A method for automatically analyzing CT images of mouse thoracic and lumbar vertebrae

Abstract

The application discloses a kind of mouse thoracolumbar vertebra CT image automatic analysis method, comprising the following steps: S1, to the original mouse thoracolumbar vertebra three-dimensional CT image is carried out automatically cutting processing;S2, to the CT image that cutting is completed in step S1 is executed automatically rotation correction processing, standardization vertebra posture;S3, to the CT slice after rotation correction carries out vertebra slice feature identification, single vertebra section division, vertebra type determination and ordered sorting;S4, based on effective slice set, complete vertebra trabecula accurate segmentation;S5, based on trabecula segmentation mask automatic batch calculation bone quantitative morphological parameter and bone density.The application, realize mouse thoracolumbar vertebra CT image from pre-processing, posture correction, vertebra automatic segmentation and typing, trabecula deep learning segmentation to bone quantitative parameter calculation Full-process automatic processing, need not artificial correction image, manually divide vertebra boundary and delineate analysis area, can batch processing experimental sample.

Description

Technical Field

[0001] This invention relates to the field of CT image analysis technology, specifically an automatic analysis method for CT images of the thoracic and lumbar vertebrae of mice. Background Technology

[0002] In the field of animal model research on osteoporosis, bone metabolism, and degenerative bone diseases, mouse thoracic and lumbar spine Micro-CT three-dimensional imaging is the core detection method for evaluating bone microstructure and bone mass levels. However, existing mouse vertebral CT analysis protocols have many limitations, which restrict the detection efficiency and data accuracy of large-scale experimental samples.

[0003] First, the existing analysis process has a low degree of automation and relies heavily on manual operation. The original 3D CT images contain large areas of air, soft tissue and other irrelevant backgrounds, requiring manual selection of the skeletal area for cropping. Due to differences in the mouse's positioning posture, the vertebrae often tilt or deflect, requiring manual adjustment of the image rotation angle to unify the posture. This manual operation is time-consuming and labor-intensive, and subjective operation introduces random errors, which is not conducive to the comparison of multiple samples.

[0004] Secondly, traditional methods struggle to automatically distinguish and segment multiple continuous thoracic and lumbar vertebrae. In mice, the T13 thoracic vertebra is closely connected to the L1-L6 lumbar vertebrae, with overlapping and blurred boundaries in the images of adjacent vertebral junctions. Existing algorithms cannot accurately identify individual vertebral segments, relying heavily on manual segmentation of each image. Furthermore, they cannot automatically identify thoracic and lumbar vertebrae types and generate ordered sequences, resulting in an extremely large workload for batch sample processing. Moreover, existing methods do not effectively screen the slices; all coronal slices are used for trabecular bone calculations. Junctional slices with unstable structures at both ends can interfere with trabecular bone index measurements, reducing the reliability of the quantitative results.

[0005] Furthermore, the segmentation accuracy of trabecular bone is insufficient. The general medical segmentation network is not optimized for the slender trabecular bone structure of mice, and fine bone texture and tissue boundaries are easily lost, making it difficult to simultaneously distinguish multiple structures such as cancellous bone, cortical bone, bone marrow, and soft tissue. The general model lacks training on a mouse thoracic and lumbar spine-specific dataset, and the segmentation mask has problems with blurred edges and region confusion, which directly causes deviations in the subsequent calculation of bone morphology parameters.

[0006] Finally, existing tools are fragmented and cannot complete the output of a full set of indicators in one stop. Preprocessing, vertebral segmentation, and parameter calculation require switching between multiple software programs and operating in steps. They can only output some bone evaluation parameters separately, making it difficult to simultaneously obtain bone mass, trabecular microstructure, cortical bone integrity morphological indicators, and bone mineral density. The data export and integration process is cumbersome, which significantly delays the progress of scientific research data statistics. Summary of the Invention

[0007] The purpose of this invention is to provide an automatic analysis method for CT images of the thoracic and lumbar vertebrae of mice, so as to solve the problems in the prior art.

[0008] To achieve the above objectives, the present invention provides the following technical solution: an automatic analysis method for CT images of the thoracic and lumbar vertebrae of mice, comprising the following steps: S1. Automatic cropping of the original three-dimensional CT images of the rat's thoracic and lumbar spine; S2. Perform automatic rotation correction on the CT images cropped in step S1 to standardize the vertebral placement posture; S3. Perform vertebral slice feature recognition, single vertebral segment division, vertebral type determination and ordered sorting on the CT slices after rotation correction; S4. Accurate segmentation of vertebral trabeculae based on the effective slice set; S5. Automatic batch calculation of quantitative morphological parameters and bone density of bones based on trabecular segmentation mask.

[0009] Preferably, in step S1, the automatic cropping process includes the following steps: S101. Downsample the original CT images; S102. The Otsu maximum inter-class variance threshold segmentation algorithm is used to perform binarization processing on the downsampled image to distinguish between the skeletal foreground and the non-skeletal background. S103. Extract the complete outer contour boundary of the skeleton based on the binarized image, crop the image based on the skeleton boundary, remove irrelevant background areas other than the skeleton, and obtain a cropped CT image that retains only the skeleton area.

[0010] Preferably, in step S2, the automatic rotation correction process includes the following steps: S201. Resample the cropped CT images; S202. The downsampled image is binarized using Otsu thresholding to obtain a skeleton binary mask; S203. Perform PCA principal component analysis on the binary mask of the skeleton to extract the principal direction feature vector of the overall skeleton contour. S204. Calculate the affine rotation matrix of the image based on the principal direction feature vector of the skeleton, and use the rotation matrix to rotate and transform all CT slices to make the vertebrae uniformly aligned, which facilitates the recognition and quantitative analysis of two-dimensional slice features.

[0011] Preferably, step S3 specifically includes the following steps: S301. Vertebral slice feature classification and identification: The coronal slice along the CT scan axis divides a single complete vertebra into three types of feature slices: Side feature slice, Mid feature slice, and Mid-Down feature slice. Among them, the Side feature slice is a slice of the uppermost and lowermost ends of a single vertebra, in which the vertebral structure is divided into three parts in the image; the Mid feature slice is a slice of the middle section of a single vertebra, in which the vertebra presents a complete closed circular structure; the Mid-Down feature slice is a slice of the middle and lower sections of a single vertebra, in which the main body of the vertebra is divided into two parts and there is a clear cavity in the lower region; the fixed order of the slice arrangement of a single vertebra along the scanning axis is Side→Mid→Mid-Down→Side; S302. Divide the single vertebral segment intervals and select effective slices for trabecular bone analysis: Define the complete axial range of each vertebral segment based on the continuous slice index intervals corresponding to various feature slices; select the Mid feature slice index interval and the Mid-Down feature slice index interval and merge them as a set of effective slices for quantitative analysis of trabecular bone; Side feature slices are not involved in the calculation of trabecular bone parameters. Define the set of contiguous slices as , recorded as Taking a CT image containing three vertebrae—T13 (thoracic), L1 (lumbar), and L2 (lumbar)—with slice numbers ranging from 1 to 892 as an example, the coronal slice corresponding to T13 contains four continuous intervals: the upper side interval... Mid section Mid-Down (lower segment) Lower Side section The lower side section of T13; The two vertebral segments overlap with the adjacent upper L1 side slice, and the boundary between the two vertebrae cannot be distinguished by the side feature alone. S303. Automatically determine vertebral type and generate ordered vertebral sequence: compare the starting index of the first vertebral mid slice. Mid-Down slice start index If satisfied < If the first segment of the image is not found, the first vertebra is determined to be the thoracic vertebra T13; otherwise, the first vertebra is determined to be the lumbar vertebra L2. Based on the axial sequence of the Mid and Mid-Down slices corresponding to each vertebra, the system automatically distinguishes between thoracic and lumbar vertebrae and outputs an ordered vertebral number sequence, which includes T13, L1, L2, L3, L4, L5, and L6. When there is an overlap at the junction of two adjacent vertebrae and the Side slice cannot distinguish their affiliation, the vertebral boundary is determined solely by their independent Mid and Mid-Down slice intervals.

[0012] Preferably, in S4, Mid-Down feature slice interval The two types of interval slices are merged as the input image for single vertebral segmentation. A U-shaped symmetric fully convolutional deep learning segmentation network is used to perform trabecular bone instance segmentation on the slice frame by frame. The segmentation output is a multi-region mask to distinguish trabecular bone, cortical bone, bone marrow and soft tissue respectively.

[0013] Preferably, in step S5, the quantitative morphological parameters of the skeleton are divided into three categories: bone mass and total volume parameters, trabecular microstructure parameters, and cortical bone structure parameters. The bone mass and total volume parameters include at least the BV / TV bone volume-tissue volume ratio, TV tissue volume, BS bone surface area, and BS / BV specific surface area. The trabecular microarchitecture parameters include at least the number of Tb.N trabeculae, the thickness of Tb.Th trabeculae, the separation of Tb.Sp trabecular bone, the trabecular pattern factor of TBPf, and the connection density of Conn.D. The cortical bone structural parameters include at least Ct.Th cortical bone thickness, Ct.Po cortical bone porosity, and Ct.Ar cortical bone area.

[0014] Preferably, the U-shaped symmetric fully convolutional deep learning segmentation network is a 21-Net network. The 21-Net network includes an encoder downsampling module, a decoder upsampling module, and a skip connection structure. The encoder extracts shallow and deep features of the skeleton, the decoder restores the original size of the image, and the skip connections preserve the fine boundary details of the trabeculae, thereby achieving high-precision trabeculae segmentation.

[0015] Preferably, in step S203, the PCA principal component analysis uses the coordinates of all bone pixels within the binary bone mask as input data, and extracts the principal direction vector with the largest variance through linear transformation dimensionality reduction. The principal direction vector represents the overall extension direction of the vertebra.

[0016] Preferably, CT samples containing thoracic vertebrae T13, lumbar vertebrae L1, and lumbar vertebrae L2, with a slice range of 1–892, are used for network training; wherein the T13 training slice range is [missing information]. L1 training slice interval is + L2 training slice interval is + The U-shaped symmetric fully convolutional deep learning segmentation network is trained using the sliced ​​data within the aforementioned merged interval as the training set.

[0017] Compared with existing technologies, the beneficial effects of this invention are: It achieves fully automated processing of mouse thoracic and lumbar spine CT images, from preprocessing, posture correction, automatic vertebral segmentation and classification, deep learning segmentation of trabecular bone, to quantitative parameter calculation of bones. This eliminates the need for manual image correction, manual division of vertebral boundaries, and delineation of analysis regions, enabling batch processing of experimental samples and significantly improving the analysis efficiency of animal skeletal experiments. Automatic image rotation correction is achieved through Otsu threshold segmentation combined with PCA principal component analysis, unifying the vertebral posture of different samples, eliminating recognition errors caused by camera tilt, and improving algorithm robustness. A unique three-slice feature recognition system (Side, Mid, and Mid-Down) is created, discarding the overlapping and unstable features of the Side slice. It accurately defines the range of a single vertebra based on the non-interfering Mid and Mid-Down intervals, and automatically distinguishes between thoracic and lumbar vertebrae based on the difference in the starting index of the two types of slices, generating an ordered vertebral sequence. This effectively solves the industry pain points of difficulty in distinguishing vertebral boundaries and the inability to automatically identify thoracic and lumbar vertebrae. Only valid slices are selected for trabecular bone quantification calculations, reducing the amount of computation on invalid data and minimizing computational power consumption while avoiding parameter deviations caused by invalid slices, thus balancing computational speed and detection accuracy. A self-developed 21-Net U-shaped symmetric fully convolutional network adapted to mouse vertebral images is used for trabecular bone segmentation. Relying on the encoder, decoder, and skip connection structure, the fine boundaries of trabecular bone are fully preserved. It can simultaneously distinguish multiple tissues such as cancellous bone, cortical bone, bone marrow, and soft tissue. It has been specifically trained based on a real mouse vertebral slice dataset, and its segmentation accuracy is superior to general models. The entire method can automatically and batch output three major categories of complete skeletal morphological parameters and bone density based on the segmentation mask: total bone volume, trabecular bone microstructure, and cortical bone structure. It meets the scientific research evaluation needs of osteoporosis and bone metabolism-related animal experiments in one stop. At the same time, the preliminary cropping and correction steps adopt lightweight image processing algorithms such as downsampling and Otsu segmentation to remove redundant background and compress data volume in advance, reducing the computational power requirements of the entire analysis process and making it more adaptable to hardware. Attached Figure Description

[0018] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of the present invention; Figure 2 This is the Side feature diagram of the present invention; Figure 3 This is the Mid feature map of the present invention; Figure 4 This is the Mid-Down feature map of the present invention; Figure 5 These are example diagrams of the thoracic vertebra T13, lumbar vertebra L1, and lumbar vertebra L2 of the present invention; Figure 6 This is a schematic diagram of trabecular bone segmentation according to the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention.

[0020] Please see Figure 1-6 In this embodiment of the invention, an automatic analysis method for CT images of the thoracic and lumbar vertebrae of mice includes the following steps: S1. Automatic cropping of the original three-dimensional CT images of the rat's thoracic and lumbar spine; S2. Perform automatic rotation correction on the CT images cropped in step S1 to standardize the vertebral placement posture; S3. Perform vertebral slice feature recognition, single vertebral segment division, vertebral type determination and ordered sorting on the CT slices after rotation correction; S4. Accurate segmentation of vertebral trabeculae based on the effective slice set; S5. Automatic batch calculation of quantitative morphological parameters and bone density of bones based on trabecular segmentation mask.

[0021] Preferably, in step S1, the automatic cropping process includes the following steps: S101. Downsample the original CT images; S102. The Otsu maximum inter-class variance threshold segmentation algorithm is used to perform binarization processing on the downsampled image to distinguish between the skeletal foreground and the non-skeletal background. S103. Extract the complete outer contour boundary of the skeleton based on the binarized image, crop the image based on the skeleton boundary, remove irrelevant background areas other than the skeleton, and obtain a cropped CT image that retains only the skeleton area.

[0022] Preferably, in step S2, the automatic rotation correction process includes the following steps: S201. Resample the cropped CT images; S202. The downsampled image is binarized using Otsu thresholding to obtain a skeleton binary mask; S203. Perform PCA principal component analysis on the binary mask of the skeleton to extract the principal direction feature vector of the overall skeleton contour. S204. Calculate the affine rotation matrix of the image based on the principal direction feature vector of the skeleton, and use the rotation matrix to rotate and transform all CT slices to make the vertebrae uniformly aligned, which facilitates the recognition and quantitative analysis of two-dimensional slice features.

[0023] Preferably, step S3 specifically includes the following steps: S301. Vertebral slice feature classification and identification: The coronal slice along the CT scan axis divides a single complete vertebra into three types of feature slices: Side feature slice, Mid feature slice, and Mid-Down feature slice. Among them, the Side feature slice is a slice of the uppermost and lowermost ends of a single vertebra, in which the vertebral structure is divided into three parts in the image; the Mid feature slice is a slice of the middle section of a single vertebra, in which the vertebra presents a complete closed circular structure; the Mid-Down feature slice is a slice of the middle and lower sections of a single vertebra, in which the main body of the vertebra is divided into two parts and there is a clear cavity in the lower region; the fixed order of the slice arrangement of a single vertebra along the scanning axis is Side→Mid→Mid-Down→Side; S302. Divide the single vertebral segment intervals and select effective slices for trabecular bone analysis: Define the complete axial range of each vertebral segment based on the continuous slice index intervals corresponding to various feature slices; select the Mid feature slice index interval and the Mid-Down feature slice index interval and merge them as a set of effective slices for quantitative analysis of trabecular bone; Side feature slices are not involved in the calculation of trabecular bone parameters. Define the set of contiguous slices as , recorded as Taking a CT image containing three vertebrae—T13 (thoracic), L1 (lumbar), and L2 (lumbar)—with slice numbers ranging from 1 to 892 as an example, the coronal slice corresponding to T13 contains four continuous intervals: the upper side interval... Mid section Mid-Down (lower segment) Lower Side section The lower side section of T13; The two vertebral segments overlap with the adjacent upper L1 side slice, and the boundary between the two vertebrae cannot be distinguished by the side feature alone. S303. Automatically determine vertebral type and generate ordered vertebral sequence: compare the starting index of the first vertebral mid slice. Mid-Down slice start index If satisfied < If the first segment of the image is not found, the first vertebra is determined to be the thoracic vertebra T13; otherwise, the first vertebra is determined to be the lumbar vertebra L2. Based on the axial sequence of the Mid and Mid-Down slices corresponding to each vertebra, the system automatically distinguishes between thoracic and lumbar vertebrae and outputs an ordered vertebral number sequence, which includes T13, L1, L2, L3, L4, L5, and L6. When there is an overlap at the junction of two adjacent vertebrae and the Side slice cannot distinguish their affiliation, the vertebral boundary is determined solely by their independent Mid and Mid-Down slice intervals.

[0024] Preferably, in S4, Mid-Down feature slice interval The two types of interval slices are merged as the input image for single vertebral segmentation. A U-shaped symmetric fully convolutional deep learning segmentation network is used to perform trabecular bone instance segmentation on the slice frame by frame. The segmentation output is a multi-region mask to distinguish trabecular bone, cancellous bone, cortical bone, bone marrow and soft tissue respectively.

[0025] Preferably, in step S5, the quantitative morphological parameters of the skeleton are divided into three categories: bone mass and total volume parameters, trabecular microstructure parameters, and cortical bone structure parameters. The bone mass and total volume parameters include at least the BV / TV bone volume-tissue volume ratio, TV tissue volume, BS bone surface area, and BS / BV specific surface area. The trabecular microarchitecture parameters include at least the number of Tb.N trabeculae, the thickness of Tb.Th trabeculae, the separation of Tb.Sp trabecular bone, the trabecular pattern factor of TBPf, and the connection density of Conn.D. The cortical bone structural parameters include at least Ct.Th cortical bone thickness, Ct.Po cortical bone porosity, and Ct.Ar cortical bone area.

[0026] Preferably, the U-shaped symmetric fully convolutional deep learning segmentation network is a 21-Net network. The 21-Net network includes an encoder downsampling module, a decoder upsampling module, and a skip connection structure. The encoder extracts shallow and deep features of the skeleton, the decoder restores the original size of the image, and the skip connections preserve the fine boundary details of the trabeculae, thereby achieving high-precision trabeculae segmentation.

[0027] Preferably, in step S203, the PCA principal component analysis uses the coordinates of all bone pixels within the binary bone mask as input data, and extracts the principal direction vector with the largest variance through linear transformation dimensionality reduction. The principal direction vector represents the overall extension direction of the vertebra.

[0028] Preferably, CT samples containing thoracic vertebrae T13, lumbar vertebrae L1, and lumbar vertebrae L2, with a slice range of 1–892, are used for network training; wherein the T13 training slice range is [missing information]. L1 training slice interval is + L2 training slice interval is + The U-shaped symmetric fully convolutional deep learning segmentation network is trained using the sliced ​​data within the aforementioned merged interval as the training set.

[0029] Automatic cropping: Step 1: Image downsampling. Step 2: Image binarization using Otsu thresholding. Step 3: Calculation of skeletal boundaries.

[0030] Automatic rotation: The first step is image downsampling. The second step is image binarization using Otsu thresholding. The third step is PCA principal component analysis to obtain the principal orientation of the bones. The fourth step is calculating the rotation matrix. The rotated vertebrae are easier to analyze in a two-dimensional cross-section.

[0031] Vertebral Side / Mid / Mid-Down Feature Recognition: Vertebral slice features are divided into three categories. If a vertebra is scanned from top to bottom, the features are in the following order: Side -> Mid -> Mid-Down -> Side. The features of the uppermost and lowermost images of a vertebra are similar, and the whole is split into three parts, which is called the Side feature.

[0032] The middle end of a vertebra, where the vertebrae are connected as a whole, presents a ring shape, known as the Mid feature.

[0033] The middle and lower segment of a vertebra is broken into two large parts, with the lower part containing a cavity, known as the Mid-Down segment.

[0034] Determine the vertebral sequence and the extent of each segment. By following the slice order of Side->Mid->Middown->Side, the extent of each vertebral segment can be identified. Furthermore, the extent of the trabecular bone segment can be used for trabecular bone analysis, characterized by the Mid+MidDown slice.

[0035] The order of the vertebrae can be determined based on the order of Mid and MidDown (e.g., to arrange them in the order of L1, L2, L3...).

[0036] by Figure 5 For example, the CT images contain three vertebral segments: thoracic T13, lumbar L1, and lumbar L2, with slice numbers 1 to 892. Let a set of consecutive slices be denoted as . , recorded as T13 coronal plane includes , , , .in It also includes some slices of L1, which are indistinguishable in terms of features.

[0037] if < The first vertebra in the current image is T13; otherwise, the first vertebra is L2.

[0038] Vertebral trabecular segmentation: Figure 5 For example, take vertebral CT slices , + and + Some of the data was used for training and segmentation.

[0039] Bone morphological parameters are quantitative indicators that describe the macroscopic / microscopic structure and metabolic activity of bones.

[0040] 1. Bone mass and gross volume parameters sheet 2. Trabecular microstructure parameters (cancellous bone core) sheet 3. Cortical bone structural parameters (core of long bone diaphysis) sheet Otsu thresholding (Otsu's method / maximum inter-class variance method) is the most classic automatic global binarization algorithm. Its core is to maximize the inter-class variance between the foreground and background and automatically calculate the optimal threshold without manual intervention.

[0041] PCA (Principal Component Analysis) is a classic unsupervised dimensionality reduction algorithm. Its core is to map high-dimensional data to a low-dimensional space through linear transformation, while preserving as much of the core information of the data as possible (the direction of the largest variance). It is often used to simplify data and remove redundancy and noise.

[0042] Net is a U-shaped symmetric fully convolutional network that relies on encoder downsampling to extract features, decoder upsampling to restore size, and skip connections to preserve details. It has become a classic basic model for medical image segmentation, and is especially suitable for tasks that require high-precision boundaries, such as cells, tissues, and bones.

[0043] YOLOv8, released by Ultralytics in 2023, is a next-generation end-to-end object detection / segmentation / pose estimation framework and the culmination of the YOLO series. In addition to detection, YOLOv8's core tasks include segmentation, classification, pose estimation, and detection-based tracking.

[0044] YOLOv8's detect module is its core module, specifically designed for object detection, which identifies the category of objects in images / videos and outputs the bounding boxes (locations) of the objects.

[0045] Pose (pose estimation) is a human / animal keypoint detection module based on the YOLOv8 framework. Its core function is to identify the skeletal keypoints of targets in images / videos (such as the nose, shoulders, knees, etc. of a human body). It can also be trained using proprietary data for localization.

[0046] The working principle of this invention is as follows: First, downsampling is performed on the original three-dimensional CT images of the mouse thoracic and lumbar spine. Otsu's maximum inter-class variance thresholding algorithm is used to binarize the image to separate the skeletal foreground from the irrelevant background. The outer contour of the skeleton is extracted for automatic cropping, eliminating redundant background areas, reducing subsequent computational pressure, and removing background noise interference. Then, the cropped image is downsampled and binarized again to obtain a binary mask of the skeleton. Principal component analysis (PCA) is performed using the coordinates of all bone pixels within the mask as input to extract the principal direction feature vector representing the overall extension direction of the vertebrae. Based on this, the affine rotation matrix is ​​solved for all CT slices. A rotational transformation was performed to standardize the vertebral placement postures of different samples, eliminating the influence of tilt and offset on feature recognition. Then, based on the inherent structural variation patterns of mouse vertebral coronal sections, the axial sections of a single vertebra were divided into three feature sections: Side, Mid, and Mid-Down. Following a fixed arrangement order of Side→Mid→Mid-Down→Side, the continuous index intervals corresponding to each type of section were identified, accurately defining the complete axial range of each vertebra. Side sections with cross-vertebral overlap or unstable features were removed, and only Mid and Mid-Down sections were merged for trabecular quantification. Analyzing the effective dataset, and distinguishing between the thoracic T13 and lumbar vertebrae by comparing the starting index positions of the mid and mid-down vertebral slices, the vertebral boundaries are delineated based on non-overlapping mid and mid-down intervals, automatically generating ordered vertebral sequences containing T13 and L1 to L6. The selected effective slices are then input into a self-developed 21-Net U-shaped symmetric fully convolutional deep learning network. The network extracts superficial and deep bone features through encoder downsampling, restores the original image size through decoder upsampling, and preserves subtle trabecular boundary details using skip connections. It performs multi-tissue instance segmentation frame by frame, outputting the final result. This system can distinguish between trabecular bone, cancellous bone, cortical bone, bone marrow, and soft tissue segmentation masks. The network is specifically trained using effective slice datasets of mouse T13, L1, and L2 vertebrae, adapting to mouse vertebral image features to ensure segmentation accuracy. Finally, based on the high-precision tissue segmentation mask, geometric quantification calculations such as region volume, contour, and distance are performed. The system automatically calculates a full set of quantitative morphological parameters and bone density of the skeleton, including total bone volume, trabecular microstructure, and cortical bone structure, in batches. The entire process requires no manual delineation, segmentation, or measurement, enabling one-stop automatic analysis of mouse thoracic and lumbar spine CT images from raw data to bone pathology evaluation indicators.

[0047] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An automatic analysis method for CT images of the thoracic and lumbar vertebrae of mice, characterized in that, Includes the following steps: S1. Automatic cropping of the original three-dimensional CT images of the rat's thoracic and lumbar spine; S2. Perform automatic rotation correction on the CT images cropped in step S1 to standardize the vertebral placement posture; S3. Perform vertebral slice feature recognition, single vertebral segment division, vertebral type determination and ordered sorting on the CT slices after rotation correction; S4. Accurate segmentation of vertebral trabeculae based on the effective slice set; S5. Automatic batch calculation of quantitative morphological parameters and bone density of bones based on trabecular segmentation mask.

2. The automatic analysis method for CT images of the thoracic and lumbar vertebrae of mice according to claim 1, characterized in that, In step S1, the automatic cropping process includes the following steps: S101. Downsample the original CT images; S102. The Otsu maximum inter-class variance threshold segmentation algorithm is used to perform binarization processing on the downsampled image to distinguish between the skeletal foreground and the non-skeletal background. S103. Extract the complete outer contour boundary of the skeleton based on the binarized image, crop the image based on the skeleton boundary, remove irrelevant background areas other than the skeleton, and obtain a cropped CT image that retains only the skeleton area.

3. The automatic analysis method for CT images of the thoracic and lumbar vertebrae of mice according to claim 1, characterized in that, In step S2, the automatic rotation correction process includes the following steps: S201. Resample the cropped CT images; S202. The downsampled image is binarized using Otsu thresholding to obtain a skeleton binary mask; S203. Perform PCA principal component analysis on the binary mask of the skeleton to extract the principal direction feature vector of the overall skeleton contour. S204. Calculate the affine rotation matrix of the image based on the principal direction feature vector of the skeleton, and use the rotation matrix to rotate and transform all CT slices to make the vertebrae uniformly aligned, which facilitates the recognition and quantitative analysis of two-dimensional slice features.

4. The automatic analysis method for CT images of the thoracic and lumbar vertebrae of mice according to claim 1, characterized in that, S3 specifically includes the following steps: S301. Vertebral slice feature classification and identification: The coronal slice along the CT scan axis divides a single complete vertebra into three types of feature slices: Side feature slice, Mid feature slice, and Mid-Down feature slice. Among them, the Side feature slice is a slice of the uppermost and lowermost ends of a single vertebra, in which the vertebral structure is divided into three parts in the image; the Mid feature slice is a slice of the middle section of a single vertebra, in which the vertebra presents a complete closed circular structure; the Mid-Down feature slice is a slice of the middle and lower sections of a single vertebra, in which the main body of the vertebra is divided into two parts and there is a clear cavity in the lower region; the fixed order of the slice arrangement of a single vertebra along the scanning axis is Side→Mid→Mid-Down→Side; S302. Divide the single vertebral segment intervals and select effective slices for trabecular bone analysis: Define the complete axial range of each vertebral segment based on the continuous slice index intervals corresponding to various feature slices; select the Mid feature slice index interval and the Mid-Down feature slice index interval and merge them as a set of effective slices for quantitative analysis of trabecular bone; Side feature slices are not involved in the calculation of trabecular bone parameters. Define the set of contiguous slices as , recorded as Taking a CT image containing three vertebrae—T13 (thoracic), L1 (lumbar), and L2 (lumbar)—with slice numbers ranging from 1 to 892 as an example, the coronal slice corresponding to T13 contains four continuous intervals: the upper side interval... Mid section Mid-Down (lower segment) Lower Side section The lower side section of T13; The two vertebral segments overlap with the adjacent upper L1 side slice, and the boundary between the two vertebrae cannot be distinguished by the side feature alone. S303. Automatically determine vertebral type and generate ordered vertebral sequence: compare the starting index of the first vertebral mid slice. Mid-Down slice start index If satisfied < If the first segment of the image is not found, the first vertebra is determined to be the thoracic vertebra T13; otherwise, the first vertebra is determined to be the lumbar vertebra L2. Based on the axial sequence of the Mid and Mid-Down slices corresponding to each vertebra, the system automatically distinguishes between thoracic and lumbar vertebrae and outputs an ordered sequence of vertebral numbers, including T13, L1, L2, L3, L4, L5, and L6. When there is an overlap between the side slices at the junction of two adjacent vertebrae and the vertebral boundaries cannot be distinguished, the vertebral boundaries are determined solely by the independent Mid and Mid-Down slice intervals.

5. The automatic analysis method for CT images of the thoracic and lumbar vertebrae of mice according to claim 1, characterized in that, In S4, Mid-Down feature slice interval The two types of interval slices are merged as the input image for single vertebral segmentation. A U-shaped symmetric fully convolutional deep learning segmentation network is used to perform trabecular bone instance segmentation on the slice frame by frame. The segmentation output is a multi-region mask to distinguish trabecular bone, cancellous bone, cortical bone, bone marrow and soft tissue respectively.

6. The automatic analysis method for CT images of the thoracic and lumbar vertebrae of mice according to claim 1, characterized in that, In S5, the quantitative morphological parameters of the skeleton are divided into three categories: bone mass and total volume parameters, trabecular microstructure parameters, and cortical bone structure parameters. The bone mass and total volume parameters include at least the BV / TV bone volume-tissue volume ratio, TV tissue volume, BS bone surface area, and BS / BV specific surface area. The trabecular microarchitecture parameters include at least the number of Tb.N trabeculae, the thickness of Tb.Th trabeculae, the separation of Tb.Sp trabecular bone, the trabecular pattern factor of TBPf, and the connection density of Conn.D. The cortical bone structural parameters include at least Ct.Th cortical bone thickness, Ct.Po cortical bone porosity, and Ct.Ar cortical bone area.

7. The automatic analysis method for CT images of the thoracic and lumbar vertebrae of mice according to claim 5, characterized in that, The U-shaped symmetric fully convolutional deep learning segmentation network is a 21-Net network. The 21-Net network includes an encoder downsampling module, a decoder upsampling module, and a skip connection structure. The encoder extracts shallow and deep features of the skeleton, the decoder restores the original size of the image, and the skip connections preserve the fine boundary details of the trabeculae, thus achieving high-precision trabeculae segmentation.

8. The automatic analysis method for CT images of the thoracic and lumbar vertebrae of mice according to claim 3, characterized in that, In step S203, the PCA principal component analysis uses the coordinates of all bone pixels within the binary bone mask as input data, and extracts the principal direction vector with the largest variance through linear transformation dimensionality reduction. The principal direction vector represents the overall extension direction of the vertebra.

9. The automatic analysis method for CT images of the thoracic and lumbar vertebrae of mice according to claim 1, characterized in that, Network training was conducted using CT samples containing thoracic vertebra T13, lumbar vertebra L1, and lumbar vertebra L2 with slice ranges from 1 to 892. The T13 training slice interval is: L1 training slice interval is + L2 training slice interval is + The U-shaped symmetric fully convolutional deep learning segmentation network is trained using the sliced ​​data within the aforementioned merged interval as the training set.

Images