Spinal degenerative nerve disease focus positioning method based on multi-scale convolutional neural network

Abstract

The invention discloses a spinal degenerative nerve disease focus positioning method based on a multi-scale convolutional neural network, and belongs to the technical field of biomedical image processing. According to the method, through multi-scale image input, subdivision label construction and a multi-model training strategy, in combination with the target detection capability of YOLOv8, refined positioning and evaluation of the spinal degenerative nerve disease are realized, and a quantitative auxiliary tool is provided for clinical diagnosis. The core innovation of the method lies in a multi-dimensional label system and label shielding training, and the detection precision of serious lesions is improved.

Description

Technical Field

[0001] This invention belongs to the field of biomedical image processing technology, specifically involving the training of single and multi-target detection networks and the post-processing of magnetic resonance image localization results. Background Technology

[0002] According to the World Health Organization, low back pain is a leading cause of disability worldwide, affecting 619 million people globally in 2020. Most people experience low back pain at some point in their lives, and as they age, pain and limited mobility are often symptoms of spinal disorders. Degenerative spinal diseases, such as foraminal stenosis (NFN), subarticular stenosis (SS), and spinal canal stenosis (SCS), are major causes of back pain and neurological dysfunction, severely impacting patients' quality of life. Magnetic resonance imaging (MRI) can provide detailed visualization of the lumbar spine, intervertebral discs, and nerves, enabling radiologists to assess the presence and severity of these conditions. Accurate diagnosis and classification of these diseases help guide treatment and potential surgical procedures, thereby alleviating low back pain and improving patients' overall health and quality of life. However, accurately classifying degenerative spinal diseases is a significant challenge due to the heterogeneity of disease presentation and the overlapping characteristics between different types of degenerative changes. Conditions such as intervertebral disc degeneration, spinal stenosis, and osteophyte formation often occur simultaneously or at different degrees of severity, further complicating the classification process. Subtle differences in intervertebral disc morphology, spinal stenosis, and neurological involvement require comprehensive and meticulous analysis for reliable differentiation. Furthermore, the diversity of patient anatomy and imaging findings introduces additional variability, necessitating methods capable of consistently handling these complexities. Currently, MRI-based classification methods for degenerative spinal diseases typically rely on traditional machine learning techniques and deep learning models, such as convolutional neural networks (CNNs) and support vector machines (SVMs), to identify disease-specific features. Some studies have explored ensemble methods and feature engineering to improve model performance. While these methods have yielded significant results in specific contexts, they still possess significant limitations. Most traditional models suffer from limited feature extraction capabilities, particularly in capturing detailed anatomical structures crucial for accurate classification, hindering generalization. Moreover, these models often fail to maintain good performance across diverse imaging conditions and patient populations, limiting their clinical application. Additionally, traditional methods typically require extensive preprocessing and manual segmentation, leading to data instability and reduced efficiency; therefore, more adaptive and robust methods are urgently needed. Summary of the Invention

[0003] This invention proposes a fusion-based lightweight model for precise lesion localization and accurate and timely identification of disease severity, especially severe cases.

[0004] The specific implementation scheme of this technology is as follows: A method for lesion localization of spinal degenerative neurological diseases based on multi-scale convolutional neural networks, wherein the spinal degenerative neurological diseases include: intervertebral foramen stenosis, subarticular stenosis, and spinal canal stenosis, and the method includes the following steps:

[0005] Step 1: Obtain magnetic resonance brain imaging data, which must include at least the original image data and the doctor's expert assessment of the location and severity of the spinal disease lesions; construct target detection data based on the obtained magnetic resonance imaging images and the fixed coordinate information of the lumbar vertebrae;

[0006] Step 2: Image scaling; Bilinear interpolation is used for images at various scaling sizes. For the original image pixel area, there is a Q-square at the top left corner. 12 (x1, y2), bottom left corner Q 11 (x1, y1), Q in the upper right corner 22 (x2, y2), bottom right corner point Q 21 (x2, y1), and their pixel values ​​Q 12 Q 11 Q 22 Q 21 The pixel values ​​of the interpolation point R2(x,y2) at y=y2, the interpolation point R1(x,y1) at y=y1, and the interpolation point P(x,y) at x=x are calculated using the following formula:

[0007]

[0008] Where f(X) represents the pixel value at point X;

[0009] Step 3: Construct the model input format;

[0010] Set a bounding box of a predetermined size on the scaled image in step 2, select the lesion location, and use it as the target location box for model input;

[0011] Step 4: Tag formatting;

[0012] 75 subtype labels were constructed; case images were divided into three categories: by orientation (left and right); by severity (mild, moderate, and severe); and by level (L1 / L2, L2 / L3, L3 / L4, L4 / L5, and L5 / S1). Spinal stenosis is not considered by orientation and includes 15 subtypes, while subarticular stenosis includes 30 subtypes. For example, the predictive label format for severe L2 / L3 spinal stenosis is "spinal canalstenosis l2 l3 severe".

[0013] Step 5: Data batching;

[0014] The data is randomly divided into two batches, labeled fold0 and fold1, and training is performed on different batches of data.

[0015] Step 6: YOLOv8 model training;

[0016] Since the YOLO model itself does not have multi-label multi-classification capabilities, a separate detection base model needs to be designed for each subtype.

[0017] Step 6.1: Training the full detection model for mild, moderate, and severe cases;

[0018] For each disease and each batch of data, a separate base model is trained at each scaling size; for example, the NFN model nfn_fold0_384 at a scaling size of 384.

[0019] Step 6.2: Tag masking training;

[0020] A single image with multiple labels is preprocessed to remove non-severely degraded categories, retaining only the most severely degraded categories. The remaining images and their categories are then used for further training based on the training completed in step 6.1, resulting in an additional trained model.

[0021] Step 7: Reasoning;

[0022] Input the MR images from three directions of each case into the corresponding model for inference, and obtain the lesion localization and probability output for each base model for each pathological section; for each subtype, the probability x i The maximum probability of a category is the probability of each category for a single case. Normalizing the obtained probabilities gives the final classification prediction result.

[0023] Furthermore, in step 2, the image is scaled to three sizes, and in steps 6.1 and 6.2, 18 base models are trained, for a total of 36 base models are trained.

[0024] Furthermore, the loss function used in step 6 when training the model. for:

[0025]

[0026] Among them, w i It is the weight of the i-th sample, LL i It is the cross-entropy loss of the i-th sample.

[0027] Furthermore, in step 7, the categories and probabilities obtained from the inference for each image are statistically analyzed, and the results for a single direction are summarized, with the highest probability for a single category being selected.

[0028]

[0029] in, This represents the probability of the m-th image corresponding to the i-th subtype in the current direction, obtained through inference. N is the total number of images in the current direction, and i represents the corresponding subtype.

[0030] Furthermore, in step 7, softmax normalization is used to obtain a probability distribution, that is, the probability p corresponding to each of the 75 subtypes. j The specific formula is as follows:

[0031]

[0032] Where x j This represents the maximum output probability of the model corresponding to the j-th category.

[0033] This invention utilizes multi-scale image input, detailed label construction, and a multi-model training strategy, combined with the target detection capabilities of YOLOv8, to achieve refined localization and assessment of spinal degenerative neurological diseases, providing a quantitative auxiliary tool for clinical diagnosis. Its core innovation lies in its multi-dimensional label system and label masking training, which improves the detection accuracy of severe lesions. Attached Figure Description

[0034] Figure 1 These are schematic diagrams of 12 basic models used in this method to detect intervertebral foramen stenosis;

[0035] Figure 2 This is the process of building the model using this method;

[0036] Figure 3 This method involves the scaling results of specific multi-label (full) image data, including box addition and size illustrations;

[0037] Figure 4 This method involves scaling specific critically ill (label-masked) image results, including box addition and size illustrations;

[0038] Figure 5 This is a schematic diagram of the post-processing steps of the inference;

[0039] Figure 6 This is a partial display of the reasoning results;

[0040] Figure 7 These are the core evaluation parameters of the full training model under a 384-size scale for intervertebral foramen stenosis.

[0041] Figure 8 The results are the core evaluation parameters for training severe cases of intervertebral foramen stenosis model with masked labels under a 384-size scaling. Detailed Implementation

[0042] The purpose of this invention is to provide a method for locating and assessing lesions in spinal degenerative neurological diseases, which can effectively improve the accuracy of lesion localization and classification in the left / right, L1 / L2, L2 / L3, L3 / L4, L4 / L5, and L5 / S1 levels for NFN, SS, and CSC.

[0043] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. See the specific implementation steps below. Figure 2 ;

[0044] Step A: Constructing target detection data

[0045] ① The dataset used in this study is a publicly available dataset provided by the Radiological Society of North America, containing 1975 patients. This data also includes diagnostic results from medical experts, including the level, grade, and diagnostic criteria (lesion coordinates). Therefore, the specific location information of the lesion must first be marked on the original image based on its coordinates. Then, a bounding box of a certain size is used to select the area containing the lesion. The bounding box size varies depending on the scaling, as follows: scaling size 384*384, target bounding box size 28*28; scaling size 512*512, target bounding box size 36*36; scaling size 768*768, target bounding box size 56*56. See... Figure 3 , Figure 4 Specifically, the labeled data should include the category label and bounding box coordinates for each target.

[0046] ② Divide the dataset into a training set and a validation set (80% training set and 20% validation set).

[0047] Step B: Loss Function Design

[0048] For spinal neurodegenerative diseases, the ultimate goal is to develop a lesion localization and detection model that can not only locate lesions in mild and moderate cases, enabling early intervention, but also serve as an expert in detecting severe cases, allowing these patients to receive timely precision medicine and reduce suffering. Therefore, we improve upon the traditional logarithmic loss in multi-class classification problems by using a weighted average logarithmic loss to measure model performance. The specific weights are: mild: 1; moderate: 2; severe: 4. This means that the model will be penalized more severely for mispredicting moderate and severe lesions, thus encouraging better performance on more clinically important cases. The specific formula is as follows:

[0049]

[0050] Among them, w i It is the weight of the i-th sample (1, 2, or 4), LL i This is the cross-entropy loss of the i-th sample. The smaller the average weighted log loss, the better the model performance, especially in identifying critically ill patients.

[0051] Step C: Model Training and Inference

[0052] ① Environment Setup. Install Python (version 3.8 or higher recommended); install PyTorch (version 2.0 or higher recommended); install the Ultralytics YOLOv8 library. You will need an Nvidia graphics card with at least 24GB of dedicated video memory and at least 100GB of storage space.

[0053] ② Modify the configuration file for data preprocessing and YOLO format generation. Edit the augment.py file to replace the original augment.py file in the Ultralytics library, and run the script file to convert the raw data into a format suitable for model training input.

[0054] ③ Train the model. Download the model weight file yolov8.pt, determine the data batch, image scaling size, and detection box size for training the base model, and train 36 base models respectively.

[0055] ④ Model performance evaluation. Evaluate the model performance on the validation set, recording the frame loss, class loss, mAP50, and mAP50-95. For example... Figure 7 , Figure 8 .

[0056] ⑤ Run the inference code to complete the inference process. The inference result is as follows: Figure 6 .

[0057] Step C: Post-processing

[0058] Because a patient has multiple MRI images from multiple angles, and each image can be classified using this method, but the final classification result can only provide one quantitative analysis for each category, the final patient assessment report needs to integrate all the model's inference results. Specific implementation steps are as follows: Figure 5 The category and probability of each image are statistically analyzed. Results for a single direction (Sagittal T1, Axial T2, Sagittal T2STIR) are summarized, and the image with the highest probability for that single category is selected.

[0059]

[0060] Where N is the total number of pathological sections in the current direction, and i represents the corresponding subtype.

[0061] The probability distribution is normalized using softmax, i.e., the probability p corresponding to each of the 75 subtypes. j The specific formula is as follows:

[0062]

[0063] Where x j This represents the maximum output probability of the model corresponding to the j-th category.

[0064] This method utilizes a multi-scale YOLOv8 deep learning model to identify and classify spinal degenerative diseases. The model leverages multi-scale feature extraction techniques to identify disease features at different resolutions and combines this with the efficient target detection capabilities of the YOLOv8 architecture, thereby improving detection accuracy. Furthermore, by training on a large labeled dataset, and given the lightweight nature of the YOLOv8 model itself, this method, by fusing 36 base models, still achieves strong generalization ability while maintaining high detection speed. The ensemble model achieves significant performance metrics on the validation set, with an average accuracy of 92.5% and a mean weighted cross-entropy loss of only 0.4821. In clinical applications, the model demonstrates strong robustness, consistently identifying degenerative lesions under different patient conditions and scanning conditions. It exhibits good adaptability to lesion size, morphology, and location, indicating its broad applicability to various imaging data and significant clinical application value.

Claims

1. A method for lesion localization in spinal degenerative neurological diseases based on multi-scale convolutional neural networks, wherein the spinal degenerative neurological diseases include: The steps of this method include: Foraminal stenosis, subarticular stenosis, and spinal canal stenosis. Step 1: Obtain magnetic resonance brain imaging data, which must include at least the original image data and the doctor's expert assessment of the location and severity of the spinal disease lesions; construct target detection data based on the obtained magnetic resonance imaging images and the fixed coordinate information of the lumbar vertebrae; Step 2: Image scaling; Bilinear interpolation is used for images at various scaling sizes. For the original image pixel area, there is a Q-square at the top left corner. 12 (x1, y2), bottom left corner Q 11 (x1, y1), Q in the upper right corner 22 (x2, y2), bottom right corner point Q 21 (x2, y1), and their pixel values ​​Q 12 Q 11 Q 22 Q 21 The pixel values ​​of the interpolation point R2(x,y2) at y=y2, the interpolation point R1(x,y1) at y=y1, and the interpolation point P(x,y) at x=x are calculated using the following formula: Where f(X) represents the pixel value at point X; Step 3: Construct the model input format; Set a bounding box of a predetermined size on the scaled image in step 2, select the lesion location, and use it as the target location box for model input; Step 4: Tag formatting; 75 subtype labels were constructed; case images were divided into three categories: by orientation (left and right); by severity (mild, moderate, and severe); and by level (L1 / L2, L2 / L3, L3 / L4, L4 / L5, and L5 / S1). Spinal stenosis does not involve orientation and includes 15 subtypes, while subarticular stenosis includes 30 subtypes. Step 5: Data batching; The data is randomly divided into two batches, labeled fold0 and fold1, and training is performed on different batches of data. Step 6: YOLOv8 model training; Since the YOLO model itself does not have multi-label multi-classification capabilities, a separate detection base model needs to be designed for each subtype. Step 6.1: Training the full detection model for mild, moderate, and severe cases; For each disease, a separate base model is trained for each batch of data at each scaling size. Step 6.2: Tag masking training; A single image with multiple labels is preprocessed to remove non-severely degraded categories, retaining only the most severely degraded categories. The remaining images and their categories are then used for further training based on the training completed in step 6.1, resulting in an additional trained model. Step 7: Reasoning; Input the MR images from three directions of each case into the corresponding model for inference, and obtain the lesion localization and probability output for each base model for each pathological section; for each subtype, the probability x i The maximum probability of a category is the probability of each category for a single case. Normalizing the obtained probabilities gives the final classification prediction result.

2. The method for lesion localization in spinal degenerative neurological diseases based on multi-scale convolutional neural networks as described in claim 1, characterized in that, In step 2, the image is scaled to three sizes. In steps 6.1 and 6.2, 18 base models are trained, for a total of 36 base models.

3. The method for lesion localization of spinal degenerative neurological diseases based on multi-scale convolutional neural networks as described in claim 1, characterized in that, Step 6: Loss function during model training for: Among them, w i It is the weight of the i-th sample, LL i It is the cross-entropy loss of the i-th sample.

4. The method for lesion localization in spinal degenerative neurological diseases based on multi-scale convolutional neural networks as described in claim 1, characterized in that, In step 7, the categories and probabilities obtained from the inference for each image are statistically analyzed, and the results for a single direction are summarized, with the highest probability for a single category being selected. in, This represents the probability of the m-th image corresponding to the i-th subtype in the current direction, obtained through inference. N is the total number of images in the current direction, and i represents the corresponding subtype.

5. The method for lesion localization in spinal degenerative neurological diseases based on multi-scale convolutional neural networks as described in claim 1, characterized in that, In step 7, softmax normalization is used to obtain a probability distribution, i.e., the probability p corresponding to each of the 75 subtypes. j The specific formula is as follows: Where x j This represents the maximum output probability of the model corresponding to the j-th category.

Images