Function convergence and cleaning method for AI-annotated data of few-sample echocardiographic measurements
Patent Information
- Application Number
- CN202410401391.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-04-03
- Publication Date
- 2024-12-06
- Estimated Expiration
- 2044-04-03
AI Technical Summary
In echocardiography examination, there are too many missing values and outliers in the preprocessing and reporting process of data, which makes it difficult for neural network training to converge and affects the prediction accuracy.
The missing values are filled by determining the normal value range, and the outlier value range is compressed into a predetermined target interval using a compression algorithm, and the data set is then normalized and segmented, adjusting the normal value range to optimize model training.
A data distribution is formed that is more suitable for training neural networks, which reduces the impact of outliers, promotes the rapid convergence of the model and improves the prediction accuracy.
Smart Images

Figure CN118312730B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of medical diagnosis technology, and in particular to a function convergence and cleaning method for AI-labeled data of echocardiogram measurements with a small number of samples. Background Art
[0002] In the prior art, multi-dimensional measurement methods are generally used in medical echocardiography to obtain the attribute characteristics of tissues and organs, including quantitative characteristics in terms of morphology, function, structure, etc. These measurement processes will be carried out in accordance with relevant medical guidelines and standards, and standardized measurement processes will be carried out at designated locations. Normal value intervals are also formulated based on factors such as race and region, and different normal value ranges are also based on factors such as age, gender, height and weight. After measurement, the echocardiography examiner refers to these complex normal value ranges to judge indicators such as organ morphology and function, and generates a text report. Due to the large number of measurement values, it is a certain burden for the examiner to judge normal and abnormal. It is usually judged by looking up a table, and then a simple table lookup method is used to implement it. The whole process is to write a complex normal value table into the program configuration in advance, and select the corresponding normal value table by combining the judgment of related variables. The situation of the measured value is judged based on this normal range table, and then the examiner manually generates a report. From the description and conclusion of the report, many conclusions come from the situation of the previous measurement values, which are interrelated.
[0003] At present, with the rapid development of neural network technology, it has become possible to process multi-dimensional information. How to mix measured values through labels, identify measured values through models, and train a multi-classification network structure to map report conclusions. By analyzing the problems of the current model and the actual echocardiography process, we found that obtaining correct and effective report conclusions involves two factors. The first aspect is the data preprocessing process, which allows the values to effectively show the correlation between the defined interval and the conclusion. This part needs to be solved through specific preprocessing methods. The second aspect is the structured definition of the report description and conclusion. This part clarifies the stability of the meaning through the structured definition of the report terms, and finally becomes the conclusion label corresponding to the measured value, which is manifested as a manifestation of medical prior knowledge.
[0004] However, it is difficult to effectively preprocess the values and turn the valid values into a data set suitable for model training. At the same time, through the analysis of the measured values to generate the description report, we found that there are many types of values obtained through the examination, and the various measured values are not complete for different patients. Generally, there will be a large number of missing states. The reason for the missing is that in the entire measured data set, the common measurement parameters only occupy about 1 / 3, and the remaining measurements are designed for different diseases and occupy most of them. In the actual examination process, the common measurements are generally measured first, and then the specific measurements are selected according to different diseases. In this way, many values will be missing due to different diseases; for a measured value, it can be divided into a valid value interval and an abnormal value interval, and the abnormal value interval includes singular and erroneous values; if the values are simply divided into segments for training in the neural network, it will cause excessively large abnormal values and singular values to occupy most of the divided intervals, making it difficult for the neural network training to converge. Summary of the invention
[0005] In order to solve the above problems, the present invention provides a function convergence and cleaning method for AI-labeled data of echocardiography measurements with a small number of samples, which is more suitable for training neural network data distribution and effectively highlights the proportion of effective value data range.
[0006] In order to achieve the above object, the present invention designs a function convergence and cleaning method for AI labeled data of echocardiogram measurements with a small number of samples, and the method comprises the following steps:
[0007] S1. Determine a common value range for acquiring echocardiographic measurement values, which step specifically includes the following sub-steps:
[0008] i. For measurement parameters with clinical value ranges defined by clinical guidelines, the clinical values defined by clinical guidelines
[0009] Range sets the normal value range;
[0010] ii. For measurement parameters that are not defined in clinical guidelines, statistical methods are used to calculate and determine the common value range;
[0011] S2. Fill in the missing values. This step specifically includes the following sub-steps:
[0012] i. Determine the common value range of the measurement parameter to which each missing value belongs, and calculate the mean, minimum, and maximum values of the missing values within the common value range;
[0013] ii. Construct a mathematical model or equation based on the average value, minimum value, and maximum value, use the constructed mathematical model or equation to generate a series of discrete points within the common value range through a random function, and use the generated discrete points to replace the corresponding missing values for filling processing;
[0014] S3. compress the range of outliers. This step specifically includes the following sub-steps:
[0015] i. Define the data points in the data set to which the acquired echocardiographic measurements belong that deviate from the normal value range as abnormal
[0016] Normal values, and determine outliers in the data set through statistical methods;
[0017] ii. compress the outlier range into a predetermined target range through a compression algorithm;
[0018] S4. normalize the compressed data set;
[0019] S5. Split the normalized data set into a training set, a test set, and a validation set. This step specifically includes the following sub-steps:
[0020] i. Shuffle the normalized data set and randomly arrange it to generate a data set with a new order;
[0021] ii. Use the randomized data set and split it into training set, test set, and validation set according to the predetermined ratio;
[0022] S6. Train the AI model and fine-tune the values of the normal value range according to convergence. This step specifically includes the following sub-steps:
[0023] i. Use AI model architecture to perform deep learning based on the characteristics of echocardiographic measurements and the requirements of the prediction task.
[0024] study;
[0025] ii. For data labels that the model prediction results fail to accurately predict, adjust the corresponding column data in the dataset.
[0026] Expand or reduce the interval corresponding to the range of common values;
[0027] iii. Reconstruct the dataset using the adjusted common value range and retrain the model.
[0028] A further technical solution is that, in step S1, the quartiles of the acquired echocardiographic measurement values are determined using a box plot method, and the interval between the upper quartile and the lower quartile is defined as a normal value range.
[0029] A further technical solution is that, in step S3, the target interval for compressing the abnormal value range is 0.5 unit intervals above and below the normal value range.
[0030] A further technical solution is that in step S3, the compression algorithm is a log function based on a specific base, so as to compress the outlier mapping into a target interval.
[0031] A further technical solution is that in step S5, the division ratio of the training set, the test set, and the validation set is 6:2:2.
[0032] The function convergence and cleaning method of AI-labeled data of echocardiographic measurements with a small number of samples designed in the present invention preprocesses the data set to form a data distribution that is more suitable for training a neural network, effectively highlights the proportion of the effective value data range, effectively compresses and narrows the range of abnormal values, reflects the abnormal area in a shorter abnormal interval, and balances the overall distribution of values, so as to promote the correct learning of the spatial mapping of numerical distribution and conclusion expression during the neural network training process, which plays a positive role in accelerating the rapid convergence of the model and improving the accuracy of prediction. BRIEF DESCRIPTION OF THE DRAWINGS
[0033] Figure 1 This is a specific implementation flow chart in Example 1;
[0034] Figure 2 is a schematic diagram of the box plot method in Example 1;
[0035] Figure 3 It is a schematic diagram of the target interval after compression in Example 1. DETAILED DESCRIPTION
[0036] The preferred embodiments of the present invention are described below in conjunction with the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.
[0037] Example 1.
[0038] like Figure 1 As shown, Figure 1 A specific implementation of the function convergence and cleaning method for AI-labeled data of echocardiographic measurements with a small number of samples is shown.
[0039] It should be noted that if substantially the same results are achieved, the method of the present invention is not limited to Figure 1 The process sequence shown is limited to the following steps:
[0040] S1. Determine a common value range for acquiring echocardiographic measurement values, which step specifically includes the following sub-steps:
[0041] i. For measurement parameters with clinical value ranges defined in clinical guidelines, set common value ranges according to the clinical value ranges defined in clinical guidelines;
[0042] Specifically, for measurement parameters defined by clinical guidelines, such as the ASE / EACVI guidelines and standards (Recommendations for the assessment of left ventricular diastolic function by echocardiography: updated by the American Society of Echocardiography and the European Society of Cardiovascular Imaging), a common value range is set, and in this embodiment, the common value range adopts the range interval currently related to the conclusion in clinical practice, including a normal range, a graded abnormal range, a singular value, etc. For example: the left atrial anteroposterior diameter measurement value range is 1-10 cm, of which 1-3.8 cm is a normal range, greater than 3.8 cm is an abnormal range, less than 1 cm and greater than 10 cm are singular values, and the aortic valve peak blood flow pressure difference value range is 5-40 mmHg, of which 1-10 mmHg is a normal range, 10-20 mmHg is mild stenosis, 20-40 mmHg is moderate stenosis, greater than 40 mmHg is severe stenosis, and less than 1 mmHg is a singular range.
[0043] It should be noted that the common value ranges corresponding to different measurement parameters are defined differently to facilitate the subsequent compression and slicing of data.
[0044] ii. For measurement parameters that are not defined in clinical guidelines, statistical methods are used to calculate and determine the common value range;
[0045] For example, for cases where there is no unified reference range for values such as mitral E peak velocity and tricuspid E peak ratio in clinical practice, the box plot method can be used to determine the quartiles of the echocardiographic measurement values, and the interval between the upper quartile and the lower quartile is defined as the common value range.
[0046] Specifically, if Figure 2 As shown in the figure, echocardiographic data of a certain number of healthy people are collected, the mitral valve E peak velocity and tricuspid valve E peak ratio data measured in the past are extracted and calculated, a data sequence (ME / TE) is constructed, and it is sorted. Then, the position Q1, Q2, Q3, upper edge, and lower edge parameters are calculated using the tertile, quartile, and quintile methods, and [Q1:Q3] is taken as the normal value range interval, and (upper edge: Q3, Q1: lower edge) is taken as the abnormal value interval.
[0047] S2. Fill in the missing values. This step specifically includes the following sub-steps:
[0048] i. Determine the common value range of the measurement parameter to which each missing value belongs, and calculate the mean, minimum, and maximum values of the missing values within the common value range;
[0049] ii. Construct a mathematical model or equation based on the average value, minimum value, and maximum value, use the constructed mathematical model or equation to generate a series of discrete points within the common value range through a random function, and replace the corresponding missing values with the generated discrete points for filling;
[0050] In this embodiment, missing values refer to measurement values that are not recorded during the data collection process. At the same time, in the specific implementation, because not all measurement values need to be measured during a patient's examination, many of the data are added according to the condition. For example, in the examination of patients with valvular stenosis, in order to further average the degree of stenosis, it is necessary to add further judgment indicators for the degree of valvular lesions. For example, a large number of missing values appear in the measurement parameter columns under the average pressure gradient of the valve orifice, velocity time integral, etc., showing sparse characteristics. Therefore, the missing values need to be replaced with specific values during the model training process. However, if they are simply replaced with extreme values or means, it will inevitably cause the model training process to not converge, and these measurement items will point to errors in the high-dimensional space of the model.
[0051] Therefore, after obtaining the normal value range of the measurement parameter through the aforementioned step S1, a mathematical model or equation is constructed based on the average value, minimum value, and maximum value calculated by the above method. This model can be linear, nonlinear, or customized according to the data distribution characteristics. For example, if the data is normally distributed, the mean and standard deviation of the normal distribution can be used to construct the model. If the data distribution is irregular, more complex statistical models may need to be used, such as piecewise linear regression or polynomial regression; then a random function is used to generate uniformly distributed discrete points within this range, and these discrete data are used to replace missing values to simulate the natural variability of real data, avoid introducing new deviations, and effectively optimize the data to improve the rapid convergence of the model training process and improve the robustness of the model. Compared with directly using extreme values or mean filling, this method can more accurately reflect the true distribution of the data and reduce the negative impact on model training.
[0052] For example: in the previous step, the average, minimum and maximum values are obtained by calculation, and then the average mean and standard deviation parameters can be calculated, and then the Gaussian distribution random number generation function (such as random.gauss() used in python) is used to generate random numbers to replace missing values. That is, the Gaussian distribution random function can be understood as publishing linearity and generating nonlinearity; when the corresponding data items show non-normal distribution, we use the maximum and minimum values to generate random numbers (such as random() in python).
[0053] S3. compress the range of outliers. This step specifically includes the following sub-steps:
[0054] i. defining data points in the data set to which the acquired echocardiographic measurement values belong that deviate from the normal value range as outliers, and determining the outliers in the data set by statistical methods;
[0055] In the aforementioned step S1, the normal value range is determined based on clinical practice and statistical analysis, while the value range of extreme abnormal values and singular values will be several times or even hundreds of times the "normal value range". For example, the "normal value range" of the anteroposterior diameter of the left atrium is 1-10cm, while the extreme abnormal value range may be 10-30cm, and the singular value may be 100cm. This will cause a large number of slices to appear in the abnormal value area when the value range is sliced equidistantly in the later stage, resulting in a reduction in the number of slices in the "normal value range" and fuzzy prediction of the corresponding range of the result value. Therefore, it is necessary to effectively compress and control the abnormal value range.
[0056] ii. compress the outlier range into a predetermined target range through a compression algorithm;
[0057] In this embodiment, the target interval for compression of the abnormal value range is 0.5 unit intervals above and below the normal value range. For example, if the normal value range is [2-6] cm, the compressed target interval will be [1.5-6.5] cm.
[0058] The specific formula is as follows: Figure 3 As shown in the figure, the extremely abnormal values and singular values are explained by taking the normal value range of the left atrial anterior-posterior diameter as an example:
[0059] a) Domain range: [0-50]
[0060] b) Common value range: [2-6]
[0061] c) Example of measurement value: 8cm
[0062] d) Segment the value range:
[0063] "Low anomaly interval": domain [0-2] range [0-0.5]
[0064] "Normal value interval": domain [2-6] range [0.5-1.5]
[0065] “High anomaly interval”: domain [6-50] range [1.5-5]
[0066] e) For each interval, apply a different compression formula to map the value to the target interval:
[0067] "Low anomaly interval": nonlinear compression maps the result to the interval [0-0.5]
[0068] Formula 1:
[0069] "Normal value interval": after linear compression, the result is mapped to the interval [0.5-1.5]
[0070] Formula 2: y=(x-2) / (6-2)+0.5
[0071] "High anomaly interval": nonlinear compression maps the result to the interval [1.5-2]
[0072] Formula 3: y=0.5*(log e (x-6)+1.5) / (50-6)+1.5
[0073] Example: x=8 calculate y=1.6443012470399492
[0074] This shows that the 8cm measurement value was compressed into the target interval [1.5-2].
[0075] In summary, segmented processing and application of specific formulas can effectively compress the range of outliers, thereby optimizing the dataset and improving the efficiency and robustness of model training. This approach helps ensure that the model can learn the distribution characteristics of the data more accurately while reducing the impact of outliers on model predictions.
[0076] S4. normalize the compressed data set;
[0077] Specifically, in order to meet the input requirements of the Transformer neural network, we need to map the compressed data set to a specified range space area. For example, we can choose the integer domain [0-100] as the normalized value range space, for example: define the normalized value range space [0-100] integer domain, range range = 100, slice interval d = 1, and the sliced value space is [0-100 / d].
[0078] Formula 4: y=int(range*x / 2 / d)
[0079] Among them, x is the original compressed data value, y is the normalized value, and the normalized data will be divided into equally spaced slices, each slice representing a numerical range. In this example, since the slice interval d is 1, the normalized value range [0-100] will be divided into 101 equally spaced slices, and the normalized data points will be used for the word vector encoding of the neural network, that is, in the Transformer model, these values will be converted into corresponding word embeddings (token embeddings) so that the model can understand and process these numerical information.
[0080] S5. Split the normalized data set into a training set, a test set, and a validation set. This step specifically includes the following sub-steps:
[0081] i. Shuffle the normalized data set and randomly arrange it to generate a data set with a new order;
[0082] Randomizing the normalized data set ensures the randomness of the data, which helps avoid overfitting during model training. For example, assuming that the data set contains 10,000 records, the DataLoader class in the PyTorch framework can be used to randomly shuffle the data set through its built-in shuffle function to achieve random arrangement of the data.
[0083] ii. Use the randomized data set and split it into training set, test set, and validation set according to the predetermined ratio;
[0084] In this embodiment, the division ratio of the training set, the test set, and the validation set is 6:2:2, that is, 60% (training set), 20% (validation set), and 20% (test set), which is used for training the neural network.
[0085] S6. Train the AI model and fine-tune the values of the normal value range according to convergence. This step specifically includes the following sub-steps:
[0086] i. Use AI model architecture to perform deep learning based on the characteristics of echocardiographic measurements and the requirements of the prediction task.
[0087] Learning; AI models can use existing mature models;
[0088] ii. For data labels that the model prediction results fail to accurately predict, adjust the corresponding column data in the dataset.
[0089] Expand or reduce the interval corresponding to the range of common values;
[0090] In specific implementation, for data labels that the model prediction results fail to accurately hit, the relevant column data in the data set is analyzed. If it is found that the common value range of a certain measurement value (such as the anteroposterior diameter of the left atrium) fails to effectively distinguish specific clinical labels (such as left atrial enlargement), the common value range needs to be adjusted. For example, if the original common value range is [2-6] cm, it can be reduced to [2-3.5] cm, or expanded to [2-10] cm to better reflect the distribution of abnormal areas.
[0091] iii. Reconstruct the dataset using the adjusted common value range and retrain the model. In specific implementation, this step may need to be iterated until the model achieves satisfactory performance on the validation set.
[0092] The function convergence and cleaning method for AI-labeled data of echocardiographic measurements of a small number of samples provided in this embodiment preprocesses the data set to form a data distribution that is more suitable for training a neural network, effectively highlights the proportion of the valid value data range, effectively compresses and narrows the range of abnormal values, reflects the abnormal area in a shorter abnormal interval, and balances the overall distribution of values, so as to promote the correct learning of the spatial mapping of numerical distribution and conclusion expression during the neural network training process, which plays a positive role in accelerating the rapid convergence of the model and improving the accuracy of prediction.
[0093] In the description of the present invention, it should be noted that the terms "vertical", "up", "down", "horizontal", etc. indicate orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation on the present invention.
[0094] In the description of the present invention, it is also necessary to explain that, unless otherwise clearly specified and limited, the terms "set", "install", "connect", and "connect" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be indirectly connected through an intermediate medium, or it can be the internal communication of two elements. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.
[0095] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art can still modify the technical solutions described in the aforementioned embodiments or replace some of the technical features therein by equivalents. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.
Claims
1. A function convergence and cleaning method for AI-annotated data of echocardiographic measurements with a small number of samples, characterized in that: The method comprises the following steps: S1. Determine a common value range for acquiring echocardiographic measurement values, which step specifically includes the following sub-steps: i. For measurement parameters with clinical value ranges defined in clinical guidelines, set common value ranges according to the clinical value ranges defined in clinical guidelines; ii. For measurement parameters that are not defined in clinical guidelines, statistical methods are used to calculate and determine the common value range; S2. Fill in the missing values. This step specifically includes the following sub-steps: i. Determine the common value range of the measurement parameter to which each missing value belongs, and calculate the mean, minimum, and maximum values of the missing values within the common value range; ii. Construct a mathematical model or equation based on the average value, minimum value, and maximum value, use the constructed mathematical model or equation to generate a series of discrete points within the common value range through a random function, and use the generated discrete points to replace the corresponding missing values for filling processing; S3. compress the range of outliers. This step specifically includes the following sub-steps: i. defining data points in the data set to which the acquired echocardiographic measurement values belong that deviate from the normal value range as outliers, and determining the outliers in the data set by statistical methods; ii. compress the outlier range into a predetermined target range through a compression algorithm; S4. normalize the compressed data set; S5. Split the normalized data set into a training set, a test set, and a validation set. This step specifically includes the following sub-steps: i. Shuffle the normalized data set and randomly arrange it to generate a data set with a new order; ii. Use the randomized data set and split it into training set, test set, and validation set according to the predetermined ratio; S6. Train the AI model and fine-tune the values of the normal value range according to convergence. This step specifically includes the following sub-steps: i. Use AI model architecture for deep learning based on the characteristics of echocardiographic measurements and the requirements of the prediction task; ii. For data labels that the model prediction results fail to accurately predict, adjust the corresponding column data in the dataset. Expand or reduce the interval corresponding to the range of common values; iii. Reconstruct the dataset using the adjusted common value range and retrain the model.
2. The function convergence and cleaning method for AI-annotated data of echocardiographic measurements of a small number of samples according to claim 1, characterized in that: In step S1, the quartiles of the acquired echocardiographic measurement values are determined using a box plot method, and the interval between the upper quartile and the lower quartile is defined as the common value range.
3. The function convergence and cleaning method for AI-annotated data of echocardiographic measurements of a small number of samples according to claim 1, characterized in that: In step S3, the target interval for compressing the abnormal value range is 0.5 unit interval above and below the normal value range.
4. The function convergence and cleaning method for AI-annotated data of echocardiographic measurements with a small number of samples according to claim 1 or 3, characterized in that: In step S3, the compression algorithm is a log function based on a specific base, so as to compress the outlier mapping into a target interval.
5. The function convergence and cleaning method for AI-annotated data of echocardiographic measurements with a small number of samples according to claim 1, characterized in that: In step S5, the split ratio of the training set, test set, and validation set is 6:2:2.
