A method for identifying benign and malignant tumors based on imaging data and deep learning
Through image data and deep learning technology, medical images are processed, organ contours are marked, tumor areas and invasion factors are determined, and features are extracted using CNN models to calculate tumor area, which solves the problems of large errors and low efficiency in traditional manual identification, and improves tumor goodness Accuracy and efficiency of malignant identification.
Patent Information
- Application Number
- CN202411025401.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-30
- Publication Date
- 2024-11-26
- Estimated Expiration
- 2044-07-30
AI Technical Summary
Traditional artificial recognition methods have problems of large errors and low efficiency in tumor benign and malignant identification, mainly due to inaccurate reliance on doctor experience and image feature recognition.
Using an imaging data and deep learning method, normalized processing and modal registration are performed by acquiring and processing medical images, labeling organ profiles, determining the tumor pending areas and invasion factors, using CNN models to extract tumor characteristics, and calculate tumor area and invade area, generating the overall impact area to assist doctors in identifying it.
It improves the accuracy and efficiency of tumor benign and malignant identification, reduces artificial errors, ensures the accuracy of tumor area and the reliability of feature extraction of CNN models.
Smart Images

Figure CN118552541B_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the field of image processing technology, and more specifically, to a method for identifying benign and malignant tumors based on image data and deep learning. Background Art
[0002] The technology of identifying benign and malignant tumors is often used by doctors to identify tumors in medical images such as CT, MRI, and PET. It is a traditional method in medical image recognition. This method relies on the experience and expertise of professionals such as radiologists. Radiologists determine the benign and malignant nature of tumors by observing image features in medical images, such as the size, shape, edge, density, signal strength, etc. of the tumor.
[0003] Traditional manual identification methods are highly subjective, and different doctors may have different results for the same impact because the naked eye cannot accurately identify the tumor area and the area of the invasion zone. Furthermore, the method relies on professional knowledge, which leads to large errors in the identification of benign and malignant tumors and low identification efficiency.
[0004] Therefore, there is an urgent need to provide a solution for identifying benign and malignant tumors based on imaging data and deep learning, which can provide accurate tumor area and area of invasion to assist doctors in identifying benign and malignant tumors, and solve the problem of large errors and low efficiency in manual identification of benign and malignant tumors. Summary of the invention
[0005] The present invention provides a method for identifying benign and malignant tumors based on image data and deep learning, which is used to solve the technical problems of large errors and low efficiency in the manual identification of benign and malignant tumors in the prior art. The method comprises:
[0006] Obtain the basic information and medical images of the patient to be identified, normalize multiple medical images, and perform registration of the processed multiple medical images in different modalities and at different times to generate time series images;
[0007] The original contours of each organ are marked on the time series images based on the basic information of the patient to be identified, and multiple tumor pending areas and the invasion factors corresponding to each tumor pending area are determined in the original contours of each organ based on the PET scan images in the medical images, and the tumor pending areas are expanded based on the invasion factors to generate multiple tumor determination areas;
[0008] Separate multiple tumor determination areas from the time series images and generate time series tumor determination areas. Use the CNN model to extract tumor features in the time series tumor determination areas.
[0009] The tumor area and invasion area are calculated based on tumor characteristics, and the overall impact area is determined by the tumor area and invasion area. The overall impact area is used to help doctors identify whether the tumor is benign or malignant.
[0010] In some embodiments of the present application, normalization processing is performed on multiple medical images, including:
[0011] Get the signal-to-noise ratio and contrast-to-noise ratio of each medical image, read the pixel values of all pixels in each medical image, traverse all pixels to determine the maximum pixel value g1 and the minimum pixel value g2 in each medical image, and define is the light-dark ratio;
[0012] The quality index of the medical image is obtained by weighted summing the signal-to-noise ratio, the contrast-to-noise ratio and the brightness-to-darkness ratio, the medical images are sorted according to the size of the quality index, each medical image is used as a pre-reference image in turn, the resolution and size of other medical images are converted to the pre-reference image, and the peak signal-to-noise ratio of each other medical image is determined to measure the image difference before and after the conversion of each other medical image, the conversion effect of the overall medical image is determined by the peak signal-to-noise ratio of each other medical image, a pre-reference image is selected as a reference image according to the conversion effect and the quality index, and the resolution and size of other medical images are converted to the reference image;
[0013] Traverse all pixels of all medical images, determine the maximum pixel value gmax and the minimum pixel value gmin among all pixels, adjust the standard zoom scale according to the difference between the pixel range of gmin-gmax and the standard pixel range, and obtain the actual zoom scale;
[0014] All medical images are normalized to the same resolution, size, and grayscale range using reference image resolution and size and actual scaling.
[0015] In some embodiments of the present application, the processed multiple medical images are registered in different modalities and at different times to generate time series images, including:
[0016] The mutual information between multiple medical images in the same period is calculated and optimized through optimization algorithms. At the same time, the image transformation parameters will be adjusted. The optimal image transformation parameters are determined through multiple iterations to maximize the mutual information, so that multiple medical images in the same period are registered to the same space.
[0017] The registered medical images are arranged in chronological order to generate time series images.
[0018] In some embodiments of the present application, before determining multiple tumor regions to be determined and the invasion factor corresponding to each tumor region to be determined in the original contour of each organ according to the PET scan image in the medical image, the method further includes:
[0019] According to the basic information of the patient to be identified, the closest person is found in the person database, the normal PET scan image of the person is retrieved, and the pixel values of all pixels within the original contour of each organ and the pixel values of all pixels except the original contour of the organ are determined according to the normal PET scan image, so as to construct the standard pixel range of the inner contour-outer contour within the original contour of the organ and the standard pixel range of other areas, so as to simulate the pixel conditions within the original contour of each organ and the pixel conditions of other areas except the organs under the healthy condition of the patient to be identified;
[0020] The pixels of the inner contour are pixels completely located within the original contour of the organ, and the pixels of the outer contour are pixels partially located within the original contour of the organ and partially beyond the original contour of the organ.
[0021] In some embodiments of the present application, multiple tumor regions to be determined and invasion factors corresponding to each tumor region to be determined are determined in the original contours of each organ according to the PET scan image in the medical image, including:
[0022] Determine the actual pixel range of the inner contour-outer contour within the original contour of each organ and the actual pixel range of other regions according to the PET scan image of the patient to be identified, completely split the other regions into several sub-regions, and determine the actual pixel range of each sub-region;
[0023] The tumor region to be determined is determined by comparing the actual pixel range of the inner contour-outer contour in the original contour of each organ with the standard pixel range, and the original contour of the corresponding organ where only the actual pixel range of the inner contour exceeds the standard pixel range is used as the first tumor region to be determined, the original contour of the corresponding organ where only the actual pixel range of the outer contour exceeds the standard pixel range is used as the second tumor region to be determined, and the original contour of the corresponding organ where both the actual pixel range of the inner contour and the actual pixel range of the outer contour exceed the standard pixel range is used as the third tumor region to be determined;
[0024] The tumor region to be determined is determined by comparing the actual pixel range of each sub-region with the standard pixel range, the sub-region whose actual pixel range exceeds the standard pixel range is merged into the original contour of the closest organ, and the original contour of the organ after the sub-region is merged is used as the fourth tumor region to be determined, and the first tumor region to be determined, the second tumor region to be determined, the third tumor region to be determined, and the fourth tumor region to be determined are all categories of tumor regions to be determined;
[0025] Determine the invasion factor according to the category of the tumor area to be determined, the actual pixel range and the standard pixel range; ;
[0026] in, As an invasion factor, is the maximum value among multiple invasion distances corresponding to all tumor types mapped to the category of the tumor region to be determined, and They are the pixel conversion coefficients of the inner contour-outer contour and the pixel conversion coefficients of other areas, and The pixels corresponding to the inner contour and the outer contour exceed the weight respectively. , , The actual pixel range of the inner contour, outer contour, and other areas exceeds the standard pixel range. When the actual pixel range does not exceed the standard pixel range, the corresponding or or is 0.
[0027] In some embodiments of the present application, the tumor area to be determined is expanded according to the invasion factor to generate multiple tumor determination areas, including:
[0028] For the expansion of the first tumor region to be determined, the second tumor region to be determined, and the third tumor region to be determined, the center point within the original outline of the organ is used as the center of the circle and the invasion factor is used as the radius to expand outward, thereby forming an invasion circular area, and the invasion circular area is used as the tumor determination area;
[0029] For the expansion of the fourth tumor area to be determined, the center point of other areas in the fourth tumor area to be determined is used as the center of the circle and the invasion factor is used as the radius to expand outward to form an invasion circular area, and the union area of the invasion circular area and the fourth tumor area to be determined is used as the tumor determination area.
[0030] In some embodiments of the present application, the tumor area and invasion area are calculated based on tumor characteristics, including:
[0031] Tumor features are arranged in chronological order, and different weights are assigned to tumor features according to the time sequence. The later the tumor feature is, the greater the weight is, so as to determine the tumor area and invasion area.
[0032] In some embodiments of the present application, the overall affected area is determined by the tumor area and the invasion area, including:
[0033] ;
[0034] Where S is the overall impact area, is the tumor area, is the invasion area, exp is the exponential function, and k is a preset constant.
[0035] By applying the above technical solutions, the basic information and medical images of the patient to be identified are obtained, multiple medical images are normalized, and the processed multiple medical images are registered in different modalities and at different times to generate time series images; the original contours of each organ are marked on the time series image by the basic information of the patient to be identified, and multiple tumor pending areas and the invasion factors corresponding to each tumor pending area are determined in the original contours of each organ according to the PET scan image in the medical image, and the tumor pending area is expanded according to the invasion factor to generate multiple tumor determination areas; multiple tumor determination areas are separated from the image on the time series image, and a time series tumor determination area is generated, and the tumor features in the time series tumor determination area are extracted by the CNN model; the tumor area and the invasion area are calculated based on the tumor features, and the overall impact area is determined by the tumor area and the invasion area, and the overall impact area is used to help doctors realize the identification of benign and malignant tumors. This application facilitates the comparison of subsequent medical images and the image adaptability of the input model by normalizing multiple medical images. According to the PET scan images in medical images, multiple tumor areas and the corresponding invasion factors of each tumor area are determined in the original contours of each organ, ensuring the accuracy of the tumor area range, thereby improving the reliability of the CNN model to extract features. Finally, the overall impact area is used to help doctors make references, thereby improving the accuracy of tumor benign and malignant judgments, reducing human errors, and improving recognition efficiency. BRIEF DESCRIPTION OF THE DRAWINGS
[0036] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For those skilled in the art, other drawings can be obtained based on these drawings without creative work.
[0037] Figure 1 A schematic diagram of a process of identifying benign and malignant tumors based on image data and deep learning proposed in an embodiment of the present invention is shown;
[0038] Figure 2 A schematic diagram of a process for registering multiple medical images in different modalities according to an embodiment of the present invention is shown. DETAILED DESCRIPTION
[0039] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.
[0040] The present application embodiment provides a method for identifying benign and malignant tumors based on image data and deep learning. Figure 1 As shown, the method comprises the following steps:
[0041] Step S101, basic information and medical images of a patient to be identified are obtained, multiple medical images are normalized, and the processed multiple medical images are registered in different modalities and at different times to generate time series images.
[0042] In this embodiment, the basic information includes height, weight, body fat percentage, age, race, etc., and the medical images include CT, MRI, X-rays (which may contain images at multiple time nodes), etc. The multiple medical images are normalized (resolution, size, and grayscale range) to facilitate the subsequent registration of multiple medical images in different modalities and at different times. Registration generally refers to aligning medical images from different time points, different devices, or different imaging modalities into the same coordinate system. This process is not simply merging multiple images into one image, but retaining the independence of each image while ensuring that they are spatially aligned so that they can be compared and analyzed with each other. The result of the registration is a set of transformation parameters that can be used to map one image into the space of another image so that they are spatially aligned. This process is not to merge images, but to ensure that they can be mapped and compared with each other in space. By registering images of different modalities (such as CT, MRI, PET) together, the advantages of each modality can be combined to obtain more comprehensive information.
[0043] In some embodiments of the present application, normalization processing is performed on multiple medical images, including:
[0044] Get the signal-to-noise ratio and contrast-to-noise ratio of each medical image, read the pixel values of all pixels in each medical image, traverse all pixels to determine the maximum pixel value g1 and the minimum pixel value g2 in each medical image, and define is the light-dark ratio;
[0045] The quality index of the medical image is obtained by weighted summing the signal-to-noise ratio, the contrast-to-noise ratio and the brightness-to-darkness ratio, the medical images are sorted according to the size of the quality index, each medical image is used as a pre-reference image in turn, the resolution and size of other medical images are converted to the pre-reference image, and the peak signal-to-noise ratio of each other medical image is determined to measure the image difference before and after the conversion of each other medical image, the conversion effect of the overall medical image is determined by the peak signal-to-noise ratio of each other medical image, a pre-reference image is selected as a reference image according to the conversion effect and the quality index, and the resolution and size of other medical images are converted to the reference image;
[0046] Traverse all pixels of all medical images, determine the maximum pixel value gmax and the minimum pixel value gmin among all pixels, adjust the standard zoom scale according to the difference between the pixel range of gmin-gmax and the standard pixel range, and obtain the actual zoom scale;
[0047] All medical images are normalized to the same resolution, size, and grayscale range using reference image resolution and size and actual scaling.
[0048] In this embodiment, the signal-to-noise ratio is the ratio of signal (useful medical information) to noise (random interference). A high signal-to-noise ratio means that there is more useful information in the image. The contrast-to-noise ratio (CNR) is the ratio of contrast to noise between different tissue types. A high CNR helps to distinguish similar tissue types. The brightness-to-darkness ratio refers to the ratio of the brightest and darkest areas that the image can capture. It is expressed logarithmically with 10 as the base, which is easy for the human eye to distinguish. A large brightness-to-darkness ratio helps to display high-contrast and low-contrast structures at the same time. A quality index of a medical image is defined by three indicators to describe the image quality.
[0049] In this embodiment, each medical image is taken as a pre-reference image in turn, so that the resolution and size of other medical images are converted to the pre-reference image. Each medical image is taken as a pre-reference image, and then the cost of converting other images to the pre-reference image and the peak signal-to-noise ratio are calculated. The image difference before and after the conversion of each other medical image is measured, and the difference of each image is integrated to obtain the conversion effect. Combined with the conversion effect and quality indicators, an image that can take both into account is selected as a reference image.
[0050] In this embodiment, the maximum pixel value gmax and the minimum pixel value gmin of all pixels are determined, and the standard zoom scale is adjusted according to the difference between the pixel range of gmin-gmax and the standard pixel range to obtain the actual zoom scale. For example, for an 8-bit grayscale image, its grayscale range is from 0 (black) to 255 (white), 0-255 is the standard pixel range, and the corresponding standard zoom scale is 0-1. When gmin-gmax is 0-230, 0-1 is adjusted accordingly to obtain the actual zoom scale, which can scale the grayscale values of all images to a reasonable range, thereby eliminating the influence of grayscale value differences under different devices or different imaging conditions, which helps to improve the accuracy and consistency of subsequent processing and analysis.
[0051] In some embodiments of the present application, the processed multiple medical images are registered in different modalities and at different times to generate time series images, including:
[0052] The mutual information between multiple medical images in the same period is calculated and optimized through optimization algorithms. At the same time, the image transformation parameters will be adjusted. The optimal image transformation parameters are determined through multiple iterations to maximize the mutual information, so that multiple medical images in the same period are registered to the same space.
[0053] The registered medical images are arranged in chronological order to generate time series images.
[0054] In this embodiment, the registration process of different modalities is as follows: Figure 2 As shown in the figure, a metric is defined to quantify the difference between images, such as mutual information, square error, etc. Mutual information is a commonly used metric that reflects the degree of information sharing between multiple images. When this degree of sharing is maximized, it means that the optimization of multiple images is completed. Use optimization algorithms (such as gradient descent, conjugate gradient method, Newton method, etc.) to maximize the similarity metric. During the optimization process, the parameters in the search space (such as image transformation parameters) will be adjusted to reduce the difference between the two images. By iteratively calculating the similarity metric and adjusting the parameters, a satisfactory registration result can be achieved. The registered images are post-processed, such as smoothing, cropping, or merging, to obtain the final registered image.
[0055] It should be noted that the specific process of the above-mentioned multimodal image registration is not the focus of protection of this solution. Other methods that can meet the above requirements are also possible and will not be repeated here.
[0056] In some embodiments of the present application, before determining multiple tumor regions to be determined and the invasion factor corresponding to each tumor region to be determined in the original contour of each organ according to the PET scan image in the medical image, the method further includes:
[0057] According to the basic information of the patient to be identified, the closest person is found in the person database, the normal PET scan image of the person is retrieved, and the pixel values of all pixels within the original contour of each organ and the pixel values of all pixels except the original contour of the organ are determined according to the normal PET scan image, so as to construct the standard pixel range of the inner contour-outer contour within the original contour of the organ and the standard pixel range of other areas, so as to simulate the pixel conditions within the original contour of each organ and the pixel conditions of other areas except the organs under the healthy condition of the patient to be identified;
[0058] The pixels of the inner contour are pixels completely located within the original contour of the organ, and the pixels of the outer contour are pixels partially located within the original contour of the organ and partially beyond the original contour of the organ.
[0059] In this embodiment, PET (positron emission tomography) is an advanced medical imaging technique that uses radiopharmaceuticals (compounds labeled with positron-emitting nuclides) to track biochemical processes in the human body. In PET scans, malignant tumors usually show higher metabolic activity because the metabolism of malignant tumor cells is more active than that of normal cells, especially glucose metabolism.
[0060] How a PET scan works:
[0061] 1. Radiopharmaceuticals: Patients first take or inject a drug containing a positron-emitting radionuclide, such as 18F-FDG (fluorodeoxyglucose). This drug competes with glucose for entry into cells and accumulates in metabolically active areas.
[0062] 2. Positron emission: When positron-emitting nuclides decay inside cells, they emit positrons, which immediately interact with electrons in the surrounding environment to produce a pair of photons (photon pairs).
[0063] 3. Photon pair detection: Detectors in the PET scanner detect these photon pairs and calculate their position and timing to create an image.
[0064] 4. Metabolic activity mapping: Since malignant tumor cells have a higher metabolic demand for glucose, they will absorb more 18F-FDG and thus appear as high metabolic areas on PET images.
[0065] Images shown on a PET scan:
[0066] PET scans produce images that show metabolically active areas in the body, which are represented by different colors or grayscale levels on the image. Generally, areas with higher metabolic activity appear in brighter colors or higher grayscale on the image.
[0067] In this embodiment, by finding a normal PET scan image of a healthy person that is closest to the patient to be identified (metabolism-related conditions), the pixel conditions within the original contours of each organ of the patient to be identified and the pixel conditions of other areas except the organs are simulated. The standard range of pixels within the original contours of each healthy organ and other areas is determined, and the area where the tumor may exist is preliminarily determined by comparing the grayscale conditions of the pixels.
[0068] Step S102, marking the original contours of each organ on the time series image using the basic information of the patient to be identified, and determining multiple tumor pending areas and the invasion factors corresponding to each tumor pending area in the original contours of each organ based on the PET scan image in the medical image, and expanding the tumor pending area based on the invasion factor to generate multiple tumor determination areas.
[0069] In this embodiment, the original outline of each organ is marked on the time series image by the basic information of the patient to be identified, which means that the size of the tissue structure of different people with different basic information such as height and weight will be different. The size and shape of the human body will affect the size of its internal organs and tissues. Therefore, the original outline of each organ is determined first. The area where the tumor may exist (tumor pending area) is preliminarily judged by comparing the grayscale of the pixels. Malignant tumors may show invasion of surrounding tissues, such as disappearance of fat gaps, displacement or destruction of adjacent structures, so an invasion factor is determined. This factor is used to measure the area where the tumor may extend, expand the tumor pending area, and generate multiple tumor determination areas.
[0070] In some embodiments of the present application, multiple tumor regions to be determined and invasion factors corresponding to each tumor region to be determined are determined in the original contours of each organ according to the PET scan image in the medical image, including:
[0071] Determine the actual pixel range of the inner contour-outer contour within the original contour of each organ and the actual pixel range of other regions according to the PET scan image of the patient to be identified, completely split the other regions into several sub-regions, and determine the actual pixel range of each sub-region;
[0072] The tumor region to be determined is determined by comparing the actual pixel range of the inner contour-outer contour in the original contour of each organ with the standard pixel range, and the original contour of the corresponding organ where only the actual pixel range of the inner contour exceeds the standard pixel range is used as the first tumor region to be determined, the original contour of the corresponding organ where only the actual pixel range of the outer contour exceeds the standard pixel range is used as the second tumor region to be determined, and the original contour of the corresponding organ where both the actual pixel range of the inner contour and the actual pixel range of the outer contour exceed the standard pixel range is used as the third tumor region to be determined;
[0073] The tumor region to be determined is determined by comparing the actual pixel range of each sub-region with the standard pixel range, the sub-region whose actual pixel range exceeds the standard pixel range is merged into the original contour of the closest organ, and the original contour of the organ after the sub-region is merged is used as the fourth tumor region to be determined, and the first tumor region to be determined, the second tumor region to be determined, the third tumor region to be determined, and the fourth tumor region to be determined are all categories of tumor regions to be determined;
[0074] Determine the invasion factor according to the category of the tumor area to be determined, the actual pixel range and the standard pixel range; ;
[0075] in, As an invasion factor, is the maximum value among multiple invasion distances corresponding to all tumor types mapped to the category of the tumor region to be determined, and They are the pixel conversion coefficients of the inner contour-outer contour and the pixel conversion coefficients of other areas, and The pixels corresponding to the inner contour and the outer contour exceed the weight respectively. , , The actual pixel range of the inner contour, outer contour, and other areas exceeds the standard pixel range. When the actual pixel range does not exceed the standard pixel range, the corresponding or or is 0.
[0076] In this embodiment, when comparing the actual pixel range with the standard pixel range, the situation of each pixel in the area is first determined, and the pixel range of the entire area is obtained after integration, and then compared. Other areas are completely divided into several sub-areas. The specific splitting method is not limited here, but all areas except organs must be completely split without omission. Tumors can grow inside organs, outside organs or on the edge. In medical imaging, the location and appearance of tumors depend on where and how they grow. The following are some common situations:
[0077] 1. Growing inside an organ:
[0078] Tumors grow in the parenchyma of an organ and may disrupt the organ's normal structure and function.
[0079] On imaging, the tumor usually appears as an abnormal area inside an organ, which may be a mass, nodule, or other abnormal shape.
[0080] 2. Growing on the edge of an organ:
[0081] The tumor may grow along the edges of an organ, which may cause a change in the shape of the organ.
[0082] On imaging, such tumors may appear as blurring or spiculation of the organ margins or as focal bulging of the organ contour.
[0083] 3. Growing outside or around an organ:
[0084] A tumor may be completely outside of an organ but still have a direct connection to it.
[0085] On imaging, this tumor may appear as a mass or nodule around an organ, with contact or connection to it.
[0086] In this embodiment, the area is divided into a first tumor pending area, a second tumor pending area, a third tumor pending area, and a fourth tumor pending area according to the growth of the tumor. Although the fourth tumor pending area is a tumor growing outside the organ, it is still related to the adjacent organs and the organs need to be included.
[0087] In this embodiment, the maximum value of multiple invasion distances corresponding to all tumor types mapped to the tumor area category means that different tumor area categories (including locations) may correspond to different tumor types, and the invasion distance corresponding to each tumor type is obtained (the historical average of the tumor invasion extension distance), and the maximum value is selected from multiple distances to avoid leakage of tumor characteristics.
[0088] In some embodiments of the present application, the tumor area to be determined is expanded according to the invasion factor to generate multiple tumor determination areas, including:
[0089] For the expansion of the first tumor region to be determined, the second tumor region to be determined, and the third tumor region to be determined, the center point within the original outline of the organ is used as the center of the circle and the invasion factor is used as the radius to expand outward, thereby forming an invasion circular area, and the invasion circular area is used as the tumor determination area;
[0090] For the expansion of the fourth tumor area to be determined, the center point of other areas in the fourth tumor area to be determined is used as the center of the circle and the invasion factor is used as the radius to expand outward to form an invasion circular area, and the union area of the invasion circular area and the fourth tumor area to be determined is used as the tumor determination area.
[0091] In this embodiment, because the edge of the malignant tumor is fuzzy and irregular, and burr-like protrusions and irregular shapes may appear, an invasion circular area is used here to describe the area that the malignant tumor may affect the surrounding area (the invasion circular area can basically cover the tumor and the extension area).
[0092] Step S103, multiple tumor determination regions are separated from the time series images, and a time series tumor determination region is generated, and tumor features in the time series tumor determination region are extracted through a CNN model.
[0093] In this embodiment, multiple tumor determination areas are separated from the images on the time series images. Through image segmentation technology, the tumor area can be separated from the background, and the features of the tumor area can be extracted separately. Here, a CNN convolutional neural network model is trained in advance to extract the image features of the separated tumor determination area as tumor features. The features here include size, shape, edge, texture features, etc. Benign tumors are usually small and regular in shape; malignant tumors may be larger and irregular in shape. Benign tumors have clear and smooth edges; malignant tumors have blurred and irregular edges, and may have burr-like protrusions. Through texture analysis, the texture complexity of the tumor area can be determined. Benign tumors usually have uniform textures, and malignant tumors may have more complex textures.
[0094] It should be noted that the specific process of training the CNN convolutional neural network model is not the focus of protection of this solution. Other methods that can meet the above requirements are also acceptable and will not be repeated here.
[0095] Step S104, calculate the tumor area and invasion area based on the tumor characteristics, determine the overall impact area through the tumor area and invasion area, and help doctors identify benign and malignant tumors through the overall impact area.
[0096] In this embodiment, the tumor area and invasion area are calculated, and the overall impact area of the two is determined. The simple addition of the two cannot fully represent all areas. The overall impact area is used to help doctors identify benign and malignant tumors, avoiding the relative error of the human eye to the area, thereby improving the accuracy of identifying benign and malignant tumors.
[0097] In some embodiments of the present application, the tumor area and invasion area are calculated based on tumor characteristics, including:
[0098] Tumor features are arranged in chronological order, and different weights are assigned to tumor features according to the time sequence. The later the tumor feature is, the greater the weight is, so as to determine the tumor area and invasion area.
[0099] In this embodiment, the tumor area and the invasion area are first determined by the CNN model, and the tumor characteristics of multiple time nodes are combined to determine the size of each area of the tumor area and the invasion area. The later the tumor characteristics are, the greater the weight is. The tumor area and the invasion area are determined by combining the tumor characteristics of all time points to reduce the error of the tumor characteristics of a single time node.
[0100] In some embodiments of the present application, the overall affected area is determined by the tumor area and the invasion area, including:
[0101] ;
[0102] Where S is the overall impact area, is the tumor area, is the invasion area, exp is the exponential function, and k is a preset constant.
[0103] In this embodiment, The value range is between 1 and 1.37, which represents the correction of the invasion area to the sum of the tumor area and the invasion area.
[0104] By applying the above technical solutions, the basic information and medical images of the patient to be identified are obtained, multiple medical images are normalized, and the processed multiple medical images are registered in different modalities and at different times to generate time series images; the original contours of each organ are marked on the time series image by the basic information of the patient to be identified, and multiple tumor pending areas and the invasion factors corresponding to each tumor pending area are determined in the original contours of each organ according to the PET scan image in the medical image, and the tumor pending area is expanded according to the invasion factor to generate multiple tumor determination areas; multiple tumor determination areas are separated from the image on the time series image, and a time series tumor determination area is generated, and the tumor features in the time series tumor determination area are extracted by the CNN model; the tumor area and the invasion area are calculated based on the tumor features, and the overall impact area is determined by the tumor area and the invasion area, and the overall impact area is used to help doctors realize the identification of benign and malignant tumors. This application facilitates the comparison of subsequent medical images and the image adaptability of the input model by normalizing multiple medical images. According to the PET scan images in medical images, multiple tumor areas and the corresponding invasion factors of each tumor area are determined in the original contours of each organ, ensuring the accuracy of the tumor area range, thereby improving the reliability of the CNN model to extract features. Finally, the overall impact area is used to help doctors make references, thereby improving the accuracy of tumor benign and malignant judgments, reducing human errors, and improving recognition efficiency.
[0105] Through the description of the above implementation methods, those skilled in the art can clearly understand that the present invention can be implemented by hardware, or by software plus a necessary general hardware platform. Based on such an understanding, the technical solution of the present invention can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (which can be a CD-ROM, a USB flash drive, a mobile hard disk, etc.), including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each implementation scenario of the present invention.
[0106] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit it. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present application.
Claims
1. A method for identifying benign and malignant tumors based on imaging data and deep learning. It is characterized in that include: Obtain the basic information and medical images of the patient to be identified, normalize multiple medical images, and perform registration of the processed multiple medical images in different modalities and at different times to generate time series images; The original contours of each organ are marked on the time series images based on the basic information of the patient to be identified, and multiple tumor pending areas and the invasion factors corresponding to each tumor pending area are determined in the original contours of each organ based on the PET scan images in the medical images, and the tumor pending areas are expanded based on the invasion factors to generate multiple tumor determination areas; Separate multiple tumor determination areas from the time series images and generate time series tumor determination areas. Use the CNN model to extract tumor features in the time series tumor determination areas. Calculate the tumor area and invasion area based on tumor characteristics, determine the overall impact area through the tumor area and invasion area, and help doctors identify benign and malignant tumors through the overall impact area; Before determining a plurality of tumor regions to be determined and the invasion factor corresponding to each tumor region to be determined in the original contour of each organ according to the PET scan image in the medical image, the method further comprises: According to the basic information of the patient to be identified, the closest person is found in the person database, the normal PET scan image of the person is retrieved, and the pixel values of all pixels within the original contour of each organ and the pixel values of all pixels except the original contour of the organ are determined according to the normal PET scan image, so as to construct the standard pixel range of the inner contour-outer contour within the original contour of the organ and the standard pixel range of other areas, so as to simulate the pixel conditions within the original contour of each organ and the pixel conditions of other areas except the organs under the healthy condition of the patient to be identified; The pixels of the inner contour are pixels completely located within the original contour of the organ, and the pixels of the outer contour are pixels partially located within the original contour of the organ and partially beyond the original contour of the organ; Based on the PET scan images in the medical images, multiple tumor areas to be determined and the invasive factors corresponding to each tumor area to be determined are determined in the original contours of each organ, including: Determine the actual pixel range of the inner contour-outer contour within the original contour of each organ and the actual pixel range of other regions according to the PET scan image of the patient to be identified, completely split the other regions into several sub-regions, and determine the actual pixel range of each sub-region; The tumor region to be determined is determined by comparing the actual pixel range of the inner contour-outer contour in the original contour of each organ with the standard pixel range, and the original contour of the corresponding organ where only the actual pixel range of the inner contour exceeds the standard pixel range is used as the first tumor region to be determined, the original contour of the corresponding organ where only the actual pixel range of the outer contour exceeds the standard pixel range is used as the second tumor region to be determined, and the original contour of the corresponding organ where both the actual pixel range of the inner contour and the actual pixel range of the outer contour exceed the standard pixel range is used as the third tumor region to be determined; The tumor region to be determined is determined by comparing the actual pixel range of each sub-region with the standard pixel range, the sub-region whose actual pixel range exceeds the standard pixel range is merged into the original contour of the closest organ, and the original contour of the organ after the sub-region is merged is used as the fourth tumor region to be determined, and the first tumor region to be determined, the second tumor region to be determined, the third tumor region to be determined, and the fourth tumor region to be determined are all categories of tumor regions to be determined; Determine the invasion factor according to the category of the tumor area to be determined, the actual pixel range and the standard pixel range; ; in, As an invasion factor, is the maximum value among multiple invasion distances corresponding to all tumor types mapped to the category of the tumor region to be determined, and They are the pixel conversion coefficients of the inner contour-outer contour and the pixel conversion coefficients of other areas, and The pixels corresponding to the inner contour and the outer contour exceed the weight respectively. , , The actual pixel range of the inner contour, outer contour, and other areas exceeds the standard pixel range. When the actual pixel range does not exceed the standard pixel range, the corresponding or or is 0.
2. A method for identifying benign and malignant tumors based on image data and deep learning as claimed in claim 1, It is characterized in that Normalize multiple medical images, including: Get the signal-to-noise ratio and contrast-to-noise ratio of each medical image, read the pixel values of all pixels in each medical image, traverse all pixels to determine the maximum pixel value g1 and the minimum pixel value g2 in each medical image, and define is the light-dark ratio; The quality index of the medical image is obtained by weighted summing the signal-to-noise ratio, the contrast-to-noise ratio and the brightness-to-darkness ratio, the medical images are sorted according to the size of the quality index, each medical image is used as a pre-reference image in turn, the resolution and size of other medical images are converted to the pre-reference image, and the peak signal-to-noise ratio of each other medical image is determined to measure the image difference before and after the conversion of each other medical image, the conversion effect of the overall medical image is determined by the peak signal-to-noise ratio of each other medical image, a pre-reference image is selected as a reference image according to the conversion effect and the quality index, and the resolution and size of other medical images are converted to the reference image; Traverse all pixels of all medical images, determine the maximum pixel value gmax and the minimum pixel value gmin among all pixels, adjust the standard zoom scale according to the difference between the pixel range of gmin-gmax and the standard pixel range, and obtain the actual zoom scale; All medical images are normalized to the same resolution, size, and grayscale range using reference image resolution and size and actual scaling.
3. The method for identifying benign and malignant tumors based on image data and deep learning as claimed in claim 1, It is characterized in that And the processed multiple medical images are registered in different modalities and at different times to generate time series images, including: The mutual information between multiple medical images in the same period is calculated and optimized through optimization algorithms. At the same time, the image transformation parameters will be adjusted. The optimal image transformation parameters are determined through multiple iterations to maximize the mutual information, so that multiple medical images in the same period are registered to the same space. The registered medical images are arranged in chronological order to generate time series images.
4. The method for identifying benign and malignant tumors based on image data and deep learning as claimed in claim 1, It is characterized in that The tumor determination area is expanded according to the invasion factor to generate multiple tumor determination areas, including: For the expansion of the first tumor region to be determined, the second tumor region to be determined, and the third tumor region to be determined, the center point within the original outline of the organ is used as the center of the circle and the invasion factor is used as the radius to expand outward, thereby forming an invasion circular area, and the invasion circular area is used as the tumor determination area; For the expansion of the fourth tumor area to be determined, the center point of other areas in the fourth tumor area to be determined is used as the center of the circle and the invasion factor is used as the radius to expand outward to form an invasion circular area, and the union area of the invasion circular area and the fourth tumor area to be determined is used as the tumor determination area.
5. The method for identifying benign and malignant tumors based on image data and deep learning as claimed in claim 1, It is characterized in that Tumor area and invasive area were calculated based on tumor characteristics, including: Tumor features are arranged in chronological order, and different weights are assigned to tumor features according to the time sequence. The later the tumor feature is, the greater the weight is, so as to determine the tumor area and invasion area.
6. A method for identifying benign and malignant tumors based on image data and deep learning as claimed in claim 5, It is characterized in that The overall affected area was determined by tumor area and invasion area, including: ; Where S is the overall impact area, is the tumor area, is the invasion area, exp is the exponential function, and k is a preset constant.
Citation Information
Patent Citations
Small sample tumor necrosis rate classification prediction device based on deep learning
CN111275691A
Computer supported review of tumors in histology images and post operative tumor margin assessment
CN113366530A
Benign and malignant tumor identification method and system based on image data and deep learning
CN113743463A
Medical image diagnosis, comparison and reading method
CN118262875A
