A CT-based L3 layer plane anatomical structure automatic segmentation and quantitative quality control method and system
Patent Information
- Application Number
- CN202611192819.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-07
- Publication Date
- 2026-10-30
AI Technical Summary
[0008]本发明的目的在于克服现有技术的不足，提供一种基于CT的L3层面解剖结构自动分割与量化质控方法及系统，以解决三维切片定位难、多结构联合分割精度低、缺乏自适应风险评估以及归档易崩溃的技术问题
[0010]本发明的有益效果在于：第一，通过“三维椎体定位+Z轴中点提取”的策略，实现了中心切片的100%全自动截取，并结合多目标神经网络完成了对骨、肌、脂组织的高精度边缘分割；第二，自动将像素面积转化为通用的临床标准化指数（如SMI、SATI），并引入多维度临床截断阈值，为代谢紊乱及肌少症提供了客观的评估依据；第三，独创的底层空字符洗消机制根治了底层接口的写盘崩溃隐患；第四，通过为图像与文档分配独立的序列标识符，完全符合DICOM协议规范，大幅提升了在主流PACS系统中的工程落地稳定性。
Smart Images
Figure CN123015451A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the fields of medical image processing, artificial intelligence and quantitative body composition analysis technology, specifically involving an automatic segmentation and quantitative quality control method and system for L3 level anatomical structures based on CT. Background Technology
[0002] Abdominal CT images at the L3 level are internationally recognized as the gold standard for quantitative assessment of body composition. By measuring skeletal muscle, visceral fat (VAT), subcutaneous fat (SAT), and vertebral cancellous bone mineral density at the L3 level, effective early screening for osteoporosis, sarcopenia, and cardiovascular metabolic risks can be achieved.
[0003] Existing body composition analysis methods have the following four main technical shortcomings.
[0004] First, manual selection and segmentation are inefficient. Traditional methods rely on physicians manually searching for L3 center slices layer by layer and manually delineating the region, which is highly subjective and difficult to promote in large-scale clinical scenarios.
[0005] Second, the spatial positioning cascade mechanism is lacking. Existing two-dimensional segmentation models cannot directly process three-dimensional scan data of the entire abdomen and lack the ability to automatically locate from three-dimensional space to the two-dimensional center slice.
[0006] Third, there is a lack of comprehensive indicator extraction and automatic quality control. Most software only calculates a single indicator and lacks a comprehensive quality control and early warning mechanism that combines a core index normalized to height and targets cancellous bone mineral density.
[0007] Fourth, clinical archiving has extremely poor compatibility. When medical images are transmitted between different systems, non-standard metadata (such as trailing null characters \x00) can easily cause errors in the underlying data interface type recognition and lead to crashes. At the same time, using the same sequence identifier for images and reports often causes stringent Picture Archiving Systems (PACS) to reject data due to modal conflict checks. Summary of the Invention
[0008] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method and system for automatic segmentation and quantitative quality control of L3-level anatomical structures based on CT, so as to solve the technical problems of difficult three-dimensional slice localization, low accuracy of multi-structure joint segmentation, lack of adaptive risk assessment and easy collapse of archiving.
[0009] To address the aforementioned technical problems, this invention provides an automatic segmentation and quantitative quality control method for L3-level anatomical structures based on CT, comprising: a three-dimensional image analysis and cascaded localization step, used to acquire the three-dimensional medical image to be analyzed, inputting it into a first-stage neural network to generate a three-dimensional mask of the L3 vertebral body, and extracting the spatial coordinate set of the mask to calculate the Z-axis center slice index to extract two-dimensional slices; a multi-target anatomical structure fine segmentation step, used to input into a second-stage neural network to generate a classification segmentation mask containing specific tissue types such as bone, muscle, and fat; a physical cross-sectional area and normalized index extraction step, used to calculate the physical cross-sectional area based on pixel spacing, and extract the normalized index and fat area ratio in combination with height parameters; an adaptive risk stratification assessment step, used to compare the quantitative numerical indicators with preset clinical cutoff thresholds to generate a risk stratification status report; and a metadata cleaning and dual protocol encapsulation step, used to perform illegal null character cleanup operations when traversing the original image label metadata, and assign independent sequence identifiers to the image and report respectively for network push.
[0010] The beneficial effects of this invention are as follows: First, by employing a strategy of "three-dimensional vertebral body localization + Z-axis midpoint extraction," 100% fully automated truncation of central slices is achieved, and high-precision edge segmentation of bone, muscle, and fat tissues is completed by combining a multi-objective neural network; Second, pixel area is automatically converted into a universal clinical standardized index (such as SMI, SATI), and a multi-dimensional clinical cutoff threshold is introduced, providing an objective assessment basis for metabolic disorders and sarcopenia; Third, the unique underlying null character eradication mechanism eliminates the hidden danger of disk writing crashes in the underlying interface; Fourth, by assigning independent sequence identifiers to images and documents, it fully complies with the DICOM protocol specification, significantly improving the stability of engineering implementation in mainstream PACS systems. Attached Figure Description
[0011] Figure 1 This is a flowchart illustrating an automatic segmentation and quantitative quality control method for L3-level anatomical structures based on CT, according to an embodiment of this disclosure.
[0012] Figure 2 This is a block diagram of a CT-based L3-level anatomical structure automatic segmentation and quantification quality control system according to an embodiment of the present disclosure.
[0013] Figure labeling: 100: Automatic segmentation and quantitative quality control system for L3-level anatomical structures based on CT; 110: 3D image analysis and cascaded localization module; 120: Fine segmentation module for multi-target anatomical structures; 130: Physical cross-sectional area and normalized index extraction module; 140: Adaptive risk stratification assessment module; 150: Metadata cleaning and dual protocol encapsulation module. Detailed Implementation
[0014] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that these embodiments are provided to enable those skilled in the art to understand and reproduce the present invention, and are not intended to limit the scope of protection of the present invention.
[0015] like Figure 1 and Figure 2 In summary, this invention provides a CT-based automatic segmentation and quantification quality control system 100 for L3-level anatomical structures, used to execute corresponding CT-based automatic segmentation and quantification quality control methods for L3-level anatomical structures. The system 100 is constructed from a computer device including a processor and memory, based on software programs, and specifically includes: a 3D image analysis and cascaded localization module 110, a multi-target anatomical structure fine segmentation module 120, a physical cross-sectional area and normalization index extraction module 130, an adaptive risk stratification assessment module 140, and a metadata cleaning and dual-protocol encapsulation module 150. The collaborative working logic and specific method steps of each module are as follows.
[0016] S110, the 3D image analysis and cascaded localization module 110 performs the 3D image analysis and cascaded localization steps: acquiring the 3D medical image to be analyzed, inputting it into a pre-trained first-stage neural network to generate a coarse 3D mask of the L3 vertebral body. Extracting the set of 3D coordinates V of the effective voxels in the 3D mask of the L3 vertebral body output by the first-stage neural network. L3 ={x i ,y i ,z i Extract the subset of coordinates Z along the Z-axis of this set. coords Calculate the median of the values in this subset of coordinates as the Z-axis center slice index (i.e., Z). center =round (median (Z coords In the original three-dimensional data array, the corresponding two-dimensional matrix is directly extracted from the extracted index slices, which serves as a precise two-dimensional L3 anatomical center slice, thereby completing the efficient dimensionality reduction and localization from three-dimensional space to two-dimensional target area.
[0017] S120, the multi-target anatomical structure fine segmentation module 120 performs the multi-target anatomical structure fine segmentation step: the two-dimensional L3 anatomical center slice is used as a tensor input to the second-stage neural network for multi-target joint inference. The specific tissue types for network inference are defined as including four key tissue structures: cancellous bone, subcutaneous fat, visceral fat, and skeletal muscle. Through the forward propagation of this neural network, high-precision classification and segmentation masks corresponding to these four tissue types are output in parallel, achieving one-time edge delineation of the core structures of the abdominal CT scan.
[0018] S130, the physical cross-sectional area and normalized index extraction module 130 performs the physical cross-sectional area and normalized index extraction steps: It reads the physical pixel spacing (Δx and Δy) of the two-dimensional slice in the X-axis and Y-axis directions. For each specific anatomical structure, it counts the total number of effective pixels N within its classification mask. k Through the multiplication algorithm (Area) k =N k The actual physical cross-sectional area of each anatomical structure is calculated using the formula (×Δx×Δy). Then, combined with the patient's height data obtained from the electronic medical record system, the physical cross-sectional area of each tissue is divided by the square of the height (Index). k =Area k / Height 2 The system calculates and generates clinically accepted skeletal muscle index (SMI), subcutaneous fat index (SATI), and visceral fat index (VATI). Finally, the physical cross-sectional area of visceral fat is divided by the physical cross-sectional area of subcutaneous fat to generate the core indicator for assessing metabolic disorders—the ratio of visceral to subcutaneous fat area.
[0019] S140, the adaptive risk stratification assessment module 140 executes the adaptive risk stratification assessment steps: extracting the quantitative numerical indicators of each anatomical structure calculated in the above steps and comparing them with the built-in preset clinical multicenter cutoff thresholds. Specific logic includes: automatically triggering a suspected osteoporosis alarm when the average pixel density in the cancellous bone region is less than 110 HU; automatically triggering a muscle loss alarm when the skeletal muscle index in male patients is less than 52.4 cm² / m² or in female patients is less than 38.5 cm² / m²; and automatically triggering a high-risk cardiovascular and metabolic alarm when the ratio of visceral to subcutaneous fat area is greater than or equal to 1.0. Based on the above comparison results, the system generates a quantitative assessment conclusion document containing the comprehensive risk stratification status.
[0020] S150, the metadata cleaning and dual-protocol encapsulation module 150 performs metadata cleaning and dual-protocol encapsulation steps: generating a DICOM report file containing an intuitive visual image overlay and quantitative evaluation conclusions. When constructing the exported file and traversing the original image metadata (such as patient name, patient ID, and sequence description fields), regular expression matching is used to identify and forcibly remove hidden null bytes (such as \x00) at the end of the string to completely avoid underlying disk write crashes caused by illegal characters. During the protocol encapsulation stage, data modalities are strictly distinguished: the category attribute of the generated secondary capture image is set to "OT" and a first independent sequence identifier (SeriesInstanceUID) is assigned; the category attribute of the report file is set to "DOC" and a second independent sequence identifier different from the first identifier is assigned. Finally, the two sets of independent sequences are mounted under the same examination instance (StudyInstanceUID) and securely pushed to the hospital's Picture Archiving and Communication System (PACS) based on the C-STORE protocol, thus cleverly circumventing the PACS's stringent modal conflict verification mechanism.
[0021] This disclosure also provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, it implements all steps S110 to S150 of the CT-based L3-level anatomical structure automatic segmentation and quantitative quality control method as detailed above, and achieves the same technical effect.
[0022] The above description is merely an embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.
Claims
1. A method for automatic segmentation and quantitative quality control of L3-level anatomical structures based on CT, characterized in that, The CT-based automatic segmentation and quantitative quality control method for L3 level anatomical structures includes: The three-dimensional image analysis and cascaded localization steps are used to obtain the three-dimensional medical image to be analyzed, input it into the first-stage neural network to generate a three-dimensional mask of the L3 vertebral body, extract the spatial coordinate set of the mask to calculate the Z-axis center slice index, and extract the corresponding two-dimensional L3 anatomical center slice from the three-dimensional medical image. The multi-target anatomical structure fine segmentation step is used to input the two-dimensional L3 anatomical center slice into the second-stage neural network for multi-target inference, and generate a classification segmentation mask containing specific tissue types; The physical cross-sectional area and normalization index extraction steps are used to calculate the physical cross-sectional area of each anatomical structure in the classification segmentation mask based on the pixel spacing of the two-dimensional L3 anatomical center slice, and to extract the normalization index and fat area ratio in combination with the patient's physiological parameters. An adaptive risk stratification assessment step is used to extract quantitative numerical indicators for each anatomical structure, compare the quantitative numerical indicators with preset clinical cutoff thresholds, and generate a quantitative assessment conclusion that includes the risk stratification status. Metadata cleaning and dual protocol encapsulation steps are used to generate a report file containing the visualized image and the quantitative evaluation conclusion. Illegal null characters are cleaned up when traversing the original image tag metadata, and independent sequence identifiers are assigned to the generated secondary captured image and the report file to perform network push.
2. The method for automatic segmentation and quantitative quality control of L3-level anatomical structures based on CT according to claim 1, characterized in that, The three-dimensional image analysis and cascaded localization steps include: extracting the set of three-dimensional coordinates of effective voxels in the three-dimensional mask of the L3 vertebral body output by the first-stage neural network; extracting a subset of the coordinates in the Z-axis direction of the three-dimensional coordinate set, and calculating the median of the values of the subset as the Z-axis center slice index; and extracting the two-dimensional matrix corresponding to the Z-axis center slice index from the original three-dimensional data array as the two-dimensional L3 anatomical center slice.
3. The method for automatic segmentation and quantitative quality control of L3-level anatomical structures based on CT according to claim 1, characterized in that, The multi-target anatomical structure fine segmentation step includes: setting the specific tissue type to include cancellous bone, subcutaneous fat, visceral fat and skeletal muscle; and outputting the classification segmentation mask corresponding to the cancellous bone, subcutaneous fat, visceral fat and skeletal muscle in parallel through the second-stage neural network.
4. The method for automatic segmentation and quantitative quality control of L3 level anatomical structures based on CT according to claim 1, characterized in that, The steps for extracting the physical cross-sectional area and normalized index include: calculating the physical cross-sectional area of each anatomical structure by multiplying the pixel spacing in the X and Y axes of the two-dimensional slice by the total number of effective pixels; dividing the physical cross-sectional area by the square of the patient's height data to generate the skeletal muscle index, subcutaneous fat index, and visceral fat index; and dividing the physical cross-sectional area of visceral fat by the physical cross-sectional area of subcutaneous fat to generate the ratio of visceral to subcutaneous fat area.
5. The method for automatic segmentation and quantitative quality control of L3 level anatomical structures based on CT according to claim 1, characterized in that, The adaptive risk stratification assessment steps include: triggering an osteoporosis abnormality alarm when the average pixel density in the cancellous bone region is less than 110 HU; triggering a muscle loss abnormality alarm when the skeletal muscle index is less than 52.4 cm² / m² in male patients or less than 38.5 cm² / m² in female patients; and triggering a cardiovascular and metabolic high-risk alarm when the ratio of visceral to subcutaneous fat area is greater than or equal to 1.
0.
6. The method for automatic segmentation and quantitative quality control of L3-level anatomical structures based on CT according to claim 1, characterized in that, The metadata cleaning and dual protocol encapsulation steps include: traversing the patient name, patient ID, and sequence description fields in the original image metadata, identifying and removing hidden null bytes at the end of the string; setting the category attribute of the generated secondary capture image to OT and assigning a first independent sequence identifier; setting the category attribute of the report file to DOC and assigning a second independent sequence identifier different from the first independent sequence identifier, and mounting it under the same examination instance and pushing it to the image archiving and communication system.
7. A CT-based automatic segmentation and quantitative quality control system for L3-level anatomical structures, characterized in that, The CT-based L3-level anatomical structure automatic segmentation and quantification quality control system includes: a three-dimensional image analysis and cascaded localization module for performing the three-dimensional image analysis and cascaded localization steps as described in claim 1; a multi-target anatomical structure fine segmentation module for performing the multi-target anatomical structure fine segmentation steps as described in claim 1; a physical cross-sectional area and normalized index extraction module for performing the physical cross-sectional area and normalized index extraction steps as described in claim 1; an adaptive risk stratification assessment module for performing the adaptive risk stratification assessment steps as described in claim 1; and a metadata cleaning and dual protocol encapsulation module for performing the metadata cleaning and dual protocol encapsulation steps as described in claim 1.
8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by the processor, it implements the method as described in any one of claims 1 to 6.