Spinal cord injury spinal column MRI image segmentation model and establishment method thereof
By applying the ResNet101 backbone network and region proposal network to spinal MRI images, combined with feature pyramids, high-precision segmentation of spinal cord high signal, spinal cord, spinal canal, and vertebral bodies was achieved, solving the problem of inconsistent diagnostic results in existing technologies and providing efficient and reliable diagnostic support.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- THE NAVAL MEDICAL UNIV OF PLA
- Filing Date
- 2026-04-28
- Publication Date
- 2026-05-29
AI Technical Summary
Existing spinal MRI image segmentation techniques lack high-level and high-standard detection and identification capabilities in the diagnosis of spinal cord injuries. In particular, they fail to achieve multi-dimensional feature mining in spinal instance segmentation, resulting in inconsistent diagnostic results and reliance on physician experience.
Using the ResNet101 standard neural network as the backbone network and combining it with the region proposal network, object region proposals are generated from T2WI sagittal slice images of scanned spinal MRI. The reconstructed feature pyramid and the region proposal network are then used for image classification and bounding box generation, enabling the detection and segmentation of spinal cord high signal, spinal cord, spinal canal, and vertebral bodies.
It improves the accuracy and efficiency of spinal MRI image segmentation, achieves a unified and highly accurate diagnostic level, provides clinicians with reliable quantitative assessment of injury severity and support for treatment decisions, and enhances the stability and consistency of diagnosis.
Smart Images
Figure CN122115482A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of surgical identification technology, specifically relating to a spinal cord injury spinal MRI image segmentation model and its establishment method. Background Technology
[0002] With societal development and changing lifestyles, the incidence of spinal cord injury is rising annually, affecting hundreds of thousands of people worldwide. It often results in varying degrees of sensory and motor dysfunction or even complete loss of function. This severely impacts patients' mental health and quality of life, placing a significant economic burden on families and the social healthcare system. High-level diagnostic methods are a prerequisite for scientifically developing treatment plans and personalized rehabilitation programs.
[0003] Spinal MRI is currently the gold standard imaging method for assessing spinal cord injury (SCI), directly reflecting the state of high-signal areas in the spinal cord, the spinal canal, and the vertebral bodies. While MRI hardware has continuously improved, resulting in higher image quality, the development of methods for measuring image features lags behind advanced imaging techniques. Consequently, diagnostic results are often limited by the evaluator's experience. Due to variations in the skill and experience of different radiologists, the level of comprehensive and detailed diagnosis varies greatly, and high-level, high-standard diagnoses are lacking.
[0004] In recent years, deep learning technology has made some progress in the imaging analysis of spinal cord injuries, especially Convolutional Neural Networks (CNNs), which have shown great potential in spinal MRI image analysis, enabling automatic detection, identification, and classification of key features in images. In 2023, Dr. Mert Karabacak and Dr. Konstantinos Margetis and their team developed a predictive system to improve the care and management of spinal cord injury patients through a personalized approach. Meanwhile, in 2020, Shaolong Ma and his team achieved the detection and identification of spinal diseases using Faster R-CNN. Currently, these methods are mainly applied to the detection and identification of spinal lesions, and have not yet achieved breakthroughs in spinal instance segmentation, nor do they possess the ability to simultaneously mine and analyze features from multiple dimensions. Therefore, this invention is derived from this. Summary of the Invention
[0005] To address at least one of the aforementioned technical problems, the objective of this invention is to provide a spinal cord injury MRI image segmentation model and its establishment method. This model enables the detection, identification, and segmentation of high-signal areas of the spinal cord, the spinal cord, the spinal canal, and each vertebra (C2-T1) in T2-weighted spinal MRI images. This aims to meet the clinical need for detailed and comprehensive assessment of relevant areas in spinal cord injury MRI images. By automating MRI image segmentation, the model improves diagnostic accuracy and efficiency, achieving a unified high-precision level. This provides clinicians with efficient, reliable, and highly accurate technical support for quantifying the severity of injury and making diagnostic and treatment decisions.
[0006] The technical solution of this invention is:
[0007] One objective of this invention is to provide a spinal cord injury MRI image segmentation model, comprising:
[0008] The backbone network uses the ResNet101 standard neural network as a feature extractor to generate suggestions of possible object regions from T2WI sagittal slice images of scanned spinal MRI, and to obtain regions that may contain high signal of spinal cord, spinal cord, spinal canal and vertebral bodies.
[0009] A region proposal network is used to propose region suggestions obtained from the backbone network to achieve image classification and generate bounding boxes and masks.
[0010] Preferably, the backbone network first detects low-level features, including edges and corners of the sagittal image, and then sequentially detects higher-level features, including each vertebral body, spinal canal, spinal cord, and high signal intensity of the spinal cord, so as to transform the original spinal MRI image into a feature map with corresponding features, which serves as the input for the next stage.
[0011] Preferably, the backbone network introduces a reconstructed feature pyramid, which allows functions at each level to access all low-level and high-level functions, enabling the second pyramid to acquire high-level features from the first pyramid and pass them down to lower layers.
[0012] Preferably, the regional proposal network scans the image in a sliding window manner, capturing regions containing high signal intensity of each vertebra, spinal canal, spinal cord, and spinal cord, refining the size and location of each object;
[0013] The finally obtained regions of each vertebral body, spinal canal, spinal cord, and high signal of spinal cord are then passed to the next stage region of interest classifier and bounding box regressor to generate a corresponding category for each object. The corresponding category includes the specific vertebral body sequence, spinal cord, spinal canal, and high signal of spinal cord. The bounding box is then refined to achieve the final size.
[0014] Finally, a corresponding mask is generated for this area.
[0015] Another objective of this invention is to provide a method for establishing a spinal cord injury MRI image segmentation model, comprising:
[0016] S1. Design a segmentation model, which includes a backbone network and a region proposal network. The backbone network uses a ResNet101 standard neural network as a feature extractor to generate suggestions that may contain object regions from T2WI sagittal slice images of scanned spinal MRI, and to obtain regions that may contain high signal intensity of the spinal cord, spinal cord, spinal canal, and each vertebral body. The region proposal network is used to propose regions obtained by the backbone network to achieve image classification and generate bounding boxes and masks.
[0017] S2. The training set, validation set, and test set are labeled in a 7:2:1 ratio to train, select parameters, and test the segmentation model to obtain the desired spinal MRI image segmentation model.
[0018] Preferably, in step S1, the backbone network first detects low-level features, including the edges and corners of the sagittal image, and then sequentially detects higher-level features, including each vertebral body, spinal canal, spinal cord, and high signal intensity of the spinal cord, so as to transform the original spinal MRI image into a feature map with corresponding features, which serves as the input for the next stage.
[0019] Preferably, the backbone network introduces a reconstructed feature pyramid, which allows functions at each level to access all low-level and high-level functions, enabling the second pyramid to acquire high-level features from the first pyramid and pass them down to lower layers.
[0020] Preferably, the regional proposal network scans the image in a sliding window manner, capturing regions containing high signal intensity of each vertebra, spinal canal, spinal cord, and spinal cord, refining the size and location of each object;
[0021] The finally obtained regions of each vertebral body, spinal canal, spinal cord, and high signal of spinal cord are then passed to the next stage region of interest classifier and bounding box regressor to generate a corresponding category for each object. The corresponding category includes the specific vertebral body sequence, spinal cord, spinal canal, and high signal of spinal cord. The bounding box is then refined to achieve the final size.
[0022] Finally, a corresponding mask is generated for this area.
[0023] Preferably, in step S2, the training set, validation set, and test set are obtained using the following criteria and after desensitization processing. The criteria include inclusion criteria and exclusion criteria. The inclusion criteria include a clear diagnosis of spinal cord injury, age ≥18 years, complete preoperative spinal MRI imaging data, and high-quality images. The exclusion criteria include age <18 years, poor quality spinal MRI images, missing or artifact-prone T2 sagittal image sequences, history of spinal surgery, severe scoliosis, ankylosing spondylitis, spinal tuberculosis, or infection.
[0024] Compared with the prior art, the advantages of the present invention are:
[0025] The spinal cord injury (SCI) spinal MRI image segmentation model of this invention exhibits stable performance, superior robustness, and accurate detection, identification, segmentation, and quantification capabilities on SCI spinal MRI. It effectively improves the efficiency, reliability, and accuracy of SCI diagnosis and severity quantification, achieving consistency and standardization in diagnosis and treatment, and helping radiologists report injury characteristics using consistent and objective methods. This enables clinicians to obtain more accurate and detailed diagnostic information, further providing data support for treatment and rehabilitation strategies. Attached Figure Description
[0026] The present invention will be further described below with reference to the accompanying drawings and embodiments:
[0027] Figure 1 This is an example diagram of image annotation in an embodiment of the present invention;
[0028] Figure 2 This is a schematic diagram of the principle framework of the segmentation model according to an embodiment of the present invention;
[0029] Figure 3 This is a simplified diagram of the backbone network of the segmentation model in an embodiment of the present invention;
[0030] Figure 4 This is a reconstruction diagram of the feature pyramid network of the segmentation model in an embodiment of the present invention;
[0031] Figure 5 This is a result image of a single-layer spinal cord high-signal mask from the segmentation model of this invention embodiment;
[0032] Figure 6 The image shows the overall (binary classification) segmentation result of the segmentation model in this embodiment of the invention (where the horizontal axis represents the horizontal pixel coordinates of the image, usually in pixels, ranging from 0 to the image width; the vertical axis represents the vertical pixel coordinates of the image, usually in pixels, ranging from 0 to the image height, with the origin at the upper left or lower left corner).
[0033] Figure 7This is a visualization of the multi-class segmentation effect of the segmentation model in an embodiment of the present invention (where the horizontal axis represents the horizontal pixel coordinates of the image; the vertical axis represents the vertical pixel coordinates of the image). Detailed Implementation
[0034] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments. It should be understood that these descriptions are merely exemplary and not intended to limit the scope of the invention. Furthermore, descriptions of well-known structures and technologies are omitted in the following description to avoid unnecessarily obscuring the concept of the invention.
[0035] An embodiment of the present invention provides a spinal cord injury spinal MRI image segmentation model, comprising:
[0036] The backbone network uses the ResNet101 standard neural network as a feature extractor to generate suggestions of possible object regions from T2WI sagittal slice images of scanned spinal MRI, and to obtain regions that may contain high signal of spinal cord, spinal cord, spinal canal and vertebral bodies.
[0037] A region proposal network is used to propose region suggestions obtained from the backbone network to achieve image classification and generate bounding boxes and masks.
[0038] Another embodiment of the present invention provides a method for establishing a spinal MRI image segmentation model for spinal cord injury, comprising:
[0039] S1. Design a segmentation model, which includes a backbone network and a region proposal network. The backbone network uses a ResNet101 standard neural network as a feature extractor to generate suggestions that may contain object regions from T2WI sagittal slice images of scanned spinal MRI, and to obtain regions that may contain high signal intensity of the spinal cord, spinal cord, spinal canal, and each vertebral body. The region proposal network is used to propose regions obtained by the backbone network to achieve image classification and generate bounding boxes and masks.
[0040] S2. The training set, validation set, and test set are labeled in a 7:2:1 ratio to train, select parameters, and test the segmentation model to obtain the desired spinal MRI image segmentation model.
[0041] Specifically, the technical implementation route of this invention is as follows:
[0042] The spine is divided into the cervical vertebrae (C1-C7), thoracic vertebrae (T1-T12), lumbar vertebrae (L1-L5), sacral vertebrae (5 vertebrae fused into the sacrum), and coccygeal vertebrae (3-4 vertebrae fused into the coccyx). The vertebral foramina in the center of each vertebra connect to form the vertebral canal, which houses the spinal cord. The upper end of the spinal cord connects to the medulla oblongata, and the lower end terminates at approximately the level of the first and second lumbar vertebrae (L1-L2) in adults. Below this point, the spinal cord gradually transforms into a slender filum terminale, surrounded by the cauda equina (spinal nerve roots). Cervical or thoracic vertebral fractures often cause cervical or thoracic spinal cord injury. Taking traumatic cervical spinal cord injury as an example: the physiological structure of the cervical spine changes drastically after trauma, resulting in abnormalities in the cervical spine's anatomy. At this time, the spinal cord not only undergoes anatomical changes (spinal cord compression) but also exhibits macroscopic manifestations of microscopic changes after injury (spinal cord hyperintensity), making the cervical spine structure even more complex. It mainly has the following characteristics: unclear boundaries of high signal areas in the spinal cord, uneven signal intensity, changes in vertebral body morphology and sequence, and alterations in the state of the spinal cord and spinal canal. This makes it difficult to detect, identify, and segment individual vertebrae (C2-T1, specifically as follows). Figure 1 The high signal intensity in C2, C3, C4, C5, C6, T1, the spinal canal, and the spinal cord presents difficulties and challenges. Therefore, we need a more effective method to detect, identify, and segment high signal intensity in each vertebra, spinal canal, and spinal cord in post-injury spinal MRI images with ultra-precise accuracy.
[0043] A retrospective study was conducted on 600 patients with a confirmed diagnosis of spinal cord injury at our affiliated hospital who met the inclusion and exclusion criteria. Spinal MRI sequences of the corresponding patients were collected from the Picture Archiving and Communication System (PACS). A total of 7,200 T2 sagittal images were obtained according to the following inclusion and exclusion criteria.
[0044] The inclusion and exclusion criteria are as follows:
[0045] (1) Inclusion criteria: patients with a confirmed diagnosis of SCI; age ≥18 years; complete preoperative spinal MRI imaging data with high image quality (clear imaging and no artifacts);
[0046] (2) Exclusion criteria: Age <18 years; poor quality spinal MRI, missing T2 sagittal sequence, artifacts; history of spinal surgery, severe scoliosis; coexisting ankylosing spondylitis; spinal tuberculosis, infection. It should be noted that both inclusion and exclusion criteria require all of the above-listed items to be met.
[0047] To externally validate the segmentation model and further evaluate its generalization ability, spinal MRIs of 100 patients with traumatic spinal cord injury from different devices obtained from other hospitals were used, employing inclusion and exclusion criteria consistent with the internal dataset.
[0048] I. Dataset Labeling: Labeling the dataset
[0049] like Figure 1 As shown, after anonymization processing (e.g., removing basic patient information such as name, age, and gender to protect patient privacy), the training set, validation set, and test set (in this embodiment, the training set, validation set, and test set are in a 7:2:1 ratio) were used to accurately annotate the high-signal areas of the spinal cord, spinal cord, spinal canal, and C2-T1 vertebral bodies on sagittal T2WI using the open-source ITK-NSP, with the joint opinion of a spinal cord injury clinician A and two senior imaging diagnostic experts (B and C). Dataset annotation was independently completed by expert B, using standards consistent with the dataset to ensure consistency and standardization in the evaluation. To assess internal consistency during the annotation process, the expert reviewed the test set four weeks after the initial annotation, thereby reducing observer-internal variability and ensuring the accuracy and reliability of the annotation.
[0050] II. Deep Learning Model Development (Ultra-precise mask)
[0051] 2.1 As Figure 2 As shown, this embodiment of the invention designs a two-stage network architecture (Ultra-preciseMask) to detect, identify, segment, and quantify high signal intensity in the spinal cord, spinal cord, spinal canal, and vertebral bodies (C2-T1): The first stage generates proposals (regions that may contain objects) by scanning T2WI sagittal slice images of spinal MRI, obtaining regions that may contain high signal intensity in the spinal cord, spinal cord, spinal canal, and vertebral bodies (C2-T1); The second stage proposes regions obtained in the first stage to achieve image classification and generate bounding boxes and masks.
[0052] 2.2 From a high-level perspective, Ultra-precise Mask consists of the following modules:
[0053] 2.2.1 Backbone Network: The ResNet101 standard neural network is used as the feature extractor. In... Figure 2 The C1 layer is shown to detect low-level features (edges and angles in the sagittal image), followed by C2-C5 layers to detect higher-level features (vertebral bodies, spinal canal, spinal cord, and high signal intensity in the spinal cord). In this way, the original spinal MRI image is transformed into a feature map with the corresponding features, such as... Figures 1 to 4 As shown, this feature map becomes the input for the following stages. Although the aforementioned backbone network performs well, there is still room for improvement, as illustrated in the embodiments of the present invention. Figure 4The diagram illustrates how a reconstructed Feature Pyramid Network (FPN) is introduced to better represent multi-scale object extensions. The FPN adds an improved standard pyramid, effectively creating a second pyramid that extracts high-level features from the first pyramid and passes them to lower layers. This aims to allow features at each level to access all low- and high-level features, eliminating aliasing introduced by upsampling, reducing potential prediction errors, and further enhancing the model's prediction accuracy. Figure 3 As shown.
[0054] 2.2.2 Region Proposal Network (RPN): This is a lightweight neural network that scans the image using a sliding window to identify regions containing high signal intensity from each vertebra, spinal canal, spinal cord, and spinal cord. This refines the size and location of each object. The resulting regions of each vertebra (including C2-T1 mentioned above), spinal canal, spinal cord, and spinal cord are then passed to the next stage, the Region of Interest (ROI) classifier and bounding box regressor. This generates a corresponding category for each object (e.g., specific vertebral sequence, spinal cord, spinal canal, high signal intensity from spinal cord), refines the bounding box to achieve the optimal size, and finally generates a corresponding mask for the region.
[0055] like Figure 5 The image shown is a result of processing a single-layer spinal cord high-signal mask using the segmentation model of an embodiment of the present invention. Figure 6 The image shown is an overall (binary classification) segmentation result diagram of the segmentation model in an embodiment of the present invention. Figure 7 The image shown is a visualization of the multi-class segmentation effect of the segmentation model according to an embodiment of the present invention. Figure 6 and Figure 7It is evident that the segmentation model (Ultra-precise Mask model) of this invention exhibits superior performance in recognizing both global (binary classification) and detailed (multi-class classification) features of the spine. Verification by the applicant confirms that the segmentation model of this invention achieves a mean average precision (mAP) of 0.876. In detecting high signal intensity in the spinal cord, spinal cord, spinal canal, and individual vertebrae, the segmentation model of this invention demonstrates the best performance at an IoU (Intersection over Union) of 0.5, with average precision (AP), recall, and f1 score all reaching 0.999. The applicant compared the segmentation model of this invention (referred to as "our model" in Table 1) with existing segmentation models Mask-RCNN and U-Net. Table 1 shows that, in terms of recognition ability, the high AP, mAP, recall, and f1 score comprehensively demonstrate that the segmentation model of this invention significantly outperforms existing Mask-RCNN and U-Net in terms of category recognition ability, global performance, result completeness, and quality. In terms of detection capability, the AP, recall, and f1 score of the model are all improved compared with Mask-RCNN and U-Net under the same threshold, which proves that the segmentation model of the present invention is significantly better than other existing models in terms of detection completeness and quality.
[0056] Table 1. Comparison of the recognition and detection capabilities of the segmentation model (denoted as our model in the table) of the present invention with existing models Mask-RCNN and U-Net.
[0057]
[0058] III. Clinical contributions of the segmentation model in this invention:
[0059] The Ultra-precise mask model of this invention not only optimizes the automated analysis capabilities of spinal MRI images, but also provides a precise segmentation tool for clinical decision-making, enabling clinicians to accurately assess injuries and develop personalized treatment and rehabilitation plans.
[0060] IV. Conclusion
[0061] The Ultra-precise mask segmentation model of this invention demonstrates stable performance, superior robustness, and accurate detection, identification, and segmentation capabilities on spinal MRI of spinal lesions with spinal cerebral infarction (SCI). It effectively improves the speed and reliability of SCI diagnosis, achieves consistency and standardization in diagnostic levels, and helps radiologists report lesion characteristics using consistent and objective methods. This enables clinicians to obtain more accurate and detailed diagnostic information, further providing data support for treatment and rehabilitation strategies.
[0062] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.
Claims
1. A spinal cord injury MRI image segmentation model, characterized in that, include: The backbone network uses the ResNet101 standard neural network as a feature extractor to generate suggestions of possible object regions from T2WI sagittal slice images of scanned spinal MRI, and to obtain regions that may contain high signal of spinal cord, spinal cord, spinal canal and vertebral bodies. A region proposal network is used to propose region suggestions obtained from the backbone network to achieve image classification and generate bounding boxes and masks.
2. The segmentation model according to claim 1, characterized in that, The backbone network first detects low-level features, including edges and corners of the sagittal image, and then sequentially detects higher-level features, including each vertebral body, spinal canal, spinal cord, and high signal intensity of the spinal cord, in order to transform the original spinal MRI image into a feature map with corresponding features, which serves as the input for the next stage.
3. The segmentation model according to claim 2, characterized in that, The backbone network introduces a reconstructed feature pyramid, which allows each level of functionality to access all low-level and high-level functionality, enabling the second pyramid to acquire high-level features from the first pyramid and pass them down to lower layers.
4. The segmentation model according to claim 1, characterized in that, The region proposal network scans the image in a sliding window manner, capturing regions containing high signals from each vertebra, spinal canal, spinal cord, and spinal cord, refining the size and location of each object; The finally obtained regions of each vertebral body, spinal canal, spinal cord, and high signal of spinal cord are then passed to the next stage region of interest classifier and bounding box regressor to generate a corresponding category for each object. The corresponding category includes the specific vertebral body sequence, spinal cord, spinal canal, and high signal of spinal cord. The bounding box is then refined to achieve the final size. Finally, a corresponding mask is generated for this area.
5. A method for establishing a spinal cord injury spinal MRI image segmentation model, characterized in that, include: S1. Design a segmentation model, which includes a backbone network and a region proposal network. The backbone network uses a ResNet101 standard neural network as a feature extractor to generate suggestions that may contain object regions from T2WI sagittal slice images of scanned spinal MRI, and to obtain regions that may contain high signal intensity of the spinal cord, spinal cord, spinal canal, and each vertebral body. The region proposal network is used to propose regions obtained by the backbone network to achieve image classification and generate bounding boxes and masks. S2. The training set, validation set, and test set are labeled in a 7:2:1 ratio to train, select parameters, and test the segmentation model to obtain the desired spinal MRI image segmentation model.
6. The method according to claim 5, characterized in that, In step S1, the backbone network first detects low-level features, including the edges and corners of the sagittal image, and then detects higher-level features in sequence, including each vertebral body, spinal canal, spinal cord, and high signal intensity of the spinal cord, so as to transform the original spinal MRI image into a feature map with corresponding features, which serves as the input for the next stage.
7. The method according to claim 6, characterized in that, The backbone network introduces a reconstructed feature pyramid, which allows each level of functionality to access all low-level and high-level functionality, enabling the second pyramid to acquire high-level features from the first pyramid and pass them down to lower layers.
8. The method according to claim 5, characterized in that, The region proposal network scans the image in a sliding window manner, capturing regions containing high signals from each vertebra, spinal canal, spinal cord, and spinal cord, refining the size and location of each object; The finally obtained regions of each vertebral body, spinal canal, spinal cord, and high signal of spinal cord are then passed to the next stage region of interest classifier and bounding box regressor to generate a corresponding category for each object. The corresponding category includes the specific vertebral body sequence, spinal cord, spinal canal, and high signal of spinal cord. The bounding box is then refined to achieve the final size. Finally, a corresponding mask is generated for this area.
9. The method according to claim 5, characterized in that, In step S2, the training set, validation set, and test set are obtained using the following criteria and after desensitization processing. The criteria include inclusion criteria and exclusion criteria. The inclusion criteria include patients with a clear diagnosis of spinal cord injury, age ≥18 years, complete preoperative spinal MRI images with high-quality images and no artifacts. The exclusion criteria include age <18 years, poor spinal MRI image quality, missing T2 sagittal image sequences, artifacts, history of spinal surgery, severe scoliosis, ankylosing spondylitis, spinal tuberculosis, or infection.