A method and program product for identifying ovarian cysts based on ultrasonic medical images
Through the recognition method based on ultrasound medical images, the imagingomics characteristics of ovarian cysts were extracted and the cascaded multi-layer perceptron model was used to solve the problem that the benign and malignant risks of ovarian cysts were not effectively identified in the prior art, and the ovarian cysts were realized. Accurate prediction and uncertainty analysis improve the accuracy and reliability of the diagnosis.
Patent Information
- Application Number
- CN202211009842.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-08-22
- Publication Date
- 2024-11-15
- Estimated Expiration
- 2042-08-22
AI Technical Summary
The prior art cannot effectively identify and predict the benign and malignant risks of ovarian cysts, resulting in the failure to detect potential malignant lesions in time, which may threaten the patient's life.
Based on ultrasound medical images, the geometric features, imagingomics and texture features of the ovarian cyst area were extracted, and the benign and malignant risks of ovarian cysts were predicted using a cascade multi-layer perceptron model, and the results were evaluated in combination with uncertainty values. and threshold setting, providing a more objective diagnostic reference.
It realizes the prediction and uncertainty analysis of benign and malignant risks of ovarian cysts, provides more objective identification results, helps users make medical judgments in different application scenarios, and improves the accuracy and reliability of the diagnosis.
Smart Images

Figure CN115376127B_ABST
Abstract
Description
Technical Field
[0001] The invention discloses an ovarian cyst recognition method based on ultrasonic medical images and a program product, belonging to the technical field of medical image intelligent analysis. Background Art
[0002] Ovarian cysts have a high prevalence in women of childbearing age, and most ovarian cysts are functional ovarian cysts. Most ovarian cysts can disappear on their own, but if severe cases are not diagnosed and treated in time, they may be secondary to infection, cyst rupture, pedicle torsion, etc., which will have an adverse effect on the patient's life and work, and even threaten their lives. Especially for cysts with malignant lesions, their prognosis is related to the stage, the size of the residual lesions after the initial surgery, and the pathological type. Therefore, if the type of ovarian cyst can be identified and predicted based on the patient's abdominal ultrasound, malignant cysts can be identified as early as possible to save the patient's life. Summary of the invention
[0003] In view of the deficiencies in the prior art, the present invention discloses an ovarian cyst recognition method based on ultrasonic medical images.
[0004] The invention also discloses a program product for realizing the identification method.
[0005] The detailed technical scheme of the present invention is as follows:
[0006] A method for identifying ovarian cysts based on ultrasonic medical images, comprising:
[0007] 1) Obtain an abdominal ultrasound image that includes the ovarian area on the affected side:
[0008] 2) Region of interest ROI extraction:
[0009] For the diseased ovarian region, extract the ovarian cyst region x1 and the ovarian region x in the ultrasound image;
[0010] 3) Radiomics feature calculation module:
[0011] Perform feature calculation on the region extracted in step 2): Use the radiomics feature calculation package pyradiomics in Python to calculate features. For each patient, multiple radiomics features can be obtained. The radiomics features that need to be calculated include the geometric features, first-order statistical features, texture features, and wavelet features of the ovarian cyst region x1 and the ovarian region x.
[0012] 4) Establish an ovarian cyst identification and prediction model:
[0013] It includes three cascaded multilayer perceptrons (MLPs); the inputs are the geometric features, first-order statistical features, texture features, and wavelet features corresponding to the ovarian cyst region x1 and the ovarian region x in step 3); the output is the risk prediction value of benign or malignant ovarian cysts;
[0014] The output also includes an uncertainty value for the risk prediction value of benign or malignant ovarian cysts.
[0015] Preferably, according to the present invention, the uncertainty value for the benign or malignant risk prediction value of the ovarian cyst ranges from [0, 1], and a threshold value P is selected between [0, 1] according to the actual application scenario:
[0016] When the uncertainty value of the benign or malignant risk prediction value of the ovarian cyst is greater than the threshold value P, it is prompted to re-identify the ovarian cyst according to step 1)-step 4). The higher the uncertainty value, the less reliable the prediction result of the model, so the result needs to be re-evaluated. Since the output range of the benign or malignant and uncertainty values are both between [0, 1], the high uncertainty threshold is set to P. If it is higher than this value, the prediction result needs to be re-evaluated. The user can also decide whether to use the benign or malignant risk prediction value of the ovarian cyst as a reference for medical judgment based on the uncertainty value of the benign or malignant risk prediction value of the ovarian cyst.
[0017] Preferably, according to the present invention, when the uncertainty value of the benign or malignant risk prediction value of the ovarian cyst is greater than the threshold value P, it is prompted to re-identify the ovarian cyst according to steps 2) to 4).
[0018] Preferably, according to the present invention, before step 4), a training model is used: the geometric features, first-order statistical features, texture features and wavelet features corresponding to the ovarian cyst region x1 and the ovarian region x are respectively input into the three cascaded multi-layer perceptrons MLP, and the output is a benign or malignant risk prediction value of the ovarian cyst;
[0019] The loss function of the training model is the cross entropy between the predicted value and the true value: when the loss function no longer decreases, the optimal training model is obtained.
[0020] Preferably, according to the present invention, the training model also outputs an uncertainty value for the benign or malignant risk prediction value of the ovarian cyst, and the uncertainty value is obtained by calculating the information entropy of the model prediction value.
[0021] Preferably, according to the present invention, in step 2), the ovarian cyst region x1 and the ovarian region x are manually depicted in the ultrasound image, and the non-cyst region x2 can be obtained by graphical calculation. The specific calculation process is:
[0022] The mask corresponding to the ovarian cyst region x1 is m1, and the mask m2 corresponding to the non-cyst region x2 is calculated as:
[0023] m2=1-m1, thus the area around the ovarian cyst is x2=x·m2.
[0024] In this embodiment, since the uncertainty value of the benign or malignant risk prediction value of ovarian cysts corresponds to a range between [0, 1], the threshold P of the uncertainty value is set to 0.1 in the early stage: when it is higher than the threshold P, the prediction result is re-evaluated; when the uncertainty value is lower than 0.1, there is no need to re-evaluate the prediction result and it can be used as a reference.
[0025] Furthermore, when the obtained benign or malignant risk prediction value of the ovarian cyst is greater than 0.5, it is judged that the ovarian cyst is likely to be malignant, otherwise it is judged to be a benign ovarian cyst.
[0026] A program product loaded with the above method comprises: the computer program product is tangibly stored on a non-transitory computer readable medium and comprises machine executable instructions, wherein the machine executable instructions are used to execute the above method.
[0027] The technical advantages of the present invention are:
[0028] The present invention discloses an ovarian cyst recognition method based on ultrasonic medical images, so as to achieve the purpose that the prior art cannot predict the benign or malignant risk of ovarian cysts. At the same time, it can also perform uncertainty analysis on the risk prediction results, and provide more objective recognition results for different application scenarios for user reference. BRIEF DESCRIPTION OF THE DRAWINGS
[0029] Figure 1 is a marked image of the region of interest extracted in the method of the present invention, wherein the dotted outline is marked by manual drawing;
[0030] Figure 2 is a flow chart of the method of the present invention; wherein the imaging omics feature corresponds to step 3), and the training model and the prediction model both correspond to the model described in step 4);
[0031] Figure 3 The imaging genomics feature in implementation example 1 of the present invention corresponds to step 3), and the prediction model corresponds to the model described in step 4). DETAILED DESCRIPTION
[0032] The following is a detailed description with reference to the embodiments and the accompanying drawings, but is not limited thereto.
[0033] Embodiment 1,
[0034] like Figure 1-3 shown.
[0035] An ovarian cyst recognition method based on ultrasonic medical images, comprising:
[0036] 1) Obtain an abdominal ultrasound image that includes the ovarian area on the affected side:
[0037] 2) Region of interest ROI extraction:
[0038] For the diseased ovarian region, extract the ovarian cyst region x1 and the ovarian region x in the ultrasound image;
[0039] In step 2), the ovarian cyst region x1 and the ovarian region x are manually depicted in the ultrasound image, and the non-cyst region x2 can be obtained by graphical calculation. The specific calculation process is:
[0040] The mask corresponding to the ovarian cyst region x1 is m1, and the mask m2 corresponding to the non-cyst region x2 is calculated as:
[0041] m2 = 1-m1, so the area around the ovarian cyst is x2 = x·m2;
[0042] 3) Radiomics feature calculation module:
[0043] Perform feature calculation on the region extracted in step 2): Use the radiomics feature calculation package pyradiomics in Python to calculate features. For each patient, multiple radiomics features can be obtained. The radiomics features that need to be calculated include the geometric features, first-order statistical features, texture features, and wavelet features of the ovarian cyst region x1 and the ovarian region x.
[0044] 4) Establish an ovarian cyst identification and prediction model:
[0045] It includes three cascaded multilayer perceptrons (MLPs); the inputs are the geometric features, first-order statistical features, texture features, and wavelet features corresponding to the ovarian cyst region x1 and the ovarian region x in step 3); the output is the risk prediction value of benign or malignant ovarian cysts;
[0046] The geometric features are descriptions of the two-dimensional size and shape of the ROI. The geometric features are independent of the gray intensity distribution in the ROI and are therefore only calculated on the manually drawn mask, including surface area (SURFA), volume (VOLUM), compactness 1 (COMPA1), compactness 2 (COMPA2), surface area to volume ratio (SU2VO), roundness (ROUND), and maximum diameter (MAXDA);
[0047] The first-order statistical features describe the distribution of voxel intensities in the image area defined by the mask through commonly used and basic metrics. The physical meaning of the first-order statistical features is described in the technical documentation of the pyradiomics package, including: maximum intensity value (MAX), minimum intensity value (MIN), mean intensity value (MEA), mean absolute deviation (MAD), median intensity value (MED), intensity value range (RAN), root mean square (RMS), standard deviation (STD), variance (VAR), uniformity (UNI), kurtosis (KUR), skewness (SKE), energy (ENE) and entropy (ENT);
[0048] The physical meaning of the texture features is explained in the pyradiomics package technical documentation:
[0049] a) Features based on gray-level co-occurrence matrix (GLCM): autocorrelation (AUTOC), contrast (CONTR), correlation 1 (CORRM), correlation 2 (CORRP), cluster prominence (CPROM), cluster shadow (CSHAD), cluster trend (CTEND), dissimilarity (DISSI), energy (ENERG), entropy (ENTRO), homogeneity 1 (HOMOG), homogeneity 2 (HOMOP), maximum probability (MAXPR), sum of squares (SOSVH), sum mean (SAVGH), total variance (SVARH), total entropy (SENTH), difference variance (DVARH), variance (VARIA), difference entropy (DENTH), related information measure 1 (INF1H), related information measure 2 (INF2H), inverse variance (INVAR), inverse difference normalization (INDNC) and inverse difference moment normalization (IDMNC).
[0050] b) Features based on gray-level run length matrix (GLRLM): short-term emphasis (SRE), long-term emphasis (LRE), gray-level non-uniformity (RLGLN), run length non-uniformity (RLN), run percentage (RP), low gray-level emphasis (LGRE), high gray-level emphasis (HGRE), short-term low gray-level emphasis (SRLGE), short-term high gray-level emphasis (SRHGE), long-term low gray-level emphasis (LRLGE), long-term high gray-level emphasis (LRHGE), gray-level variance (RLLGV) and run-length variance (RLV).
[0051] c) Features based on grayscale size zone matrix (GLSZM): small zone emphasis (SZE), large zone emphasis (LZE), grayscale non-uniformity (SZGLN), zone size non-uniformity (ZSN), zone percentage (ZP), low grayscale zone emphasis (LGZE), high grayscale zone emphasis (HGZE), small zone low grayscale emphasis (SZLGE), small zone high grayscale emphasis (SZHGE), large zone low grayscale emphasis (LZLGE), large zone high grayscale emphasis (LZHGE), grayscale variance (SZGLV), and zone size variance (ZSV).
[0052] d) Features based on the neighborhood grayscale difference matrix (NGTDM): roughness (COARS), contrast (CTRT), busyness (BUSS), complexity (CMPXY) and intensity (STRH);
[0053] The wavelet features, wavelet transform uses different filters to convert the original image into a version that focuses on information at different scales. In wavelet transform, high-pass filtering is used to emphasize local details in the original image, while low-pass filtering can refine the general shape of the entire image. Therefore, the wavelet transform can highlight information at different scales, thereby generating transformed image features that can be used to quantify tumor heterogeneity. The specific operation steps are as follows: a two-dimensional haar wavelet transform is applied to each mask area. The two-dimensional wavelet transform decomposes the original image space into eight decompositions, and the decomposition obtained has the same size as the original image; for each decomposition, 14 first-order statistical features and 56 texture features are calculated, thereby obtaining the corresponding 560 wavelet features.
[0054] The output also includes an uncertainty value for the risk prediction value of benign or malignant ovarian cysts.
[0055] The uncertainty value for the benign or malignant risk prediction value of ovarian cysts ranges from [0, 1], and a threshold value P is selected between [0, 1] according to the actual application scenario:
[0056] When the uncertainty value of the benign or malignant risk prediction value of the ovarian cyst is greater than the threshold value P, it is prompted to re-identify the ovarian cyst according to steps 1) to 4).
[0057] Embodiment 2,
[0058] The difference between the ovarian cyst identification method based on ultrasonic medical images as described in Example 1 is that when the uncertainty value of the benign or malignant risk prediction value of the ovarian cyst is greater than the threshold P, it is prompted to re-identify the ovarian cyst according to steps 2) to 4).
[0059] Embodiment 3,
[0060] In the method for identifying ovarian cysts based on ultrasonic medical images as described in Examples 1 and 2, before step 4), a training model is provided: the geometric features, first-order statistical features, texture features and wavelet features corresponding to the ovarian cyst region x1 and the ovarian region x are respectively input into the three cascaded multi-layer perceptrons MLP, and the output is a benign or malignant risk prediction value of the ovarian cyst;
[0061] The loss function of the training model is the cross entropy between the predicted value and the true value: when the loss function no longer decreases, the optimal training model is obtained.
[0062] The training model also outputs an uncertainty value for the benign or malignant risk prediction value of the ovarian cyst, and the uncertainty value is obtained by calculating the information entropy of the model prediction value.
[0063] In this embodiment, since the uncertainty value of the benign or malignant risk prediction value of ovarian cysts corresponds to a range between [0, 1], the threshold P of the uncertainty value is set to 0.1 in the early stage: when it is higher than the threshold P, the prediction result is re-evaluated; when the uncertainty value is lower than 0.1, there is no need to re-evaluate the prediction result and it can be used as a reference.
[0064] Furthermore, when the obtained benign or malignant risk prediction value of the ovarian cyst is greater than 0.5, it is judged that the ovarian cyst is likely to be malignant, otherwise it is judged to be a benign ovarian cyst.
[0065] Embodiment 4,
[0066] A program product loaded with the method described in Examples 1-3, comprising: the computer program product is tangibly stored on a non-transitory computer-readable medium and includes machine executable instructions, and the machine executable instructions are used to execute the above-mentioned method.
Claims
1. A method for identifying ovarian cysts based on ultrasonic medical images, characterized in that: include: 1) Obtain an abdominal ultrasound image that includes the ovarian area on the affected side: 2) Region of interest ROI extraction: For the diseased ovarian region, firstly, the ovarian cyst region x1 and the ovarian region x in the ultrasound image are extracted; then, the ovarian cyst region x1 and the ovarian region x are manually depicted in the ultrasound image, and the non-cyst region x2 is obtained by graphic calculation; 3) Radiomics feature calculation module: Feature calculation is performed on the region extracted in step 2): the imaging omics feature calculation package pyradiomics in Python is used to perform feature calculation. The imaging omics features that need to be calculated include: geometric features, first-order statistical features, texture features, and wavelet features of the ovarian cyst region x1 and the non-cyst region x2; 4) Establish an ovarian cyst identification and prediction model: It includes three cascaded multi-layer perceptrons (MLPs); the inputs are the geometric features, first-order statistical features, texture features, and wavelet features corresponding to the ovarian cyst region x1 and the non-cyst region x2 in step 3); The output is the risk prediction value of benign or malignant ovarian cysts; The output also includes an uncertainty value for the benign or malignant risk prediction value of the ovarian cyst. When the uncertainty value for the benign or malignant risk prediction value of the ovarian cyst is greater than a threshold value P, it is prompted to re-identify the ovarian cyst according to steps 1) to 4).
2. The method for identifying ovarian cysts based on ultrasonic medical images according to claim 1, characterized in that: The uncertainty value for the benign or malignant risk prediction value of the ovarian cyst ranges from [0, 1], and a threshold value P is selected between [0, 1] according to an actual application scenario.
3. The method for identifying ovarian cysts based on ultrasonic medical images according to claim 1, characterized in that: When the uncertainty value of the benign or malignant risk prediction value of the ovarian cyst is greater than the threshold value P, it is prompted to re-identify the ovarian cyst according to steps 2) to 4).
4. The method for identifying ovarian cysts based on ultrasonic medical images according to claim 1, characterized in that: Before step 4), a training model is provided: the geometric features, first-order statistical features, texture features and wavelet features corresponding to the ovarian cyst region x1 and the non-cyst region x2 are respectively input into the three cascaded multi-layer perceptrons MLP, and the output is a benign or malignant risk prediction value of the ovarian cyst; The loss function of the training model is the cross entropy between the predicted value and the true value: when the loss function no longer decreases, the optimal training model is obtained.
5. The method for identifying ovarian cysts based on ultrasonic medical images according to claim 4, characterized in that: The training model also outputs an uncertainty value for the benign or malignant risk prediction value of the ovarian cyst, and the uncertainty value is obtained by calculating the information entropy of the model prediction value.
6. The method for identifying ovarian cysts based on ultrasonic medical images according to claim 1, characterized in that: In step 2), the ovarian cyst region x1 and the ovarian region x are manually depicted in the ultrasound image, and the non-cyst region x2 can be obtained by graphical calculation. The specific calculation process is: The mask corresponding to the ovarian cyst region x1 is m1, and the mask m2 corresponding to the non-cyst region x2 is calculated as: m2=1-m1, thus the area around the ovarian cyst is x2=x·m2.
7. A program product loaded with the method according to any one of claims 1 to 6, characterized in that: include: The program product is tangibly stored on a non-transitory computer-readable medium and includes machine-executable instructions for performing the above-described method.
