the grading results have a high correlation with the high-risk pathological characteristics of clinical prognosis, which plays a guiding role in the clinical treatment of IMT.
[0099] Table: Correlation between grading and high-risk prognostic features
[0100]
[0101]
[0102] The above table shows the correlation between the classification results obtained by using the classical convolutional network, the neural network of the present invention (Step 2), and the neural network after the improved preprocessing of the present invention (Steps 1 and 2) for 63 groups of slices in this hospital and the high-risk features obtained by actual biochemical detection. It can be seen from the table that the present invention can significantly improve the accuracy of classification through Steps 1 and 2, and it can basically distinguish the slices related to high-risk features, which has clinical guiding significance.
[0103] In each embodiment of the present application, the functional modules can be integrated together to form an independent part, or each module can exist alone, or two or more modules can be integrated to form an independent part. If the function is implemented in the form of a software functional module and sold or used as an independent product, it can be stored in a computer-readable storage medium. Therefore, the technical solution of the present application can be embodied in the form of a software product, or this computer software product is stored in a storage medium, or this computer software product runs on a computer device, and such a computer device includes not only personal computers and servers but also mobile terminals.
[0104] In several embodiments provided by the present application, it should be understood that the disclosed device and method can also be implemented in other ways. The device embodiments described above are only illustrative. Therefore, the above description is only for the embodiments of the present application and is not used to limit the protection scope of the present application. For those skilled in the art, the present application can have various changes and modifications. Any changes or substitutions that can be easily thought of by those skilled in the art within the technical scope disclosed by the present application should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.
Claims
1. A digital information automatic quantification and grading system for ovarian immature teratoma, characterized in that: It includes a processing unit that runs the following algorithm: Step 1: Preprocessing of digital scanned slices: S1.1 Calculate the equilibrium coordinates of each unit of the original slice data; S1.2 Calculate the linear smoothing index map; S1.3 Perform Gaussian smoothing denoising on the original slice data according to the linear smoothing index σ(x,y), where the variance of the Gaussian function is determined according to the linear smoothing index; Step 2: Construct a slice grading neural network model based on key regions: The input of the model includes the denoised slice data I(x, y) obtained in step 1, and several pre-labeled sub-regions S i (x′, y′), The first layer of the model is: where α 1j is the convolution kernel acting on the key area, β 0j is the corresponding linear bias, P ij is the feature space of the same size as the key area, the subscripts i, j mark its independent numbers, i is the sub-area mark, and μ is the activation function of the neural network; The second layer of the model is: ω 21 、ω 22 are fully connected linear coefficients acting globally, β1 is the corresponding linear bias, Q is the fully connected output vector of the second layer, and z is the vector label; After passing through the output layer, the grading result is output.
2. The system according to claim 1, wherein: It also includes a tumor slice, a scanning camera, a memory, and a display device.
3. The system according to claim 2, wherein: The tumor slice is a sample slice of ovarian immature teratoma from a clinical patient with ovarian immature teratoma.
4. The system according to claim 2, wherein: The scanning camera is used to collect digital images of the ovarian immature teratoma sample slices.
5. The system according to claim 4, wherein: The scanning camera includes a microscope.
6. The system according to claim 5, wherein: The microscope includes an optical microscope, an electron microscope, a fluorescence microscope, or a phase contrast microscope.
7. The system according to claim 2, wherein: The processing unit is also used to display the digital image of the training sample slice on the display device so that the operator can mark the digital image according to the image displayed on the display device to form a training sample that can be utilized by the neural network.
8. The system according to claim 2, wherein: The memory is used to store the digital images of the ovarian immature teratoma sample slices collected by the scanning camera and the digital images of the ovarian immature teratoma training sample slices marked by the operator.
9. The system according to claim 2, wherein: The display device is used to display the digital images of the ovarian immature teratoma sample slices for the operator to observe and mark.
10. The system according to claim 1, characterized in that: The cost function adopted during the training process of the neural network model is where S(n) is the actual output, is the marked output.