Lumbar intervertebral disc protrusion MRI image classification method

Abstract

The invention relates to the technical field of medical image analysis and processing, in particular to a lumbar disc herniation MRI image classification method, which comprises the following steps of: collecting MRI image data of the spine of a patient with lumbar disc herniation, and constructing a sagittal plane data set; the method comprises the following steps: identifying a T2WI sequence for lumbar disc herniation diagnosis aiming at an MRI image of a spine of a patient with lumbar disc herniation to be classified, and screening a diagnosis layer in the T2WI sequence; on the basis of a trained optimized YOLOv3 model, performing automatic segmentation and positioning of the intervertebral disc on the diagnosis layer, and determining a geometric center point of the intervertebral disc; and obtaining an intervertebral disc degeneration grading result based on the trained optimized YOLOv4 model, and classifying the lumbar disc herniation MRI image based on the intervertebral disc degeneration grading result. According to the method, the deep learning algorithm and the image processing technology are combined, so that the speed and the accuracy of image analysis processing are improved, and the degenerative change of the intervertebral disc is accurately graded.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image analysis and processing, and in particular to a lumbar disc herniation MRI image classification method. Background Art

[0002] Lumbar disc herniation is a very common degenerative spinal disease in clinical practice, which is usually caused by excessive strain or trauma of the spine and damage to the cushioning and connective tissues between the vertebrae. This spinal disease not only seriously affects the quality of life of patients, but also brings a heavy burden to the national finances. At present, the main methods for diagnosing spinal diseases are magnetic resonance imaging (MRI) and computed tomography (CT). However, CT can only obtain cross-sections perpendicular to the long axis of the human body, so diagnosis is very dependent on MRI images.

[0003] How to locate the required observation level in magnetic resonance imaging and electronic computer tomography images and determine whether there is lumbar disc herniation and the location of the disease is a major challenge to imaging examination methods and the technical level, physical strength and energy of clinical physicians. Due to the unbalanced distribution of medical resources in my country today, the diagnostic accuracy of lumbar disc herniation is acceptable in areas with developed medical resources, but it is prone to misdiagnosis and missed diagnosis in underdeveloped areas.

[0004] It can be seen that whether it is possible to provide a lumbar disc herniation MRI image classification method that can accurately identify the Pfirrmann disc degeneration grading results and realize lumbar disc herniation MRI image classification based on the grading results has become a technical problem that needs to be urgently solved by technical personnel in this field. Summary of the invention

[0005] The embodiment of the present invention provides a lumbar disc herniation MRI image classification method, which solves the defects of the prior art such as low analysis accuracy of lumbar disc herniation MRI images, complex analysis and calculation process, and inability to accurately and quickly obtain Pfirrmann disc degeneration grading results.

[0006] The present invention provides a lumbar disc herniation MRI image classification method, characterized in that it comprises the following steps:

[0007] Step 1: Collect MRI imaging data of the spine of patients with lumbar disc herniation and construct a sagittal dataset containing T1WI, T2WI, and FatOnly sequences;

[0008] Step 2: identifying a T2WI sequence for diagnosing lumbar disc herniation based on the MRI image of the spine of the patient with lumbar disc herniation to be classified, and selecting a diagnostic layer in the T2WI sequence;

[0009] Step 3: Optimize the YOLOv3 model, use the sagittal images of the T1WI, T2WI, and FatOnly sequences in step 1 as the model training data set, train the optimized YOLOv3 model, and automatically segment and locate the intervertebral disc on the diagnostic layer based on the trained optimized YOLOv3 model to determine the geometric center point of the intervertebral disc;

[0010] Step 4: Optimize the YOLOv4 model, use the sagittal images of the T1WI, T2WI, and FatOnly sequences described in step 1 to annotate the lumbar disc degeneration level according to the Pfirrmann grading standard, use the annotated images as the model training data set, train the optimized YOLOv4 model, and predict the diagnostic layer for determining the geometric center point of the intervertebral disc based on the trained optimized YOLOv4 model to obtain the intervertebral disc degeneration grading result, and classify the lumbar disc herniation MRI image based on the intervertebral disc degeneration grading result.

[0011] Furthermore, in step 2, the method for screening the diagnostic layer in the T2WI sequence includes the following steps:

[0012] A cross-platform computer vision library is used to extract the trunk as the region of interest using the abdomen and back boundaries as the outline. In the subsequent image processing steps, pixels outside the region of interest are ignored and only pixels in the region of interest are processed.

[0013] The left side boundary of the spinal cord was extracted using an edge detection algorithm;

[0014] Whether the left border of the spinal cord is detected is determined based on the number of final curves. If two curves are detected, the current image belongs to the T2WI sequence. The length of the left border of the spinal cord of the images numbered between 40% and 70% of the total number of images in the current image set is counted, and the image with the longest border is selected as the diagnosis layer. If one curve is detected, it means that the left border of the spinal cord is not extracted, and the current image sequence is not used as a basis for diagnosis, and the error information is fed back to the input end.

[0015] Furthermore, in step 3, the optimization method of the YOLOv3 model includes the following steps:

[0016] The CBAM attention module is fused after the 4 residual blocks of the Darknet-53 feature extraction module of the YOLOv3 network;

[0017] Detection of diagnostic layers from multiple prediction box scales;

[0018] Use FCM clustering algorithm to optimize anchor box size;

[0019] Use GIoU-loss as the bounding box loss function to calculate the prediction error and get the optimized YOLOv3 model.

[0020] More preferably, the multiple prediction box scales include prediction scales and sizes of 13x13, 26x26, 52x52, and 104x104 prediction boxes of 82 layers, 94 layers, 106 layers, and 118 layers, respectively.

[0021] Furthermore, in step 3, the method for determining the geometric center point of the intervertebral disc includes the following steps:

[0022] Based on the trained optimized YOLOv3 model, the coordinates of the center point of the intervertebral disc were preliminarily predicted;

[0023] The diagnostic layer is input, and the coordinates of the center point of the intervertebral disc are identified using the trained optimized YOLOv3 model, and the coordinates of the center point of each prediction box are recorded in a set, which is the original set;

[0024] The coordinates of the center points of the prediction boxes in the set are traversed to determine whether each point satisfies the geometric relationship constraints. If the geometric relationship constraints are not satisfied, the current coordinate point is a false positive point and needs to be deleted from the set. Otherwise, switch to the next coordinate point and repeat the above steps. After traversing all the coordinate points, the current set is a precise coordinate point set. The number corresponding to each intervertebral disc is marked according to the prior knowledge of the spinal model to complete the intervertebral disc positioning.

[0025] More preferably, the method for determining whether each point satisfies the geometric relationship constraint comprises the following steps:

[0026] Calculate the angle value formed by connecting the center points of adjacent intervertebral discs in the diagnosis layer in the vertical direction, wherein the angle value is 140° to 180°;

[0027] Calculate the distance value between adjacent intervertebral disc center points in the diagnosis layer, where the distance value is 0.6 to 1.2 pixels;

[0028] The positions of the center points of the intervertebral discs in the diagnostic layer are calculated, and the center points of multiple intervertebral discs are arranged in sequence, which conforms to the natural arrangement of the spine.

[0029] Furthermore, in step 4, the method for optimizing the YOLOv4 model includes the following steps:

[0030] The backbone network in YOLOv4 uses MobileNetv3 to replace CSPDarknet-53;

[0031] The number of anchor boxes is set to 3, and the cluster center sizes of the three types of anchor boxes are set to (72, 55), (77, 38), and (398, 508), respectively;

[0032] The prediction error is calculated based on Focal-loss as the bounding box loss function to obtain the optimized YOLOv4 model.

[0033] One or more technical solutions provided by the present invention have at least the following technical effects or advantages:

[0034] 1. The present invention combines deep learning algorithms and image processing technology, especially the improved YOLO series model, to quickly and accurately identify and locate lumbar intervertebral discs from MRI images, improve the speed and accuracy of MRI image analysis and processing of lumbar disc herniation, improve the accuracy of classification through geometric constraint verification and Pfirrmann hierarchical classification, introduce data enhancement and attention mechanisms, and enhance the model's adaptability to different imaging conditions and individual differences. The method of the present invention can be easily integrated into the existing medical imaging system, provide technical support for grassroots hospitals, and contribute to the balanced allocation of medical resources.

[0035] 2. The YOLOv3 and YOLOv4 algorithms trained and optimized by the method of the present invention enhance the recognition ability of small targets and improve the accuracy of image analysis and detection. By integrating the Pfirrmann grading standard, the degenerative changes of the intervertebral disc are accurately graded, thereby accurately realizing the classification of lumbar disc herniation MRI images.

[0036] 3. When the optimized YOLOv3 model automatically segments and locates the intervertebral disc on the diagnostic layer, the precision, recall rate, and mAP value can reach more than 80%, which are higher than YOLOv3 and other networks to varying degrees, and the detection frame rate per second can reach 32.2FPS, which meets the needs of daily diagnosis.

[0037] 4. The introduction of data enhancement and attention mechanism in the method of the present invention further improves the robustness of the model and its adaptability to complex medical images, making the MRI image analysis results of lumbar disc herniation and the Pfirrmann grading results more reliable. Overall, this technology provides an innovative, fast and accurate auxiliary means for the diagnosis of lumbar disc herniation. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 This is a flow chart of step 2 of a method for classifying MRI images of lumbar disc herniation according to an embodiment of the present invention;

[0039] Figure 2 This is a flow chart of determining the geometric center point of the intervertebral disc in step 3 of a method for classifying MRI images of lumbar disc herniation in an embodiment of the present invention. DETAILED DESCRIPTION

[0040] In order to better understand the above technical solution, the above technical solution will be described in detail below in conjunction with the accompanying drawings and specific implementation methods.

[0041] Example:

[0042] The embodiment of the present invention provides a method for classifying lumbar disc herniation MRI images, comprising the following steps:

[0043] Step 1: Collect MRI imaging data of the spine of patients with lumbar disc herniation and construct a sagittal dataset containing T1WI, T2WI, and FatOnly sequences;

[0044] Step 2: Using computer vision technology to identify the T2WI sequence used for diagnosing lumbar disc herniation from the MRI image of the spine of the patient with lumbar disc herniation to be classified, and selecting the best diagnostic layer in the T2WI sequence. The process is as follows: Figure 1 As shown, the specific steps include:

[0045] A cross-platform computer vision library is used to extract the trunk as the region of interest using the abdomen and back boundaries as the outline. In the subsequent image processing steps, pixels outside the region of interest are ignored and only pixels in the region of interest are processed.

[0046] The Canny operator edge detection algorithm was used to extract the left side boundary of the spinal cord;

[0047] Determine whether the left border of the spinal cord is detected based on the number of final curves. If two curves are detected, the current image belongs to the T2WI sequence. Count the length of the left border of the spinal cord of the images numbered between 40% and 70% of the total number of images in the current image set. For example, if the total number of images in this sequence is 12, count the length of the left border of the spinal cord of the images numbered 5-8 (rounded to 5 after the decimal point) in the current image set. Select the image with the longest border as the best diagnostic layer; if one curve is detected, it means that the left border of the spinal cord is not extracted, and the current image sequence is T1WI or FatOnly. These two sequences are not used as a basis for diagnosis, and the error information needs to be fed back to the input end;

[0048] Step 3: Fusing the CBAM attention module after the four residual blocks of the Darknet-53 feature extraction module of the YOLOv3 network; detecting the diagnostic layer from multiple prediction box scales, including 13x13, 26x26, 52x52, and 104x104 prediction boxes with prediction scales and sizes of 82 layers, 94 layers, 106 layers, and 118 layers respectively; multi-scale prediction enables the network to more accurately detect intervertebral discs of different sizes; using the FCM clustering algorithm instead of the K-means algorithm to optimize the anchor box size, thereby further improving the accuracy of detection; using GIoU-loss as the loss function to obtain the optimized YOLOv3 model, GIoU-loss provides a more comprehensive overlap area evaluation than IoU, which helps to optimize the prediction results; using the sagittal plane dataset described in step 1 to train the optimized YOLOv3 model, and automatically segmenting and locating the intervertebral disc on the diagnostic layer based on the trained optimized YOLOv3 model to determine the geometric center point of the intervertebral disc;

[0049] The method for determining the geometric center point of the intervertebral disc includes the following steps. The method flow chart is as follows: Figure 2 As shown:

[0050] Based on the trained optimized YOLOv3 model, the coordinates of the center point of the intervertebral disc are preliminarily predicted; the diagnostic layer is input, the coordinates of the center point of the intervertebral disc are identified using the trained optimized YOLOv3 model, and the coordinates of the center point of each prediction box are recorded in a set, which is the original set and may contain some false positive points, which need to be screened; the coordinates of the center point of the prediction box in the set are traversed to determine whether each point satisfies the geometric relationship constraint; if the geometric relationship constraint is not satisfied, the current coordinate point is a false positive point and needs to be deleted from the set, otherwise switch to the next coordinate point and repeat the above steps. After traversing all coordinate points, the current set is a precise coordinate point set, and the number corresponding to each intervertebral disc is marked according to the prior knowledge of the spine model to complete the intervertebral disc positioning;

[0051] The method for determining whether each point satisfies the geometric relationship constraint includes the following steps:

[0052] Angle constraint: Calculate the angle value formed by connecting the center points of adjacent intervertebral discs in the diagnosis layer in the vertical direction, and the angle value is 140° to 180°;

[0053] Distance constraint: Calculate the distance between adjacent intervertebral disc center points in the diagnostic layer, and the distance value is 0.6 to 1.2 times the standard distance of the intervertebral disc per pixel;

[0054] Continuity check: Calculate the position of the disc center point in the diagnostic layer. Multiple disc center points are arranged in sequence, which is consistent with the natural arrangement of the spine.

[0055] Step 4: The backbone network in YOLOv4 uses MobileNetv3 to replace CSPDarknet-53 to reduce the amount of calculation and improve the detection speed; the number of anchor boxes is set to 3, and the cluster center sizes of the three types of anchor boxes are set to (72, 55), (77, 38), and (398, 508) respectively to adapt to the changes in the size of the intervertebral disc and improve the diagnostic accuracy; Focal-loss is used as the loss function to solve the problem of class imbalance and improve the detection accuracy of small targets, thus obtaining the optimized YOLOv4 model;

[0056] The sagittal plane dataset described in step 1 is used to train and optimize the YOLOv4 model, and the disc degeneration grading result is obtained based on the trained optimized YOLOv4 model, and the lumbar disc herniation MRI images are classified based on the disc degeneration grading result. Specifically, the lumbar disc herniation MRI images with the same disc degeneration grading result are classified into the same category.

[0057] In summary, the MRI image classification method for lumbar disc herniation proposed in the present invention shows significant advantages in automation, efficiency, accuracy, robustness, etc., and is expected to be widely used in clinical practice. Compared with the traditional manual diagnosis method that relies on the doctor's experience, the disc degeneration grading results obtained by the method of the present invention. The image analysis can be completed automatically, reducing the workload of doctors. The use of deep learning algorithms, especially the improved YOLO series models, can quickly and accurately identify and locate the intervertebral disc, improving the efficiency of diagnosis. Through geometric constraint verification and Pfirrmann graded diagnosis, the accuracy of disc degeneration grading is improved, and the possibility of misdiagnosis and missed diagnosis is reduced. The introduction of data enhancement and attention mechanisms enhances the adaptability of the model to different imaging conditions and individual differences, and improves the robustness of disc degeneration grading results. The method can be easily integrated into the existing medical imaging system, providing technical support for grassroots hospitals and contributing to the balanced allocation of medical resources. Automated diagnosis reduces the dependence on highly skilled doctors, which helps to reduce medical costs in the long run.

[0058] Although the preferred embodiments of the present invention have been described, those skilled in the art may make other changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention.

[0059] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.

Claims

1. A lumbar disc herniation MRI image classification method, characterized in that: The following steps are involved: Step 1: Collect MRI imaging data of the spine of patients with lumbar disc herniation and construct a sagittal dataset containing T1WI, T2WI, and FatOnly sequences; Step 2: identifying a T2WI sequence for diagnosing lumbar disc herniation based on the MRI image of the spine of the patient with lumbar disc herniation to be classified, and selecting a diagnostic layer in the T2WI sequence; Step 3: Optimize the YOLOv3 model, use the sagittal images of the T1WI, T2WI, and FatOnly sequences in step 1 as the model training data set, train the optimized YOLOv3 model, and automatically segment and locate the intervertebral disc on the diagnostic layer based on the trained optimized YOLOv3 model to determine the geometric center point of the intervertebral disc; Step 4: Optimize the YOLOv4 model, use the sagittal images of the T1WI, T2WI, and FatOnly sequences described in step 1 to annotate the lumbar disc degeneration level according to the Pfirrmann grading standard, use the annotated images as the model training data set, train the optimized YOLOv4 model, and predict the diagnostic layer for determining the geometric center point of the intervertebral disc based on the trained optimized YOLOv4 model to obtain the intervertebral disc degeneration grading result, and classify the lumbar disc herniation MRI image based on the intervertebral disc degeneration grading result.

2. The MRI image classification method for lumbar disc herniation according to claim 1, characterized in that: In step 2, the method for screening the diagnostic layer in the T2WI sequence includes the following steps: A cross-platform computer vision library is used to extract the trunk as the region of interest using the abdomen and back boundaries as the outline. In the subsequent image processing steps, pixels outside the region of interest are ignored and only pixels in the region of interest are processed. The left side boundary of the spinal cord was extracted using an edge detection algorithm; Whether the left border of the spinal cord is detected is determined based on the number of final curves. If two curves are detected, the current image belongs to the T2WI sequence. The length of the left border of the spinal cord of the images numbered between 40% and 70% of the total number of images in the current image set is counted, and the image with the longest border is selected as the diagnostic layer. If one curve is detected, it means that the left border of the spinal cord is not extracted, and the current image sequence is not used as a basis for diagnosis, and the error information is fed back to the input end.

3. The MRI image classification method for lumbar disc herniation according to claim 1, characterized in that: In step 3, the optimization method of the YOLOv3 model includes the following steps: The CBAM attention module is fused after the 4 residual blocks of the Darknet-53 feature extraction module of the YOLOv3 network; Detection of diagnostic layers from multiple prediction box scales; Use FCM clustering algorithm to optimize anchor box size; Use GIoU-loss as the bounding box loss function to calculate the prediction error and get the optimized YOLOv3 model.

4. The MRI image classification method for lumbar disc herniation according to claim 3, characterized in that: Multiple prediction box scales include 13x13, 26x26, 52x52, 104x104 prediction boxes with prediction scales and sizes of 82 layers, 94 layers, 106 layers, and 118 layers respectively.

5. The MRI image classification method for lumbar disc herniation according to claim 1, characterized in that: In step 3, the method for determining the geometric center point of the intervertebral disc comprises the following steps: Based on the trained optimized YOLOv3 model, the coordinates of the center point of the intervertebral disc were preliminarily predicted; The diagnostic layer is input, and the coordinates of the center point of the intervertebral disc are identified using the trained optimized YOLOv3 model, and the coordinates of the center point of each prediction box are recorded in a set, which is the original set; The coordinates of the center points of the prediction boxes in the set are traversed to determine whether each point satisfies the geometric relationship constraints. If the geometric relationship constraints are not satisfied, the current coordinate point is a false positive point and needs to be deleted from the set. Otherwise, switch to the next coordinate point and repeat the above steps. After traversing all the coordinate points, the current set is a precise coordinate point set. The number corresponding to each intervertebral disc is marked according to the prior knowledge of the spinal model to complete the intervertebral disc positioning.

6. The MRI image classification method for lumbar disc herniation according to claim 5, characterized in that: The method for determining whether each point satisfies the geometric relationship constraint comprises the following steps: Calculate the angle value formed by connecting the center points of adjacent intervertebral discs in the diagnosis layer in the vertical direction, wherein the angle value is 140° to 180°; Calculate the distance value between adjacent intervertebral disc center points in the diagnosis layer, where the distance value is 0.6 to 1.2 pixels; The positions of the center points of the intervertebral discs in the diagnostic layer are calculated, and the center points of multiple intervertebral discs are arranged in sequence, which conforms to the natural arrangement of the spine.

7. The MRI image classification method for lumbar disc herniation according to claim 1, characterized in that: In step 4, the method for optimizing the YOLOv4 model includes the following steps: The backbone network in YOLOv4 uses MobileNetv3 to replace CSPDarknet-53; The number of anchor boxes is set to 3, and the cluster center sizes of the three types of anchor boxes are set to (72, 55), (77, 38), and (398, 508), respectively; The prediction error is calculated based on Focal-loss as the bounding box loss function to obtain the optimized YOLOv4 model.

Images