Computer vision based medical image processing method and system

Abstract

The application relates to the technical field of image processing, and discloses a medical image processing method and system based on computer vision. The method comprises the following steps: acquiring continuous two-dimensional tomographic images; defining a three-dimensional space formed by stacking two-dimensional images, constructing a conversion rule, mapping an original value to a color value and a visual prominence value, and setting a visual prominence adjustment coefficient for different positions in the three-dimensional space; for each pixel point, determining the color value and the visual prominence value according to the original value and the conversion rule, combining position information to determine the adjustment coefficient, and obtaining a final visual prominence value; constructing a voxel and generating a three-dimensional model, and generating a two-dimensional visualized image through simulation projection. The application can improve the diagnostic value of the three-dimensional visualized image.

Description

Technical Field

[0001] This application relates to the field of image processing technology, and in particular to a medical image processing method and system based on computer vision. Background Technology

[0002] In medical imaging diagnosis, constructing three-dimensional models based on two-dimensional tomographic images acquired by CT equipment has become an important means to assist doctors in observing the morphology of lesions and assessing the relationship between lesions and surrounding tissues.

[0003] However, in actual clinical applications, especially in the detection and delineation of deep tumors in abdominal organs such as the liver, the differentiation of pulmonary nodules and blood vessels in chest imaging, the localization of small lesions in the brain in neurological imaging, and complex cases where multiple tissues such as tumors, calcifications and normal parenchyma need to be observed simultaneously, traditional three-dimensional reconstruction methods have many limitations.

[0004] Current technologies typically employ fixed color and transparency mapping rules to uniformly process all voxels. When doctors need to observe deep lesions located in the central region of an organ, superficial tissues become visually prominent in 3D rendering, causing occlusion and blurring of the lesions. When the original values ​​(such as CT values) of the lesion and surrounding normal tissue are highly similar, traditional linear mapping cannot effectively distinguish between them, and the lesion is easily submerged in the background tissue. Many non-critical tissues, such as fat, bone, and air, are treated the same as critical lesions, occupying the doctor's visual attention and interfering with diagnostic efficiency. Furthermore, traditional methods ignore the positional differences of voxels in 3D space. Interfering tissues located at the organ's edge but with similar original values ​​to the lesion cannot be effectively suppressed using spatial information, making it difficult to achieve clear presentation when multiple tissues of interest need to be observed simultaneously. These problems directly affect the practicality and accuracy of 3D visualization images in clinical diagnosis. Therefore, a medical image processing method that can intelligently distinguish tissue types, highlight key structures, and suppress interfering information is needed to improve the clinical practicality and diagnostic efficiency of 3D visualization images.

[0005] Therefore, the present invention provides a medical image processing method and system based on computer vision. Summary of the Invention

[0006] The embodiments in this specification provide the following technical solutions: Step S1: Obtain continuous two-dimensional tomographic images. The two-dimensional tomographic images are obtained by scanning the target part of the human body layer by layer. They are composed of multiple pixels, and each pixel corresponds to an original value. Step S2: Define a three-dimensional space, which is composed of all two-dimensional images stacked and aligned. Construct transformation rules, which are used to map the original values ​​of different ranges to the corresponding color values ​​and visual prominence values. Based on the transformation rules, set the corresponding visual prominence adjustment coefficients for different positions in the three-dimensional space. Step S3: For each pixel in each two-dimensional image, determine the corresponding color value and visual prominence value according to the original value and conversion rules. Determine the corresponding visual prominence adjustment coefficient according to the position of the pixel in three-dimensional space. Multiply the visual prominence adjustment coefficient by the corresponding visual prominence value to obtain the final visual prominence value. Step S4: Construct a corresponding voxel in three-dimensional space based on the position of the pixel, assign the color value and final visual prominence value of the pixel to the corresponding voxel, generate a three-dimensional model composed of voxels, simulate projection of the three-dimensional model from a preset viewing direction, calculate and generate a two-dimensional visualization image.

[0007] A computer vision-based medical image processing system is also provided, comprising the following modules: The data acquisition module is used to acquire continuous two-dimensional tomographic images. The two-dimensional tomographic images are obtained by scanning the target part of the human body layer by layer. They are composed of multiple pixels, and each pixel corresponds to an original value. The rule building module is used to build transformation rules. The transformation rules are used to map the original values ​​of different ranges to the corresponding color values ​​and visual prominence values, define the three-dimensional space and set the corresponding visual prominence adjustment coefficients for different positions in the three-dimensional space. The three-dimensional space is composed of all two-dimensional images stacked and aligned. The data conversion module is used to determine the corresponding color value and visual prominence value for each pixel on each two-dimensional image based on the original value and conversion rules, determine the corresponding visual prominence adjustment coefficient based on the position of the pixel in three-dimensional space, and multiply the visual prominence adjustment coefficient by the corresponding visual prominence value to obtain the final visual prominence value. The projection generation module is used to construct a corresponding voxel in three-dimensional space based on the position of the pixel, assign the color value and final visual prominence value of the pixel to the corresponding voxel, generate a three-dimensional model composed of voxels, simulate the projection of the three-dimensional model from a preset viewing direction, and calculate and generate a two-dimensional visualization image.

[0008] Compared with the prior art, the beneficial effects of the present invention are at least as follows: The technical solution provided in this application introduces a spatial location weighting mechanism to give lesions located at the center of the target area higher visual prominence, effectively avoiding obscuring by superficial tissues and improving diagnostic accuracy. A dynamic visual prominence allocation function is constructed, combining original values ​​with spatial location information to effectively distinguish different tissues with similar original values. By statistically analyzing the frequency distribution and visual contribution of original values, the visual prominence function is iteratively optimized to automatically match the actual data distribution, improving adaptability to different tissue types. When multiple tissues of interest exist, a multi-peak processing mechanism is used to give the visual prominence function a multi-peak shape, meeting complex clinical needs. By assigning voxel-level color and visual prominence values, combined with simulated projection technology, the generated two-dimensional visualization image retains key structural information while reducing interference areas, enhancing the diagnostic value of the image. Attached Figure Description

[0009] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0010] Figure 1 This is a flowchart illustrating the steps of a computer vision-based medical image processing method in an embodiment of this application. Figure 2 This is a structural diagram of the computer vision-based medical image processing system in the embodiments of this application. Detailed Implementation

[0011] This application provides a computer vision-based medical image processing method and system. The terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and accompanying drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in a sequence other than that illustrated or described herein. Furthermore, the terms "comprising" or "having" and any variations thereof are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or device that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or devices.

[0012] For ease of understanding, the specific process of the embodiments of this application is described below. Please refer to [link / reference]. Figure 1 One embodiment of the computer vision-based medical image processing method in this application includes: Step S1: Obtain continuous two-dimensional tomographic images. The two-dimensional tomographic images are obtained by scanning the target part of the human body layer by layer. They are composed of multiple pixels, and each pixel corresponds to an original value.

[0013] Specifically, in the three-dimensional visualization diagnosis of CT images (such as CT images of liver tumors), traditional methods, when dealing with deep lesions, often result in lesions located in the central region of the liver appearing blurred in the three-dimensional image due to obstruction by superficial tissues, making it difficult to accurately observe their shape and boundaries. To solve the above technical problems, continuous two-dimensional tomographic images are first acquired. Two-dimensional tomographic images are cross-sectional medical images obtained by scanning the target part of the human body (such as the liver) layer by layer using a multi-slice spiral CT device. Each pixel corresponds to a raw value, which is the raw data obtained by scanning a certain part of the human body with medical equipment. It represents the degree of X-ray absorption by different tissues in the human body. After acquiring these continuous two-dimensional tomographic images, the basic raw data support is provided for subsequent construction of three-dimensional space, pixel-to-voxel conversion, and calculation of visual prominence.

[0014] Step S2: Define a three-dimensional space, which is composed of all two-dimensional images stacked and aligned. Construct transformation rules, which are used to map the original values ​​of different ranges to the corresponding color values ​​and visual prominence values. Based on the transformation rules, set the corresponding visual prominence adjustment coefficients for different positions in the three-dimensional space.

[0015] Specifically, after acquiring two-dimensional tomographic images, traditional methods typically establish simple linear mapping relationships, assigning the same display attributes to all raw values ​​under these linear mapping relationships. However, this approach leads to the following problems: 1. When the lesion the doctor needs to observe (e.g., a soft tissue tumor with a raw value of 40-60 HU) and the surrounding normal tissue (e.g., liver parenchyma with a raw value of 60-80 HU) are within similar raw value ranges, traditional methods cannot effectively distinguish between them, causing the lesion to be "submerged" in the three-dimensional image; 2. A large number of non-critical raw values ​​are treated the same as the raw values ​​of critical lesions, occupying a significant amount of visual attention and interfering with diagnosis; 3. Mechanical mapping based solely on the numerical magnitude of raw values ​​makes it impossible to understand which raw value ranges correspond to the lesions the doctor truly wants to see and which raw value ranges correspond to... The background organization can be weakened; to solve the above problems, a three-dimensional space is first defined. The three-dimensional space is a three-dimensional data field composed of all acquired two-dimensional tomographic images stacked after being precisely aligned according to the actual scan layer thickness and interlayer spacing. For example, for a 512×512 pixel CT image, if the scan layer thickness is 1mm and the number of layers is 200, a 512×512×200 three-dimensional space is formed, providing a precise three-dimensional coordinate reference system for each pixel. Then, a transformation rule is constructed. The transformation rule consists of two parts. The first part is the visual prominence assignment function, which is used to establish the correspondence between the original value and the visual prominence value. Visual prominence refers to the degree of significance of a voxel being observed in three-dimensional rendering. The value range is from 0 to 1, where 0 means completely transparent and invisible, and 1 means completely opaque. The second part is the color value assignment set, which is used to establish the correspondence between the original values ​​and a set of display color component values ​​(such as RGB values), so that tissues with different original value ranges can be presented with different colors in the final image for easy differentiation. Finally, based on the above conversion rules, corresponding prominence adjustment coefficients are set for different positions in three-dimensional space. By setting the prominence adjustment coefficients, voxels located in the central region of the target area receive larger adjustment coefficients, while voxels located in the edge region receive smaller adjustment coefficients, thereby realizing the differentiated control of visual prominence by spatial position. Through the conversion rules and visual prominence adjustment coefficients, the color amplitude and visual prominence correction when each pixel is converted into a voxel are provided, realizing the differentiated processing of tissues with different original value ranges. This allows the lesion areas that doctors are concerned about to obtain higher visual prominence and striking colors, solving the problem of visual confusion caused by the homogenization of all original value ranges in traditional methods.

[0016] Step S3: For each pixel in each two-dimensional image, determine the corresponding color value and visual prominence value according to the original value and conversion rules. Determine the corresponding visual prominence adjustment coefficient according to the position of the pixel in three-dimensional space. Multiply the visual prominence adjustment coefficient by the corresponding visual prominence value to obtain the final visual prominence value.

[0017] Specifically, to address the problem of deep lesions being obscured by superficial tissues due to the traditional method's neglect of voxel spatial location and application of the same processing mechanism to all regions, the following operations are performed on each pixel in each 2D tomographic image: First, based on the pixel's original value, the corresponding visual salience value is determined using a visual salience allocation function. Simultaneously, the color allocation set is searched to determine the corresponding color value. Then, based on the pixel's specific location in 3D space (e.g., at layer 100, row 256, column 256), the corresponding salience adjustment coefficient is determined. The calculation method for the salience adjustment coefficient will be explained in detail later. Finally, the visual salience adjustment coefficient is multiplied by the corresponding visual salience value to obtain the final visual salience value. Through these operations, voxels located in areas of interest (such as the central lesion of the liver) retain high visual salience and are thus clearly presented in subsequent rendering, while the visual salience of voxels located in non-areas of interest (such as the liver edge) is effectively attenuated and appropriately faded. This solves the visual interference problem caused by the indiscriminate use of spatial location information and provides precisely corrected voxel attribute data for the subsequent generation of 3D models with spatial focusing effects.

[0018] Step S4: Construct a corresponding voxel in three-dimensional space based on the position of the pixel, assign the color value and final visual prominence value of the pixel to the corresponding voxel, generate a three-dimensional model composed of voxels, simulate projection of the three-dimensional model from a preset viewing direction, calculate and generate a two-dimensional visualization image.

[0019] Specifically, after obtaining the color values ​​and final visual salience values ​​of all voxels, each voxel is assigned a corresponding lateral value and final visual salience value to generate a 3D model. To address the issue that the simple pure copper projection method cannot fully utilize the color and visual salience information of voxels, resulting in the spatial focusing effect not being fully presented in the final visualization image, a simulated projection is performed on the 3D model from a preset observation direction, such as from the patient's ventral side to the non-observation direction. A 2D visualization image is then calculated and generated. For example, a virtual ray is emitted from each pixel on the observation plane, penetrating the entire 3D voxel model. Multiple sampling points are collected along the ray path at preset sampling intervals (e.g., 0.5mm). The color value (e.g., RGB value) and final visual salience value (i.e., the visual salience after spatial weighting correction) of the voxel at each sampling point are obtained. Then, the image is projected in a front-to-back or back-to-front order. The color values ​​of all sampling points are accumulated and synthesized based on their corresponding final visual prominence values. For example, when the ray passes through the central region of the liver, it encounters a voxel with a final visual prominence of 0.8, which contributes significantly to the color and appears as a clear red tone in the final image. When the ray passes through the edge region of the liver, it encounters a voxel with a final visual prominence of 0.24, which contributes less to the color and appears as a faded effect in the final image. The display color of the pixel is then calculated. After traversing all pixels, a two-dimensional visualization image is generated. This image can clearly display the lesion in the central region of the target area (high final visual prominence) while fading or hiding non-critical tissues in the edge region (low final visual prominence). This allows the two-dimensional visualization image to clearly display the lesion in the central region of the target area while fading non-critical tissues, effectively solving the problem of deep lesions being obscured and key information not being effectively presented.

[0020] In one specific embodiment, the conversion rules for color and opacity are constructed, which specifically includes the following steps: A visual prominence allocation function and a color value allocation set are constructed. The visual prominence allocation function is used to establish the correspondence between the original values ​​and the visual prominence values, and the color value allocation set is used to establish the correspondence between the original values ​​and a set of display color component values.

[0021] Specifically, to address the problem that traditional methods only perform simple linear mapping of raw values ​​and cannot achieve differentiated display of different tissues, a visual salience allocation function and a color allocation set are constructed. The visual salience allocation function establishes the correspondence between raw values ​​and visual salience values. Based on this function, a curve can be generated to set different visual salience values ​​for lesion areas with different raw values. For example, the visual salience allocation function sets the visual salience value of 0.9 for lesion areas with raw values ​​of 40-60 HU and the visual salience value of 0.6 for normal liver parenchyma with raw values ​​of 60-80 HU. Through differentiated mapping, the lesion areas that doctors are concerned about achieve higher visual salience in subsequent rendering. The color allocation set establishes the correspondence between raw values ​​and a set of display color component values. The color component values ​​are usually RGB three-channel values. For example, the color value corresponding to lesion areas with raw values ​​of 40-60 HU is set to red (R=255, G=0, B=0), and the color value corresponding to normal liver parenchyma with raw values ​​of 60-80 HU is set to grayish-brown (R=180, G=120). B=80), the color value corresponding to the bone region with an original value greater than 300HU is white (R=255, G=255, B=255). Through the above color mapping, tissues with different original value ranges are presented with different colors in the final image, which makes it easier for doctors to distinguish them intuitively. By constructing a visual prominence allocation function and a color value allocation set, a clear mapping basis is provided for pixel point conversion, so that each voxel can be assigned a visual prominence value and color value corresponding to its original value, thereby realizing the differentiated display of different tissues.

[0022] In one specific embodiment, setting corresponding visual salience adjustment coefficients for different positions in three-dimensional space specifically includes the following steps: A reference axis is defined in three-dimensional space. The reference axis is perpendicular to the two-dimensional tomographic image and passes through the center of the target area. For any position in three-dimensional space, the weighted offset from the reference axis is calculated. After calculating the weighted offset of all positions, all weighted offsets are normalized to obtain the normalized standard value. The standard value is used as the visual prominence adjustment coefficient of the corresponding position.

[0023] Specifically, to address the issue that when the original values ​​of lesions and surrounding interfering tissues (such as normal liver parenchyma) are highly similar, the original value mapping alone cannot effectively distinguish between them, potentially causing the lesions to be submerged or confused in 3D images. Therefore, a spatial position dimension is introduced as a distinguishing factor. This involves setting corresponding visual prominence adjustment coefficients for different locations in 3D space. First, a reference axis is defined in 3D space. This reference axis is perpendicular to the 2D tomographic image plane and passes through the center of the target area. For example, in a liver CT image sequence, the geometric center of the liver is automatically identified based on its anatomical morphology, and a straight line parallel to the scanning direction is drawn from this center as the reference axis. The projection point of the reference axis on each 2D tomographic image is the corresponding reference center point. Then, for any location in 3D space (a subsequently generated voxel), the weighted offset from the reference axis to the reference center point of the corresponding 2D tomographic image is calculated. It is important to note that the weighted offset is not a simple straight-line distance, but a comprehensive value calculated by weighting the importance of each direction. The degree of deviation of coordinate points from the reference axis is considered. For example, in the liver region, due to the asymmetry of the anatomical structure, the importance of deviation from the center in the left-right direction (first direction) and the up-down direction (second direction) on the plane of the two-dimensional tomographic image may differ. Therefore, a weighted offset is calculated. The specific method for calculating the weighted offset will be explained in detail later. After calculating the weighted offset for all positions, all weighted offsets are normalized. The normalized labeled value is used as the visual prominence adjustment coefficient for the corresponding position. The purpose of normalization is to map the offset to a uniform numerical range (e.g., between 0 and 1), so that positions near the reference center point receive a larger adjustment coefficient, and positions far from the reference center point receive a smaller adjustment coefficient. Through the above method, differentiated processing of different positions in three-dimensional space is achieved. Even if the original values ​​of lesions and interference are exactly the same, different positions can obtain different visual prominence values ​​due to the difference in visual prominence adjustment coefficients, thereby highlighting the lesion and effectively reducing interference, solving the problem that relying solely on the original values ​​cannot distinguish similar tissues.

[0024] In one specific embodiment, calculating its weighted offset from the reference axis includes the following steps: For each position in three-dimensional space, determine the first and second projection distances from the corresponding position to the reference axis in the plane of the two-dimensional tomographic image. The first and second directions are perpendicular to each other. Determine the first and second weights. Use the first weight to weight the projection distance in the first direction and the second weight to weight the projection distance in the second direction. Combine the two weighted projection distances into a comprehensive offset.

[0025] Specifically, to address the issue of accurately quantifying the degree of pixel deviation from the reference axis, thereby providing a more precise spatial distinction between lesions and interference items with similar original values, the calculation process of the weighted offset is further explained. First, for each location in three-dimensional space, the projection distance from that location to the reference axis in the plane of the two-dimensional tomographic image is determined in the first direction (the previously mentioned left-right direction) and the second direction (the previously mentioned up-down direction). The intersection of the reference axis and the corresponding two-dimensional tomographic image is taken as the reference center point, and the first weight and the second weight are determined. The specific method for determining the first weight and the second weight will be explained in detail later. The first weight is used to adjust the first direction... The projection distances are weighted, and the projection distances in the second direction are weighted using a second weight. Assuming the weighted first projection distance is A and the weighted second projection distance is B, the formula for calculating the comprehensive offset D is D=A²+B². By calculating the comprehensive offset, even if two voxels have the same original value and the same projection distance to the reference center point, as long as they deviate from the reference axis in different directions, their comprehensive offsets will also differ due to the different weights in the two directions. This results in different visual prominence adjustment coefficients in the subsequent normalization process, leading to different final visual prominence values ​​and further improving the ability to distinguish between voxels with the same or similar original values.

[0026] In one specific embodiment, determining the first weight and the second weight specifically includes the following steps: For each location in the two-dimensional tomographic image, the boundary contour of the target part in the corresponding two-dimensional tomographic image is identified. Based on the boundary contour, the first span value and the second span value of the target part in the two-dimensional tomographic image in the first direction and the second span value in the second direction are calculated. The first weight and the second weight are determined according to the ratio of the first span value and the second span value.

[0027] Specifically, to address the issue that fixed weights cannot adapt to the anatomical differences among different patients, locations, and even different tomographic images, resulting in limited spatial discrimination ability when the original values ​​are similar, the following method is used to determine the first and second weights: For each location's two-dimensional tomographic image, the boundary contour of the target area in the two-dimensional tomographic image is identified. For example, in a CT image of the liver, the boundary of the liver region is automatically extracted using an image segmentation algorithm to obtain the precise contour of the liver on that layer of the image. Then, based on the identified boundary contour, the first span value in the first direction and the second span value in the second direction of the target area on the corresponding two-dimensional tomographic image are calculated. The first span value is the maximum width of the liver contour in the first direction, and the second span value is the maximum width of the liver in the second direction. The first weight and the second weight are determined according to the ratio of the first span value and the second span value. For example, if the first span value is 200mm and the second span value is 100mm, then the span ratio is 2: 1 means that the tissue distribution range is wider in the left-right direction, and more interfering tissue may accumulate when it deviates from the center. Therefore, the first weight is set to 2 / (2+1)=0.67, and the second weight is set to 1 / (2+1)=0.33, so that the direction with a larger span receives a higher weight. Through the above dynamic weight determination method, the calculation of the visual prominence adjustment coefficient can adaptively match the real anatomical morphology of the target part on each tomographic image. The above dynamic adaptive capability plays a key role in solving the problem of distinguishing different tissues when the original values ​​are similar. Suppose there is a lesion in the central region of the liver with an original value of 50 HU, and... There are two interference terms at the edge of the liver (with similar original values), one located at the distal end in the left-right direction (X-axis) and the other at the distal end in the front-back direction (Y-axis). Due to the different anatomical spans of different layers, the weights of the two directions also change dynamically, allowing the spatial weighting mechanism to more accurately focus on the anatomical center of the current layer. This results in lesions located in the center receiving a higher visual prominence adjustment coefficient, while the interference terms at the edge, regardless of their direction, are appropriately attenuated due to their deviation from the anatomical contour of the current layer. Ultimately, when obtaining differentiated final visual prominence values, effective differentiation is achieved when the original values ​​are similar.

[0028] In one specific embodiment, constructing the visual salience assignment function also involves performing the following steps: Select several key points, set corresponding visual prominence values ​​for the key points based on their original values, generate a smooth curve based on the key points using a difference algorithm to fit the visual prominence values ​​of the key points, obtain the function corresponding to the smooth curve as the initial visual prominence allocation function, and optimize the initial visual prominence allocation function to generate the final visual prominence allocation function.

[0029] Specifically, to address the problem that traditional methods often employ fixed, preset forms for visual salience allocation functions (such as simple linear ramps or fixed piecewise functions), which fail to adapt to individual differences among patients, scanning equipment, and lesion types, and thus may fail to adequately highlight lesions even with spatial weighting when CT values ​​are similar, the construction process of the visual salience allocation function is specifically defined to make it customizable and optimizable. First, several key points are selected. These key points can be chosen based on clinical experience and the anatomical features of the target site. For example, to highlight the lesion area, the visual salience corresponding to the lesion area is set to 0.9, air to 0, fat to 0.05, normal liver parenchyma to 0.3, and bone to 0.5. It is important to note that the correspondence between the original values ​​of the key points and their corresponding visual salience values ​​is not fixed but can be adjusted according to specific clinical conditions. The requirements can be flexibly set. For example, when it is necessary to focus on observing early lesions, the key points of the corresponding original value range can be assigned higher visual salience values. Then, based on several selected key points, a smooth curve is generated using an interpolation algorithm. The visual salience values ​​of these key points are used as the basis for the curve. Common interpolation algorithms include linear interpolation, spline interpolation, etc. For example, cubic spline interpolation is used to make the curve transition smoothly between key points, avoiding visual artifacts caused by abrupt changes. The function corresponding to the smooth curve is then used as the initial visual salience assignment function. After that, the initial visual salience function is optimized to generate the final visual salience assignment function. The specific method of optimizing the initial visual salience assignment function to generate the final visual salience assignment function will be explained in detail later. The above method, through the design of the visual salience function, enables the original value range of the tissues that need to be focused on (such as lesions) to obtain higher visual salience, while the original value range of the interference terms obtains lower visual salience, thereby improving the ability to distinguish between different tissues.

[0030] In one specific embodiment, optimizing the initial visual salience allocation function to generate the final visual salience allocation function specifically includes the following steps: The frequency of occurrence of the original values ​​corresponding to all pixels in all two-dimensional tomographic images is counted to generate an original value frequency histogram. For each original value, the frequency corresponding to the original value in the original value frequency histogram is multiplied by the visual salience value corresponding to the same original value in the initial visual salience allocation function. A visual contribution distribution histogram is generated based on the result of the multiplication operation. The peak value of the visual contribution distribution histogram is obtained as the first original value. The peak value of the curve corresponding to the initial visual prominence allocation function is also analyzed as the second original value. Based on the difference between the first and second original values, the initial prominence allocation function is adjusted. The goal of the adjustment is to reduce the difference between the recalculated first and second original values ​​until the difference between the first and second original values ​​is less than a preset threshold. The adjusted initial prominence allocation function is then used as the final visual prominence allocation function.

[0031] Specifically, to address the issue that the initial visual salience allocation function may not accurately match the current patient's data distribution, leading to insufficient highlighting of the focus tissue when raw values ​​are similar, the initial visual salience allocation function is adaptively optimized to automatically adjust based on the statistical characteristics of the actual data. First, the frequency of the raw values ​​corresponding to all pixels in all two-dimensional tomographic images is counted, generating a raw value frequency histogram. For example, in a liver CT image sequence, the number of pixels with each raw value in the range of -500HU to 1000HU is counted, resulting in a histogram with raw values ​​on the horizontal axis and frequency on the vertical axis. The histogram reflects the proportion of different tissue types in the data, such as peak regions corresponding to fat, liver parenchyma, lesions, and bone. For each raw value, the corresponding frequency value of the raw value is multiplied by the visual salience value of the same raw value corresponding to the initial visual salience allocation function. Based on the result of this multiplication operation, a frequency histogram is generated. The visual contribution distribution histogram is generated. It should be noted that the visual contribution distribution histogram reflects the potential visual influence of different original value ranges on the final rendered image under the combined effect of the current data distribution and the current visual salience allocation function. For example, if the frequency of lesions with an original value of 50 HU is high (e.g., 1000 pixels) and the visual salience value assigned by the initial function is also high (e.g., 0.8), then the product is 800, and the contribution is significant. If the frequency of normal liver parenchyma with an original value of 60 HU is even higher (e.g., 5000 pixels) but the visual salience value assigned by the initial function is low (e.g., 0.3), then the product is 1500, and the contribution is still considerable. Based on the visual salience distribution histogram, the visual influence of tissues with different original value ranges on the final rendered image under the current visual salience allocation function is predicted, and this is used as a feedback basis until the visual salience allocation function is adjusted in the direction that can highlight the tissues of most interest (e.g., lesions, liver parenchyma, etc.).

[0032] Analyze the histogram of visual contribution distribution to obtain the first raw value corresponding to the peak of the histogram. Simultaneously, analyze the curve corresponding to the initial visual salience allocation function to obtain the second raw value corresponding to the peak of that curve. For example, the peak of the visual contribution distribution histogram may appear at 52 HU (possibly corresponding to the central region of the lesion), while the peak of the initial visual salience allocation function may appear at 60 HU (possibly corresponding to the doctor's pre-defined focus point on normal liver parenchyma), resulting in a difference of 8 HU. Based on the difference between the first and second raw values, the initial visual salience allocation function is adjusted and optimized. For instance, the control points corresponding to the peak points on the initial function curve are identified and moved towards the direction of the first raw value. Simultaneously, to ensure curve smoothness, the positions of adjacent control points are also appropriately adjusted. After adjustment, a new adjusted curve is refitted to obtain a new visual salience function, and the second raw value corresponding to its peak is obtained again. Finally, the new visual contribution distribution is recalculated. A histogram is plotted, and the first original value corresponding to its peak is obtained again. The newly obtained first original value and the second original value are compared. The adjustment steps are repeated so that the difference between the first original value and the second original value obtained in each iteration gradually decreases until the difference is less than a preset threshold. It is then considered that the visual contribution allocation function has matched the current visual distribution characteristics well. Through the above optimization process, the peak of the salience visual allocation function can automatically align with the range of original values ​​with the highest visual contribution. In scenarios where the original values ​​are similar, even if the original values ​​of the organization of interest and the interference items are very close, through multiple iterations of optimization, the peak of the function will gradually move towards the region with higher visual contribution. This results in the original value range where the organization of interest is located obtaining relatively higher visual salience, while the original value range where the interference items are located has relatively lower visual salience. This works in conjunction with subsequent steps to jointly improve the ability to distinguish between the organization of interest and the interference items when the original values ​​are similar.

[0033] In one specific embodiment, when the visual contribution distribution histogram has multiple peaks, the following steps are included: Sort all peaks according to their height, select the top N peaks with the highest height, and use the original values ​​corresponding to the top N peaks as multiple first original values. Adjust the initial prominence assignment function accordingly so that the curve corresponding to the initial prominence assignment function has the shape of multiple peaks, so that the adjusted function can simultaneously highlight multiple different original value ranges.

[0034] Specifically, to address the issue of multiple tissues requiring simultaneous observation in clinical settings, such as two different lesions (tumors or calcifications), when the visual contribution distribution histogram exhibits multiple peaks, peak detection is performed after obtaining the histogram. Assuming multiple peaks exist (e.g., the image simultaneously contains early-stage small hepatocellular carcinoma with an original value of 40-50 HU and calcifications with an original value of 200-250 HU, while the CT value of normal liver parenchyma is 50-70 HU, in which case the visual contribution distribution histogram may show three peaks), peak detection is performed when the visual contribution distribution histogram exhibits multiple peaks. When there are multiple peaks, all peaks are sorted according to their height. The peak height reflects the potential visual contribution of the original value range to the final image. For example, if the peak height at 45HU is 800, the peak height at 60HU is 1500, and the peak height at 220HU is 600, then the sorting result is: 60HU (first), 45HU (second), and 220HU (third). The top N peaks with the highest heights are selected, and the original values ​​corresponding to each of the top N peaks are used as multiple first original values. The value of N can be preset according to clinical needs. For example, if N=2, then... 60HU and 45HU were selected as the two initial raw values. The initial visual salience allocation function was adjusted accordingly to give the curve of the adjusted function a multi-peak shape. This allows the adjusted visual salience allocation function to simultaneously highlight multiple different raw value ranges. The specific adjustment method is as follows: Assuming that the initial visual salience allocation function originally had only one peak at 60HU (corresponding to normal liver parenchyma), and now it is necessary to simultaneously highlight small hepatocellular carcinomas at 45HU, a new control point is added at 45HU and assigned a higher visual salience value (e.g., 0.9). At the same time, the visual salience value of the control point at 60HU is adjusted (it can be appropriately reduced to 0.5 to avoid over-highlighting normal tissue). A smooth double-peak curve is generated by spline interpolation, so that there is a local maximum near both 45HU and 60HU. Through the above multi-peak processing, the final generated visual salience allocation function can simultaneously focus on multiple raw value ranges. This method avoids the problem that a single-peak function can only highlight one type of tissue and ignore multiple tissues that need attention, thus solving the clinical need to observe multiple tissues together when the raw values ​​are similar.

[0035] According to another aspect of the embodiments of the present invention, reference is made to... Figure 2 As shown, a computer vision-based medical image processing system is also provided, comprising modules for implementing the computer vision-based medical image processing method described above, with the specific functions of each module as follows: The data acquisition module is used to acquire continuous two-dimensional tomographic images. The two-dimensional tomographic images are obtained by scanning the target part of the human body layer by layer. They are composed of multiple pixels, and each pixel corresponds to an original value. The rule building module is used to build transformation rules. The transformation rules are used to map the original values ​​of different ranges to the corresponding color values ​​and visual prominence values, define the three-dimensional space and set the corresponding visual prominence adjustment coefficients for different positions in the three-dimensional space. The three-dimensional space is composed of all two-dimensional images stacked and aligned. The data conversion module is used to determine the corresponding color value and visual prominence value for each pixel on each two-dimensional image based on the original value and conversion rules, determine the corresponding visual prominence adjustment coefficient based on the position of the pixel in three-dimensional space, and multiply the visual prominence adjustment coefficient by the corresponding visual prominence value to obtain the final visual prominence value. The projection generation module is used to construct a corresponding voxel in three-dimensional space based on the position of the pixel, assign the color value and final visual prominence value of the pixel to the corresponding voxel, generate a three-dimensional model composed of voxels, simulate the projection of the three-dimensional model from a preset viewing direction, and calculate and generate a two-dimensional visualization image.

[0036] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0037] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0038] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions 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 this application.

Claims

1. A medical image processing method based on computer vision, characterized in that, The method includes: Step S1: Obtain continuous two-dimensional tomographic images. The two-dimensional tomographic images are obtained by scanning the target part of the human body layer by layer. They are composed of multiple pixels, and each pixel corresponds to an original value. Step S2: Define a three-dimensional space, which is composed of all two-dimensional images stacked and aligned. Construct transformation rules, which are used to map the original values ​​of different ranges to the corresponding color values ​​and visual prominence values. Based on the transformation rules, set the corresponding visual prominence adjustment coefficients for different positions in the three-dimensional space. Step S3: For each pixel in each two-dimensional image, determine the corresponding color value and visual prominence value according to the original value and conversion rules. Determine the corresponding visual prominence adjustment coefficient according to the position of the pixel in three-dimensional space. Multiply the visual prominence adjustment coefficient by the corresponding visual prominence value to obtain the final visual prominence value. Step S4: Construct a corresponding voxel in three-dimensional space based on the position of the pixel, assign the color value and final visual prominence value of the pixel to the corresponding voxel, generate a three-dimensional model composed of voxels, simulate projection of the three-dimensional model from a preset viewing direction, calculate and generate a two-dimensional visualization image.

2. The method according to claim 1, characterized in that, Construct rules for color and opacity conversion, including: A visual prominence allocation function and a color value allocation set are constructed. The visual prominence allocation function is used to establish the correspondence between the original values ​​and the visual prominence values, and the color value allocation set is used to establish the correspondence between the original values ​​and a set of display color component values.

3. The method according to claim 1, characterized in that, Assign corresponding visual prominence adjustment coefficients to different locations in three-dimensional space, including: A reference axis is defined in three-dimensional space. The reference axis is perpendicular to the two-dimensional tomographic image and passes through the center of the target area. For any position in three-dimensional space, the weighted offset from the reference axis is calculated. After calculating the weighted offset of all positions, all weighted offsets are normalized to obtain the normalized standard value. The standard value is used as the visual prominence adjustment coefficient of the corresponding position.

4. The method according to claim 3, characterized in that, Calculate its weighted offset from the reference axis, including: For each position in three-dimensional space, determine the first and second projection distances from the corresponding position to the reference axis in the plane of the two-dimensional tomographic image. The first and second directions are perpendicular to each other. Determine the first and second weights. Use the first weight to weight the projection distance in the first direction and the second weight to weight the projection distance in the second direction. Combine the two weighted projection distances into a comprehensive offset.

5. The method according to claim 4, characterized in that, Determining the first and second weights includes: For each location in the two-dimensional tomographic image, the boundary contour of the target part in the corresponding two-dimensional tomographic image is identified. Based on the boundary contour, the first span value and the second span value of the target part in the two-dimensional tomographic image in the first direction and the second span value in the second direction are calculated. The first weight and the second weight are determined according to the ratio of the first span value and the second span value.

6. The method according to claim 2, characterized in that, Construct a visual salience assignment function, including: Select several key points, set corresponding visual prominence values ​​for the key points based on their original values, generate a smooth curve based on the key points using a difference algorithm to fit the visual prominence values ​​of the key points, obtain the function corresponding to the smooth curve as the initial visual prominence allocation function, and optimize the initial visual prominence allocation function to generate the final visual prominence allocation function.

7. The method according to claim 6, characterized in that, The initial visual salience assignment function is optimized to generate the final visual salience assignment function, including: The frequency of occurrence of the original values ​​corresponding to all pixels in all two-dimensional tomographic images is counted to generate an original value frequency histogram. For each original value, the frequency corresponding to the original value in the original value frequency histogram is multiplied by the visual salience value corresponding to the same original value in the initial visual salience allocation function. A visual contribution distribution histogram is generated based on the result of the multiplication operation. The peak value of the visual contribution distribution histogram is obtained as the first original value. The peak value of the curve corresponding to the initial visual prominence allocation function is also analyzed as the second original value. Based on the difference between the first and second original values, the initial prominence allocation function is adjusted. The goal of the adjustment is to reduce the difference between the recalculated first and second original values ​​until the difference between the first and second original values ​​is less than a preset threshold. The adjusted initial prominence allocation function is then used as the final visual prominence allocation function.

8. The method according to claim 7, characterized in that, Obtain the first raw value corresponding to the peak of the visual contribution distribution histogram, including: When there are multiple peaks in the visual contribution distribution histogram, all peaks are sorted according to their height. The top N peaks with the highest height are selected, and the original values ​​corresponding to the top N peaks are used as multiple first original values. The initial prominence allocation function is adjusted accordingly so that the curve corresponding to the initial prominence allocation function has the shape of multiple peaks, so that the adjusted function can simultaneously highlight multiple different original value ranges.

9. A computer vision-based medical image processing system for implementing the computer vision-based medical image processing method as described in any one of claims 1-8, characterized in that, Includes the following modules: The data acquisition module is used to acquire continuous two-dimensional tomographic images. The two-dimensional tomographic images are obtained by scanning the target part of the human body layer by layer. They are composed of multiple pixels, and each pixel corresponds to an original value. The rule building module is used to build transformation rules. The transformation rules are used to map the original values ​​of different ranges to the corresponding color values ​​and visual prominence values, define the three-dimensional space and set the corresponding visual prominence adjustment coefficients for different positions in the three-dimensional space. The three-dimensional space is composed of all two-dimensional images stacked and aligned. The data conversion module is used to determine the corresponding color value and visual prominence value for each pixel on each two-dimensional image based on the original value and conversion rules, determine the corresponding visual prominence adjustment coefficient based on the position of the pixel in three-dimensional space, and multiply the visual prominence adjustment coefficient by the corresponding visual prominence value to obtain the final visual prominence value. The projection generation module is used to construct a corresponding voxel in three-dimensional space based on the position of the pixel, assign the color value and final visual prominence value of the pixel to the corresponding voxel, generate a three-dimensional model composed of voxels, simulate the projection of the three-dimensional model from a preset viewing direction, and calculate and generate a two-dimensional visualization image.

Images