TWI781438B - Medical image analysis system and its training method - Google Patents

Abstract

A medical image analysis system and its training method, which obtains multiple image data, and establishes a three-dimensional data data based on the image data at previous and subsequent time points; re-slices the three-dimensional data data to unify the interval distance of the slices; re-slices The original window value of each pixel in the subsequent image data is remapped to obtain a new window value for each pixel; the slices with the new window value are stacked to obtain a new three-dimensional data, and at least one In specific blocks, different body structures are distinguished with different colors, and a tumor location is marked with a specific color. After receiving the new image data, the above steps can be used to determine whether the new image data contains a tumor, and find out the location of the tumor. The invention can improve the resolution of medical images, and improve the speed and accuracy of doctor's interpretation.

Description

Medical image analysis system and its training method

The invention relates to an image processing and analysis technology, in particular to a medical image analysis system and its training method.

Medical image interpretation is very important for doctors and patients, because X-rays, tomographic scans, nuclear magnetic resonance and other means can know the internal conditions of the body under the premise of non-invasive detection. The current smart medical image aided judgment system captures medical images, converts them from DICOM format into bmp, jpg, and png image formats, and manually marks tumors in the images. Then build the AI model, set parameters and training goals, then use the marked pictures to train the AI model, and conduct a blind test with samples to determine whether the AI model has been trained.

However, a large amount of important information will be lost after image conversion, and it is easier to reduce the resolution of the image sample. In addition, the training parameters cannot effectively refer to medical related information. The system only relies on the training model of deep learning, and in the case of very limited image samples for training, it is even more impossible to achieve the goal of 100% zero error.

In view of this, the present invention proposes a medical image analysis system and its training method to effectively solve the above-mentioned problems. The specific architecture and its implementation will be described in detail below:

The main purpose of the present invention is to provide a medical image analysis system and its training method, which superimposes the image data at the front and rear time points into a complete three-dimensional data and then slices again, so that all the image data can be unified into the same slice interval distance , to achieve the purpose of normalization.

Another object of the present invention is to provide a medical image analysis system and its training method, which uses the method of remapping the window value of each pixel to make the image clearer, without converting the image format, and will not lose the important information contained in the image .

Another object of the present invention is to provide a medical image analysis system and its training method, which uses different colors to distinguish different body structures, so that professional medical personnel can find tumor locations more quickly during image interpretation.

To achieve the above purpose, the present invention provides a medical image analysis system, including: a database, including a plurality of image data, each pixel in the image data has an original window value; a medical judgment module, connected to the data A library for training the ability to judge the tumor location in the image data, including: a normalization module, which creates a three-dimensional data data according to the image data at previous and subsequent time points, and inserts at least one window into the three-dimensional data data After making the 3D data conform to the actual human body structure, slice the 3D data according to the new standard, so as to unify the interval distances of the slices of the image data to form a plurality of sliced images data; a data remodeling group, which remaps the original window values in the sliced image data after re-slicing according to the window values of each human body structure, so that each of the pixels in the sliced image data obtains a A new window value is used to form a new slice image data, and then the slices of the new slice image data are stacked to obtain a new three-dimensional data; and a tumor marker module is used to find at least one from the new three-dimensional data. Specific blocks, with different colors to distinguish different body structures, each of the specific blocks is The section between a high risk value 1 and a high risk value 2 is the window value of tumor occurrence, and according to this, those who meet the high risk value 1 and high risk value 2 are marked with a specific color for a tumor location; and a mixed The rejudgment module is connected to the medical judgment module, receives a plurality of new image data and uses the medical judgment module to perform normalization, data reconstruction and tumor marking, and then judges whether the new image data contains tumors, and finds out the Tumor location.

According to an embodiment of the present invention, the specific block is marked as a tumor block.

According to an embodiment of the present invention, the image data are tomographic images.

According to an embodiment of the present invention, the hybrid review module superimposes the continuous new image data and gives them different colors, flips the superimposed new image data to detect whether there is a change in the specific color, and Judging whether the new image data contains a tumor, and finding out the location of the tumor.

The present invention also provides a training method for a medical image analysis system, which includes the following steps: a normalization module in a medical judgment module obtains multiple image data from a database, and the normalization module obtains multiple image data according to the time points before and after A three-dimensional data is established from the image data; the normalization module inserts at least one window value into the three-dimensional data to make the three-dimensional data conform to the actual human body structure, and then re-cuts the three-dimensional data according to a new standard. Slicing, so as to unify the interval distance of the slices of the image data to form a plurality of slice image data; a data reconstruction group in the medical judgment module divides each pixel in the slice image data Having an original window value according to the window value of each human body structure, remapping to a color spectrum, so that each of the pixels in the slice image data obtains a new window value, respectively forming a new slice image data; the data is reconstructed The group will stack the slices of the new slice image data to obtain a new three-dimensional data; and one of the tumor marking modules in the medical judgment module finds at least one characteristic from the new three-dimensional data. Define blocks, use different colors to distinguish different body structures, and each specific block is based on a section between a high risk value 1 and a high risk value 2 as the window value for tumor occurrence, and according to this will meet the high risk A value of 1 and a high risk value of 2 mark a tumor location with a specific color.

According to an embodiment of the present invention, the 3D data is inserted into the at least one window value by an interpolation method, so that the 3D data conforms to the actual human body structure, and then re-sliced.

According to the embodiment of the present invention, a hybrid review module is used to flip the specific block from multiple angles, so as to increase the number of image samples including the tumor location.

According to an embodiment of the present invention, such bodily structures include organ walls, air vessels, and tumors or nodules.

According to an embodiment of the present invention, the original window values are mapped from a preset density value interval to an interval of 0-255 to obtain the new window values.

According to an embodiment of the present invention, the original window values are mapped to the new window values using a linear or nonlinear method.

10:Medical image analysis system

12: Database

14:Medical judgment module

142: Normalization module

144:Data reconstruction group

146: Tumor Marking Module

16: Mixed review module

Fig. 1 is a block diagram of the medical image analysis system of the present invention.

Fig. 2 is a flowchart of the training method of the medical image analysis system of the present invention.

FIG. 3 is a schematic diagram of the remapping steps in the medical image analysis system and its training method of the present invention.

Fig. 4A is a schematic diagram of a tomographic image of the prior art, and Fig. 4B is an image processed by the medical image analysis system of the present invention.

The present invention provides a medical image analysis system and its training method, which are used to normalize medical images, reconstruct data and mark tumors, improve the resolution of image data without losing the information contained in the images, and An AI model trained to determine whether an image contains a tumor. In this way, professional medical personnel can quickly and accurately interpret images by using the system of the present invention to assist judgment, and find out the location of the tumor therefrom.

Please refer to FIG. 1 , which is a perspective view of an embodiment of the medical image analysis system of the present invention. The medical image analysis system 10 of the present invention includes a database 12 , a medical judgment module 14 and a hybrid review module 16 . The database 12 includes multiple medical image data, such as tomography (CT) image data, and each pixel in the image data has an original window value (CT-number). The medical judgment module 14 is an AI model, which is connected to the database 12, can obtain the image data in the database 12, and trains the ability to judge the tumor location in the image data. The mixed rejudgment module 16 is connected to the medical judgment module 14. When new image data enters the medical image analysis system 10, the mixed rejudgment module 16 can quickly judge whether the new image data is Contains the tumor, and finds out where the tumor is located.

The medical judgment module 14 further includes a normalization module 142 , a data reconstruction module 144 and a tumor marker module 146 . Among them, the normalization module 142 combines the image data at the front and back time points to create a complete three-dimensional data, and inserts at least one window value into the three-dimensional data, so that the three-dimensional data conforms to the actual human body structure, Then re-slice the 3D data with a new standard cutting to unify the interval distance of the slices of the image data to form multiple slice image data. For example, some tomographic scanning instruments are all 0.5 cm, and some are all 0.2 cm After the normalization module 142 forms the three-dimensional data from the image data of these slices, it is cut again with a new standard, such as 0.1 centimeters, so that all the slices of the image data can be normalized. To form a plurality of slice image data, no matter which tomographic scanning instrument is used for detection, the final storage in the medical image analysis system 10 has a uniform separation distance.

In particular, because the three-dimensional data may have a large gap in the window value due to different human body structures, for example, the window values of bones and blood vessels are quite different. Therefore, the present invention uses an interpolation method to insert at least one window value into the 3D data, so that the 3D data conforms to the actual human body structure, and then slices again.

The data reconstruction module 144 is connected to the normalization module 142 . After the image data is re-sliced, the original original window value of the sliced image data is remapped according to the window value of each human body structure, so that each pixel in the sliced image data obtains a new window value, forming a new sliced image data respectively, The slices of the new sliced image data are stacked again to obtain a new three-dimensional data. The tumor marking module 146 is connected to the data reconstruction module 144 for finding at least one specific block from the new 3D data. This specific block is all the blocks that are suspected to contain tumors. The tumor marking module 146 uses different colors to distinguish different body structures, such as organ walls, air vessels, tumors or nodules, etc., and then marks the specific block with a specific color - tumor location. Since the medical judgment module 14 does not yet have the ability to accurately judge the tumor location, the marking of the tumor location is performed by professional medical personnel to train the medical judgment module 14 to judge the tumor.

After the medical judgment module 14 is trained, the hybrid rejudgment module 16 can assist in judging whether there is a tumor on the new image data. The hybrid review module 16 can also superimpose continuous new image data into 3D image blocks, and give different colors to distinguish different body structures, and then flip the superimposed new image data to observe from different angles , to detect whether there is a specific color change. For example, if the color of the tumor is red, the image data does not appear red at the current angle, but red may appear after flipping, indicating that the tumor is hidden behind the organ. In this way, The mixed rejudgment module 16 can judge whether the new image data contains a tumor, and find out the position of the tumor.

In addition, since the image data of the tumor sample is quite limited, the present invention can increase the number of image samples including the tumor position after the specific block is flipped from multiple angles.

Fig. 2 is a flowchart of the training method of the medical image analysis system of the present invention. First in step S10, one of the normalization modules in the medical judgment module first obtains multiple image data from a database, such as tomographic images, and then builds a three-dimensional data data according to the image data at the front and rear time points; and then proceeds as in the steps As described in S12, the normalization module inserts at least one window value into the 3D data, so that the 3D data conforms to the actual human body structure, and then cuts the 3D data according to a new standard to slice the image data The interval distances of these slices are unified, so that all image data have the same slice interval distance, forming a plurality of slice image data; each pixel in each slice image data has an original window value, and then in the step In S14, the original window value of each pixel in each slice image data is remapped to a color spectrum, so that each pixel in each slice image data obtains a new window value to form a new slice image data respectively; then In step S16, the slices with the image data of each new slice are stacked to obtain a new three-dimensional data; finally, as described in step S18, one of the tumor marker modules in the medical judgment module finds out at least A specific block, the specific block is all blocks that may contain tumors, different body structures in the specific block are distinguished with different colors, and a tumor location is marked with a specific color. By repeating steps S10-S18 for multiple image data, the medical judgment module 14 of the present invention can be better trained, which is helpful for assisting doctors in judging the location of tumors.

Please refer to FIG. 3 for a specific embodiment of remapping to obtain a new window value in step S14. The density value (Hu) of the tomographic window value is between -1000 and 3000Hu, below -1000 is air, above 3000 is a high-density material. Assuming that the tomographic window value of the human body is concentrated between -900~600Hu, the present invention remaps it to the 0~255 interval of the color spectrum to obtain better image intensity. The probability of tumor occurrence is shown in the right curve of Figure 3. Therefore, the median value of -900~600Hu, high risk value 1 and high risk value 2 can be set, and the section between high risk value 1 and 2 is statistical Later, it is considered that the window value of tumor is more likely to occur. In the present invention, a linear or nonlinear method is used for mapping, and the original window value of -900~600Hu is mapped to a new window value of 0~255.

Fig. 4A is a schematic diagram of a tomographic image of the prior art, and Fig. 4B is an image processed by the medical image analysis system of the present invention. It can be seen from Figure 4A that the black part is darker, and the white filaments and dots are grayish and not obvious. However, the black parts, white filaments and dots in Figure 4B are obviously brighter, and there are more white filaments visible to the naked eye. Obviously, the image data processed by the medical image analysis system of the present invention is more conducive to image interpretation by professional medical personnel.

To sum up, a medical image analysis system and its training method provided by the present invention is to superimpose the image data at the front and rear time points into a complete three-dimensional data and then re-slice, so that all the image data can be unified into the same slice The interval distance makes the image data scanned by any medical institution or any instrument all have the same specifications; secondly, the image is clearer by remapping the window value of each pixel without converting the image format Therefore, the important information contained in the image will not be lost; moreover, the present invention uses different colors to distinguish different body structures, so that professional medical personnel can find the tumor location more quickly during image interpretation.

The above descriptions are only preferred embodiments of the present invention, and are not intended to limit the scope of the present invention. Therefore, all equivalent changes or modifications based on the features and spirit described in the scope of the application of the present invention shall be included in the scope of the patent application of the present invention.

Claims (12)

A medical image analysis system, comprising: a database containing multiple image data, each pixel in the image data has an original window value; a medical judgment module connected to the database for training the images The ability to judge the position of the tumor in the data includes: a normalization module, which creates a three-dimensional data according to the image data at the front and rear time points, and inserts at least one window value into the three-dimensional data, so that the three-dimensional data conforms to After the actual human body structure, the three-dimensional data is sliced again with a new standard, so as to unify the interval distances of the slices of the image data to form a plurality of slice image data; a data reconstruction group, The original window values in the sliced image data after re-slicing are remapped to a color spectrum according to the window value of each human body structure, so that each of the pixels in the sliced image data obtains a new window value, respectively forming a new slice image data, and then stacking the image data to obtain a new three-dimensional data; and a tumor marking module, finding at least one specific block from the new three-dimensional data, and using different colors to distinguish different Body structure, each of the specific blocks is based on a section between a high risk value 1 and a high risk value 2 as the window value for tumor occurrence, and based on this, those that meet the high risk value 1 and high risk value 2 are classified as a A specific color marks the location of a tumor; and A hybrid re-judgment module, connected to the medical judgment module, receiving multiple new image data, and using the medical judgment module to perform normalization, data reconstruction and tumor marking, to determine whether the new image data contains tumors, and Find out where the tumor is. The medical image analysis system as claimed in claim 1, wherein the specific block is marked as a tumor block. The medical image analysis system as described in Claim 1, wherein the image data are tomographic images. The medical image analysis system as described in claim 1, wherein the mixed review module superimposes the continuous new image data and gives them different colors, flips the superimposed new image data to detect whether there is the specific There is a change in color to determine whether the new image data contains a tumor, and to find out the location of the tumor. A training method for a medical image analysis system, comprising the following steps: a normalization module in a medical judgment module obtains multiple image data from a database, and the normalization module establishes a system based on the image data at previous and subsequent time points A three-dimensional data; the normalization module inserts at least one window value into the three-dimensional data, so that the three-dimensional data conforms to the actual human body structure, and then re-cuts the three-dimensional data according to a new standard to slice the three-dimensional data. A distance between the slices of the image data is unified to form a plurality of slice image data; a data reconstruction module in the medical judgment module uses an original window value of each pixel in the slice image data remapped to a color spectrum according to the window values of each anatomy, obtaining a new window value for each of the pixels in the sliced image data to form a new sliced image data; the data reconstruction group stacks the slices of the new sliced image data to obtain a new three-dimensional data; and One of the tumor marking modules in the medical judgment module finds at least one specific block from the new three-dimensional data, uses different colors to distinguish different body structures, and marks each specific block with a high risk value 1 and a The section between the high risk value 2 is the window value for tumor occurrence, and according to this, those meeting the high risk value 1 and high risk value 2 are marked with a specific color for a tumor location. The training method of the medical image analysis system according to claim 5, wherein the specific block is marked as a tumor block. The training method for a medical image analysis system as described in Claim 5, wherein the image data are tomographic images. The training method of the medical image analysis system as described in claim 5, wherein the three-dimensional data is inserted into the at least one window value by an interpolation method. The training method of the medical image analysis system as described in claim 5, wherein a hybrid review module is used to flip the specific block from multiple angles, so as to increase the number of image samples including the tumor location. The training method of the medical image analysis system as described in Claim 5, wherein the body structures include organ walls, air vessels, tumors or nodules. The training method of the medical image analysis system as described in Claim 5, wherein the original window values are mapped from a preset density value range to a range of 0-255 to obtain the new window values. The training method of the medical image analysis system as claimed in item 5 or 11, wherein the original window values are mapped to the new window values using a linear or nonlinear method.

Images