A tumor detection model training set expansion method, system, device and medium
Patent Information
- Application Number
- CN202410286749.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-03-13
- Publication Date
- 2026-10-27
AI Technical Summary
[0005]本发明的目的在于针对上述现有技术的不足，提供一种肿瘤检测模型训练集扩充方法、系统、设备与介质，以解决现有技术中扩充的数据量和对肿瘤定位的提升效果都很有限的问题
[0035]本发明通过获取每个肿瘤图像在肿瘤超声图像集中随机出现的概率值，且判断概率值与合成阈值的大小关系，并在概率值大于阈值时，将正常图像缩放至肿瘤图像大小，设定边缘融合像素宽度参数，将肿瘤边缘外侧部分的融合权重值进行更新，能有效提高肿瘤超声图像肿瘤定位准确率，生成最终肿瘤的合成图像集，对于超声检测模型的临床应用具有重要的推动作用，且合成图像在肿瘤区域和非肿瘤区域都来自于现有的真实图像，肿瘤边缘部分经过融合时权重值的平滑处理后没有造成明显的合成痕迹特征，能够起到真实训练数据集的扩充效果。
Smart Images
Figure CN122990368A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of tumor model training technology, and in particular to a method, system, device and medium for expanding the training set of a tumor detection model. Background Technology
[0002] Deep neural networks have achieved great success in image target localization applications, but their performance in the field of medical ultrasound imaging is far inferior to that in the field of natural images, failing to meet the requirements of clinical applications.
[0003] This is partly because the quality of medical ultrasound images is limited by the instruments and acquisition equipment, resulting in a much lower quality compared to natural images. Another reason is the scarcity of labeled medical ultrasound image datasets compared to natural datasets. Expanding the available labeled medical ultrasound image datasets is crucial for improving the performance of deep neural networks in tumor localization within medical ultrasound images.
[0004] Existing methods for expanding medical ultrasound image annotation datasets typically involve flipping, distorting, or scaling the original images at different angles. However, the amount of data expanded and the improvement in tumor localization are both limited. Summary of the Invention
[0005] The purpose of this invention is to address the shortcomings of the prior art by providing a method, system, device, and medium for expanding the training set of a tumor detection model, thereby solving the problem that the amount of expanded data and the improvement effect on tumor localization are both limited in the prior art.
[0006] This invention specifically provides the following technical solution: a method for expanding the training set of a tumor detection model, comprising the following steps:
[0007] Acquire an existing set of tumor ultrasound images, which includes tumor images and normal images, and the tumor images are marked with the location of the tumor region.
[0008] Set a synthesis threshold for tumor images and normal images, obtain the probability value of selecting a tumor image from the tumor ultrasound image set, and determine the relationship between the probability value and the synthesis threshold.
[0009] If the probability value is greater than the synthesis threshold, then a normal image is selected from the tumor ultrasound image set, and the normal image is scaled to the same size as the tumor image.
[0010] The pixels at the location of the tumor region are compared with the pixels in the scaled normal image. If the comparison results are consistent, the weight value of the pixel location is obtained.
[0011] Set the edge fusion pixel width parameter, and use the edge fusion pixel width parameter to update the weight value of the outer edge of the tumor region;
[0012] The product of the pixels in the tumor region and the weight value, and the product of the pixels in the normal image and the updated weight value are summed to generate a synthetic tumor image. Multiple synthetic tumor images generated at multiple times are used as an expanded training set for the tumor detection model.
[0013] Preferably, updating the weight value of the outer edge of the tumor region using the edge fusion pixel width parameter includes the following steps:
[0014] If pixel [row,col] is the left edge of the tumor, then the weight value decreases by a pixels to the left until it reaches 0, that is, weight[row,col-i]=1-i / a, where i∈[1,a];
[0015] If pixel [row,col] is the right edge of the tumor, then the weight value decreases by a pixels to the right until it reaches 0, that is, weight[row,col+i]=1-i / a, where i∈[1,a];
[0016] If pixel [row,col] is the upper edge of the tumor, then the weight value decreases by a pixels upwards until it reaches 0, that is, weight[row-i,col]=1-i / a, where i∈[1,a];
[0017] If pixel [row,col] is the lower edge of the tumor, then the weight value decreases by a pixels downwards until it reaches 0, that is, weight[row+i,col]=1-i / a, where i∈[1,a].
[0018] Preferably, the synthetic image set for generating tumors is specifically expressed as follows:
[0019] image new =image orgin *weight+image normal *(1-weight)
[0020] Among them, image new For synthesized images, image orgin For a tumor image, weight represents the values of the weight array, and image... normal This is a normal image.
[0021] Preferably, the tumor marker data corresponding to the synthesized image is a mask. new =mask orgin .
[0022] Preferably, after acquiring the existing tumor ultrasound image set, the tumor ultrasound image set is randomly divided into a training set and a test set according to a certain proportion, wherein the training set includes existing real tumor ultrasound images and tumor location annotations, normal ultrasound images without tumors and location annotations, and synthetic tumor ultrasound images and location annotations.
[0023] Preferably, when generating the tumor composite image, the pixels near the periphery of the tumor region in the tumor composite image are fused from the pixel information of the tumor image and the normal image in a certain proportion, wherein the proportion changes linearly according to the distance of the tumor region from the edge of the normal image.
[0024] Preferably, the generated multiple synthetic tumor images and tumor image samples are divided into a new training set for the tumor detection model at a 1:1 ratio.
[0025] This invention provides a system for expanding the training set of a tumor detection model, comprising:
[0026] The data acquisition module is used to acquire an existing set of tumor ultrasound images, which includes tumor images and normal images, and the tumor images are marked with the location of the tumor region.
[0027] The probability acquisition module is used to set the synthesis threshold of tumor images and normal images, acquire the probability value of selecting a tumor image from the tumor ultrasound image set, and determine the relationship between the probability value and the synthesis threshold.
[0028] A scaling module is used to select a normal image from the tumor ultrasound image set and scale the normal image to the same size as the tumor image if the probability value is greater than the synthesis threshold.
[0029] The weight acquisition module is used to compare the pixels at the location of the tumor region with the pixels in the scaled normal image, and when the comparison results are consistent, to acquire the weight value of the pixel location.
[0030] The outer weight update module is used to set the edge fusion pixel width parameter and update the weight value of the outer edge of the tumor region using the edge fusion pixel width parameter;
[0031] The synthesis module is used to perform a summation operation on the product of the pixels and weight values of the tumor region and the product of the pixels and updated weight values of the normal image to generate a synthesized tumor image, and to use the multiple synthesized tumor images generated multiple times as an expanded training set for the tumor detection model.
[0032] The present invention provides a computer device, including a memory and a processor, wherein the memory stores a program, and when the program is executed by the processor, the processor performs the steps of the method for expanding the training set of a tumor detection model.
[0033] The present invention provides a storage medium storing a computer program thereon, characterized in that the computer program, when executed by a processor, implements the steps of the method for expanding the training set of a tumor detection model.
[0034] Compared with the prior art, the present invention has the following significant advantages:
[0035] This invention obtains the probability value of each tumor image randomly appearing in a tumor ultrasound image set, determines the relationship between the probability value and the synthesis threshold, and scales the normal image to the size of the tumor image when the probability value is greater than the threshold. It sets the edge fusion pixel width parameter and updates the fusion weight value of the outer part of the tumor edge. This can effectively improve the accuracy of tumor localization in tumor ultrasound images and generate a final synthetic image set of tumors. This has an important role in promoting the clinical application of ultrasound detection models. Moreover, the synthetic images in both tumor and non-tumor areas are derived from existing real images. After the tumor edge part is smoothed by the weight value during fusion, there are no obvious synthetic traces. It can also expand the real training dataset. Attached Figure Description
[0036] Figure 1 These are examples of synthesized tumor images; where (a) is a tumor image, (b) is an image showing the location of the tumor, (c) is a normal image without a tumor, and (d) is a synthesized tumor image.
[0037] Figure 2 This is a dataset image used for training a breast tumor localization and detection model; where (a) is the existing tumor image set, (b) is the location information of the existing tumor image set, (c) is the synthetic tumor image set, (d) is the location information of the synthetic tumor image set, (e) is the normal image set without tumor, and (f) is the tumor location information of the normal image set.
[0038] Figure 3 This is the overall flowchart. Detailed Implementation
[0039] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.
[0040] like Figure 3 As shown, an embodiment of this application provides a method for expanding the training set of a tumor detection model, including the following steps:
[0041] Step S1: Obtain the existing tumor ultrasound image set D(I,M), which includes tumor images and normal images, and the tumor images are marked with the location of the tumor region.
[0042] The tumor ultrasound image set D(I,M) is divided into a training set D according to a certain ratio. train (I train M train ) and test set D test (I test M test ).
[0043] In this step, the tumor ultrasound image set includes benign tumors, malignant tumors, and normal images. The normal images do not contain tumor regions, and the images in the tumor ultrasound image set have benign / malignant markers and tumor region markers.
[0044] The training set includes existing real tumor ultrasound images and tumor location annotations, normal ultrasound images without tumors and location annotations, and synthetic tumor ultrasound images and location annotations.
[0045] Step S2: Set the synthesis threshold for tumor images and normal images. For each tumor image (sample) in the training set... orgin ,mask orgin )∈(I train M train The algorithm obtains the probability value of selecting a tumor image from a set of tumor ultrasound images and determines the relationship between the probability value and the synthesis threshold.
[0046] Step S3: If the probability value is greater than the threshold K, then select a normal image from the normal image set N in the tumor ultrasound image set. normal normal image normal Zoom to the image of the tumor orgin The same size is used to initialize the composite weight array, which is an array of all zeros, with the same size as the Image.
[0047] Step S4: Compare the pixels at the tumor region with the pixels in the scaled normal image. If the comparison results are consistent, obtain the weight value of the pixel location.
[0048] Specifically:
[0049] Tumor marker data mask in scanned tumor images orgin The values at each pixel in a normal image are determined by the following: when the value is true, the weight array (weight value) weight at that pixel position is updated to 1, i.e., weight[i,j]=1; if the value is true, the weight array (weight value) weight[i,j]=true.
[0050] Step S5: Set the edge fusion pixel width parameter, and use the edge fusion pixel width parameter to update the weight value of the outer edge of the tumor region.
[0051] In this step, the edge fusion pixel width parameter is set, and the fusion weight value of the outer part of the tumor edge is updated. When generating the final synthetic image set of the tumor, the pixels near the periphery of the tumor area in the synthetic breast tumor ultrasound image are fused with the pixel information of the original breast tumor ultrasound image and the normal ultrasound image according to the ratio, which changes linearly according to the distance from the edge.
[0052] In this step, the weight values on the outer edge of the tumor region are updated using the edge fusion pixel width parameter, including the following steps:
[0053] If pixel [row,col] is the left edge of the tumor, then the weight value decreases by a pixels to the left until it reaches 0, that is, weight[row,col-i]=1-i / a, where i∈[1,a].
[0054] If pixel [row,col] is the right edge of the tumor, then the weight value decreases by a pixels to the right until it reaches 0, that is, weight[row,col+i]=1-i / a, where i∈[1,a].
[0055] If pixel [row,col] is the upper edge of the tumor, then the weight value decreases by a pixels upwards until it reaches 0, that is, weight[row-i,col]=1-i / a, where i∈[1,a].
[0056] If pixel [row,col] is the lower edge of the tumor, then the weight value decreases by a pixels downwards until it reaches 0, that is, weight[row+i,col]=1-i / a, where i∈[1,a].
[0057] Step S6: Perform a summation operation on the product of the pixels in the tumor region and the weight value, and the product of the pixels in the normal image and the updated weight value, to generate a synthetic tumor image. Use the multiple synthetic tumor images generated multiple times as an expanded training set for the tumor detection model.
[0058] The final synthetic image set of the tumor is generated using the following expression:
[0059] image new =image orgin *weight+image normal *(1-weight)
[0060] Among them, imagenew For synthesized images, image orgin For a tumor image, weight represents the values of the weight array, and image... normal This is a normal image.
[0061] The tumor marker data corresponding to the synthesized image is a mask. new =mask orgin .
[0062] Among them, see Figure 1 (a)-(d) are example images of the tumor before and after tumor synthesis.
[0063] See Figure 2 (a)-(f) use the original training dataset, the synthetic dataset, and the normal dataset together as training data for the breast ultrasound image tumor detection model.
[0064] The generated multiple synthetic tumor images and the tumor image samples are divided into a 1:1 ratio as the training set for the new tumor detection model. That is, for each original tumor ultrasound image used in training, a normal image is randomly selected to generate the synthetic tumor ultrasound image.
[0065] Case Study:
[0066] A method for expanding the training data of a breast ultrasonography tumor detection model based on the integration of tumor images and normal images (EDITIN) is presented below, using the publicly available dataset BUSI as an example (including 487 benign breast ultrasound images and tumor localization information, 210 malignant breast ultrasound images and tumor localization information, and 133 normal breast ultrasound images). The specific steps are as follows:
[0067] 1) The 487 benign images and 210 malignant images were randomly divided into training and test sets at a ratio of 70% and 30%, respectively, resulting in a total of 488 training images and 209 test images.
[0068] 2) Set the threshold K to 0 and the edge fusion pixel width parameter a to 20. That is, for each sample in the training set, an image is randomly selected from the normal image for synthesis. The tumor region pixels of the synthesized image come from the original tumor image. The pixels within 20 pixels outside the image edge are generated by superimposing the pixel values of the original tumor image and the normal image according to the weight settings. The remaining pixels come from the normal image. The tumor localization information of the original tumor image is used as the tumor localization information of the synthesized image.
[0069] 3) The synthesized 488 tumor images and localization information, the original 488 tumor images and localization information, and 133 normal images were used together as the training set for the breast ultrasound image tumor detection model to train the breast ultrasound image tumor detection model.
[0070] 4) The trained breast ultrasound image tumor detection model was tested on 209 test set images.
[0071] Experimental Analysis:
[0072] To verify the effectiveness of the method of the present invention, the dataset expansion method proposed in this invention is compared with the method that only uses the original training set, mainly in the comparison of IOU and F1 index of 209 test set images.
[0073] In this invention, the breast ultrasound image tumor detection model adopts CMUneXt-S.
[0074] Table 1 Comparison of accuracy between the two methods
[0075] Expanding the dataset 71.48％ 77.25％ Original dataset 68.84％ 77.33％
[0076] Table 1 shows that the dataset expansion method proposed in this invention significantly improves the tumor localization accuracy of breast tumor ultrasound images compared to the original training dataset. This is due to two main reasons: (1) the synthesized images are derived from real images in both tumor and non-tumor regions, and the smoothed tumor edges do not exhibit obvious synthetic traces, thus effectively expanding the real training dataset; (2) the normal images in the expanded dataset also enhance the network's ability to distinguish between tumor and non-tumor regions, further expanding the dataset. In conclusion, this invention effectively improves the tumor localization accuracy of breast tumor ultrasound images, significantly promoting the clinical application of ultrasound detection models.
[0077] Based on the above methods and statements, the present invention provides a tumor detection model training set expansion system, comprising: a data acquisition module, a probability acquisition module, a scaling module, a weight acquisition module, an outer weight update module, and a synthesis module.
[0078] The system comprises the following modules: a data acquisition module for acquiring an existing set of tumor ultrasound images, including tumor and normal images, with tumor regions marked on the tumor images; a probability acquisition module for setting a synthesis threshold for tumor and normal images, obtaining the probability value for selecting a tumor image from the set, and determining the relationship between the probability value and the synthesis threshold; a scaling module for selecting a normal image from the set and scaling it to the same size as the tumor image if the probability value is greater than the synthesis threshold; a weight acquisition module for comparing pixels at the tumor region with pixels in the scaled normal image, obtaining the weight value for that pixel when the comparison is consistent; an outer weight update module for setting an edge fusion pixel width parameter and updating the weight value of the outer edge of the tumor region using this parameter; and a synthesis module for summing the product of the tumor region's pixels and weight value, and the product of the normal image's pixels and the updated weight value, generating a synthesized tumor image, and using multiple generated synthesized tumor images as an expanded training set for the tumor detection model.
[0079] The present invention also provides a computer device, including a memory and a processor, wherein the memory stores a program, and when the program is executed by the processor, the processor performs the steps of a method for expanding a training set for a tumor detection model.
[0080] The present invention also provides a storage medium storing a computer program thereon, wherein the computer program, when executed by a processor, implements the steps of a method for expanding a training set for a tumor detection model.
[0081] The above description, in conjunction with specific preferred embodiments, provides a more detailed explanation of the present invention. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such deductions or substitutions should be considered to fall within the scope of protection of the present invention.
Claims
1. A method for expanding the training set of a tumor detection model, characterized in that, Includes the following steps: Acquire an existing set of tumor ultrasound images, which includes tumor images and normal images, and the tumor images are marked with the location of the tumor region. Set a synthesis threshold for tumor images and normal images, obtain the probability value of selecting a tumor image from the tumor ultrasound image set, and determine the relationship between the probability value and the synthesis threshold. If the probability value is greater than the synthesis threshold, then a normal image is selected from the tumor ultrasound image set, and the normal image is scaled to the same size as the tumor image. The pixels at the location of the tumor region are compared with the pixels in the scaled normal image. If the comparison results are consistent, the weight value of the pixel location is obtained. Set the edge fusion pixel width parameter, and use the edge fusion pixel width parameter to update the weight value of the outer edge of the tumor region; The product of the pixels in the tumor region and the weight value, and the product of the pixels in the normal image and the updated weight value are summed to generate a synthetic tumor image. Multiple synthetic tumor images generated at multiple times are used as an expanded training set for the tumor detection model.
2. The method for expanding the training set of a tumor detection model as described in claim 1, characterized in that, The step of updating the weight value of the outer edge of the tumor region using the edge fusion pixel width parameter includes the following steps: If pixel [row,col] is the left edge of the tumor, then the weight value decreases by a pixels to the left until it reaches 0, that is, weight[row,col-i]=1-i / a, where i∈[1,a]; If pixel [row,col] is the right edge of the tumor, then the weight value decreases by a pixels to the right until it reaches 0, that is, weight[row,col+i]=1-i / a, where i∈[1,a]; If pixel [row,col] is the upper edge of the tumor, then the weight value decreases by a pixels upwards until it reaches 0, that is, weight[row-i,col]=1-i / a, where i∈[1,a]; If pixel [row,col] is the lower edge of the tumor, then the weight value decreases by a pixels downwards until it reaches 0, that is, weight[row+i,col]=1-i / a, where i∈[1,a].
3. The method for expanding the training set of a tumor detection model as described in claim 2, characterized in that, The specific expression for generating the synthetic image set of tumors is as follows: image new =image orgin *weight+image normal *(1-weight) Among them, image new For synthesized images, image orgin For a tumor image, weight represents the values of the weight array, and image... normal This is a normal image.
4. The method for expanding the training set of a tumor detection model as described in claim 3, characterized in that, The tumor marker data corresponding to the synthesized image is a mask. new =mask orgin .
5. The method for expanding the training set of a tumor detection model as described in claim 1, characterized in that, After acquiring the existing tumor ultrasound image set, the tumor ultrasound image set is randomly divided into a training set and a test set according to a certain ratio. The training set includes existing real tumor ultrasound images and tumor location annotations, normal ultrasound images without tumors and location annotations, and synthetic tumor ultrasound images and location annotations.
6. The method for expanding the training set of a tumor detection model as described in claim 1, characterized in that, When generating the tumor composite image, the pixels near the periphery of the tumor region in the tumor composite image are fused from the pixel information of the tumor image and the normal image in a certain proportion, wherein the proportion changes linearly according to the distance of the tumor region from the edge of the normal image.
7. The method for expanding the training set of a tumor detection model as described in claim 1, characterized in that, The generated multiple synthetic tumor images and the tumor image samples are divided into a new training set for the tumor detection model at a 1:1 ratio.
8. A system for expanding the training set of a tumor detection model, characterized in that, include: The data acquisition module is used to acquire an existing set of tumor ultrasound images, which includes tumor images and normal images, and the tumor images are marked with the location of the tumor region. The probability acquisition module is used to set the synthesis threshold of tumor images and normal images, acquire the probability value of selecting a tumor image from the tumor ultrasound image set, and determine the relationship between the probability value and the synthesis threshold. A scaling module is used to select a normal image from the tumor ultrasound image set and scale the normal image to the same size as the tumor image if the probability value is greater than the synthesis threshold. The weight acquisition module is used to compare the pixels at the location of the tumor region with the pixels in the scaled normal image, and when the comparison results are consistent, to acquire the weight value of the pixel location. The outer weight update module is used to set the edge fusion pixel width parameter and update the weight value of the outer edge of the tumor region using the edge fusion pixel width parameter; The synthesis module is used to perform a summation operation on the product of the pixels and weight values of the tumor region and the product of the pixels and updated weight values of the normal image to generate a synthesized tumor image, and to use the multiple synthesized tumor images generated multiple times as an expanded training set for the tumor detection model.
9. A computer device, characterized in that, The device includes a memory and a processor, wherein the memory stores a program that, when executed by the processor, causes the processor to perform the steps of the tumor detection model training set expansion method as described in any one of claims 1 to 7.
10. A storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the tumor detection model training set expansion method according to any one of claims 1 to 7.