Detection method, device and equipment based on lumbar spine area in MRI images

Abstract

The embodiment of the present invention discloses a detection method, device and equipment based on the lumbar spine area in MRI images. First, a lumbar spine MRI image dataset of a patient to be detected is obtained, and the obtained MRI image dataset is input into the YOLOv5 model for target detection. YOLOv5 automatically extracts ROI related to the lumbar spine and generates a feature image dataset. The feature image dataset output by YOLOv5 is input into the Mix‑inresNet model, and the Mix‑inresNet model performs in-depth classification on the feature image dataset and outputs a classification result dataset. The classification results are soft-voted and combined with the classification results of different models. The final scoring result is output to assist doctors in diagnosing lumbar spine diseases. Through automated image processing and analysis, the time for doctors to manually review and analyze images is reduced, thereby improving work efficiency.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image detection, and in particular to a detection method, device and equipment based on a lumbar spine area in MRI images. Background Art

[0002] Lumbar spine disorders refer to various conditions and problems affecting the lumbar spine, typically involving the bones, discs, nerves, muscles, or ligaments in the lower spine. These conditions may cause lower back pain, numbness in the lower limbs, muscle weakness, or other symptoms. Lumbar spine disorders include: lumbar disc herniation, lumbar spinal stenosis, lumbar degeneration, spondylolisthesis, scoliosis, and lumbar muscle strain.

[0003] When diagnosing lumbar spine diseases, imaging examinations are often performed first. Images are taken through X-rays, magnetic resonance imaging (MRI), CT scans, and bone scans to obtain imaging images of the lumbar spine. Doctors then use medical imaging images to diagnose whether the lumbar spine has a disease.

[0004] MRI can show soft tissue structures in detail, such as the intervertebral disc, spinal cord, nerve roots, and surrounding ligaments. After receiving the MRI (magnetic resonance imaging) results of the lumbar spine, the doctor will conduct a detailed analysis and diagnosis based on the specific manifestations of the image.

[0005] Doctors often evaluate the lumbar disc, nerve roots, etc. in the lumbar spine MRI to determine lumbar spine disease.

[0006] With the development of image recognition technology, some image segmentation techniques are being used to obtain clearer medical images, helping doctors better view and analyze lumbar spine images. However, existing image recognition analysis only uses object detection to extract features from lumbar spine images, resulting in poor accuracy and susceptibility to noise, making it difficult for doctors to diagnose the condition. Summary of the Invention

[0007] Based on this, it is necessary to address the above problems and propose a detection method, device and equipment based on the lumbar spine area in MRI images.

[0008] A method for detecting a lumbar vertebrae region based on MRI images, the method comprising:

[0009] Obtain an MRI image dataset of the lumbar spine of the patient to be tested;

[0010] Constructing a YOLOv5 model, inputting the lumbar MRI image dataset into the YOLOv5 model to obtain a feature image dataset;

[0011] Constructing a Mix-inresNet model, inputting the feature image dataset into the Mix-inresNet model, and the Mix-inresNet model outputting a classification result dataset;

[0012] The Mix-inresNet model performs soft voting on the classification result dataset and outputs the scoring results to assist doctors in lumbar spine examinations.

[0013] In at least one embodiment of the present application, the step of constructing a YOLOv5 model, inputting the lumbar MRI image dataset into the YOLOv5 model, and obtaining a feature image dataset includes:

[0014] The YOLOv5 model extracts the features of the central spinal canal, lateral recess, neural foramen, and intervertebral disc from the lumbar MRI image dataset to obtain the feature image dataset, wherein the feature image dataset includes: each detected ROI bounding box in the lumbar MRI image dataset, the classification result dataset, the ROI image, and the confidence score.

[0015] In at least one embodiment of the present application, the steps of constructing a Mix-inresNet model, inputting the feature image dataset into the Mix-inresNet model, and outputting a classification result dataset from the Mix-inresNet model include:

[0016] The ROI image is input into the Mix-inresNet model, which classifies the ROI image into normal, mild, moderate, and severe categories to obtain a classified result dataset.

[0017] In at least one embodiment of the present application, the method further includes:

[0018] The MRI images of the lumbar spine of the patient to be tested are labeled to obtain a labeled dataset.

[0019] In at least one embodiment of the present application, the steps of constructing a Mix-inresNet model, inputting the feature image dataset into the Mix-inresNet model, and the Mix-inresNet model outputting a classification result dataset further include:

[0020] Construct the training set, validation set, and test set of the Mix-inresNet model, and divide the images and labels of the original dataset into the training set, validation set, and test set in a ratio of 8:1:1, so that the training set accounts for 80% of the total dataset, and the validation set and test set each account for 10% of the total dataset.

[0021] In at least one embodiment of the present application, the Mix-inresNet model includes a backbone network module, which integrates a mixed convolutional layer, an average pooling layer, a dropout layer, and a softmax layer;

[0022] The hybrid convolution layer includes: depthwise separable convolution, dilated convolution and standard convolution;

[0023] The backbone network module performs multi-scale processing and parallel processing on the hybrid convolution layer, average pooling layer, and dimensionality reduction module to obtain the ROI image features after parallel processing.

[0024] In at least one embodiment of the present application, the detection method based on the lumbar region in the MRI image further includes:

[0025] The Mix-inresNet model also includes a self-attention module, which is placed before the average pooling layer;

[0026] The output of the self-attention module is calculated as follows:

[0027] F g ∈R(C′×H×W)F h ∈R(C′×H×W)F g F h G∈R(C′×N)H∈R(C′×N)N=H×Wo∈R(C′×N);

[0028] O=softmax((G T )*H / sqrt(d))*H;

[0029] Get the final feature map.

[0030] In at least one embodiment of the present application, the detection method based on the lumbar region in the MRI image further includes:

[0031] Receive the result correction data of the external input Mix-inresNet model to train the Mix-inresNet model.

[0032] A device for detecting the lumbar region based on MRI images, applied to any of the above methods for detecting the lumbar region based on MRI images, comprising:

[0033] A data acquisition module is used to acquire an MRI image dataset of the lumbar spine of the patient to be tested;

[0034] The YOLOv5 model is used to extract features from the lumbar spine MRI image dataset;

[0035] Mix-inresNet model, which processes feature image datasets and outputs scoring results;

[0036] The detection device based on the lumbar spine area in the MRI image performs the following steps:

[0037] Obtain an MRI image dataset of the lumbar spine of the patient to be tested;

[0038] Constructing a YOLOv5 model, inputting the lumbar MRI image dataset into the YOLOv5 model to obtain a feature image dataset;

[0039] Constructing a Mix-inresNet model, inputting the feature image dataset into the Mix-inresNet model, and the Mix-inresNet model outputting a classification result dataset;

[0040] The Mix-inresNet model performs soft voting on the classification result dataset and outputs the scoring results to assist doctors in lumbar spine examinations.

[0041] A computer device includes a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the following steps:

[0042] Obtain an MRI image dataset of the lumbar spine of the patient to be tested;

[0043] Constructing a YOLOv5 model, inputting the lumbar MRI image dataset into the YOLOv5 model to obtain a feature image dataset;

[0044] Constructing a Mix-inresNet model, inputting the feature image dataset into the Mix-inresNet model, and the Mix-inresNet model outputting a classification result dataset;

[0045] The Mix-inresNet model performs soft voting on the classification result dataset and outputs the scoring results to assist doctors in lumbar spine examinations.

[0046] The implementation of the present invention will have the following beneficial effects:

[0047] The detection method, device and equipment based on the lumbar spine area in MRI images in this embodiment first obtain the lumbar spine MRI image dataset of the patient to be detected, and input the obtained MRI image dataset into the YOLOv5 model for target detection. YOLOv5 automatically extracts the ROI related to the lumbar spine and generates a feature image dataset. The feature image dataset output by YOLOv5 is input into the Mix-inresNet model, and the Mix-inresNet model performs in-depth classification of the feature image dataset and outputs the classification result dataset. The classification results are soft-voted and combined with the classification results of different models. The final scoring results are output to assist doctors in diagnosing lumbar spine diseases. Through automated image processing and analysis, the time for doctors to manually review and analyze images is reduced, thereby improving work efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0049] in:

[0050] Figure 1 is a flow chart of a method for detecting the lumbar spine region based on MRI images in one embodiment;

[0051] Figure 2 is another flow chart of a method for detecting a lumbar spine region based on MRI images in one embodiment;

[0052] Figure 3 is another flow chart of a method for detecting a lumbar spine region based on MRI images in one embodiment;

[0053] Figure 4 is a full flow chart of a method for detecting the lumbar spine region based on MRI images in one embodiment;

[0054] Figure 5 is a structural block diagram of a detection device based on the lumbar spine area in MRI images in one embodiment;

[0055] Figure 6 is a structural block diagram of a computer device in one embodiment;

[0056] Figure 7 It is the MRI characteristic image in the scoring result.

[0057] 200. Detection device based on the lumbar spine area in MRI images; 210. Data acquisition module; 220. YOLOv5 model; 230. Mix-inresNet model. DETAILED DESCRIPTION

[0058] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0059] A method for detecting a lumbar vertebrae region based on MRI images, the method comprising:

[0060] S101, obtaining an MRI image dataset of the lumbar spine of a patient to be tested;

[0061] S102, constructing a YOLOv5 model, inputting the lumbar MRI image dataset into the YOLOv5 model, and obtaining a feature image dataset;

[0062] S103, constructing a Mix-inresNet model, inputting the feature image dataset into the Mix-inresNet model, and the Mix-inresNet model outputting a classification result dataset;

[0063] S104 and Mix-inresNet models perform soft voting on the classification result dataset and output the scoring results to assist doctors in lumbar spine examinations.

[0064] Please refer to Figure 1-4 In this embodiment, first, a lumbar spine MRI image dataset of the patient to be detected is obtained, and the obtained MRI image dataset is input into the YOLOv5 model for target detection.

[0065] YOLOv5 automatically extracts ROIs related to the lumbar spine and generates a feature image dataset.

[0066] Input the feature image dataset output by YOLOv5 into the Mix-inresNet model.

[0067] The Mix-inresNet model performs in-depth classification on the feature image dataset and outputs the classification result dataset.

[0068] Perform soft voting on the classification results and combine the classification results of different models.

[0069] Output the final scoring results to assist doctors in diagnosing lumbar diseases.

[0070] Through automated image processing and analysis, the time doctors spend manually reviewing and analyzing images is reduced, thereby improving work efficiency.

[0071] The combination of YOLOv5 and Mix-inresNet models can improve the recognition rate of lumbar diseases and reduce the probability of misdiagnosis and missed diagnosis.

[0072] The scoring results and classification information provided can help doctors make better decisions, especially in complex cases, and provide valuable reference information.

[0073] It should be noted that, first, an MRI image dataset taken of a patient is obtained, and the MRI image dataset is an MRI image of the lumbar region.

[0074] A Yolov5 model was constructed, and the lumbar MRI image dataset was input into the Yolov5 model to obtain a feature image dataset.

[0075] The Yolov5 model output includes the features of each region of interest (ROI) in the MRI image, specifically: the bounding box of the detected ROI, the classification result for each ROI, the confidence score and the feature image.

[0076] Next, the feature image dataset extracted by YOLOv5 is input into the Mix-inresNet model for further processing.

[0077] The Mix-inresNet model adopts a mixed convolution method (depthwise separable convolution, dilated convolution, and standard convolution) and enhances feature extraction through the self-attention mechanism.

[0078] The Mix-inresNet model classifies each ROI image and outputs the classification results.

[0079] Through deeper learning, the Mix-inresNet model can perform detailed classification of detected ROIs, helping doctors quickly identify the severity of lesions (normal, mild, moderate, severe).

[0080] The classification results of different models are combined and the weighted average method is used to improve the classification accuracy.

[0081] Finally, the scoring results of each detected ROI are output to assist doctors in making clinical judgments.

[0082] Through the soft voting mechanism, the possible deviations of a single model can be reduced, the reliability of the overall diagnosis can be improved, and the decision-making efficiency of doctors in the diagnosis of lumbar diseases can be improved.

[0083] The scoring results include MRI characteristic images, corresponding scores and levels.

[0084] Based on the feature proportions of the lesion characteristics, the classification result data set includes normal ROI images, mild ROI images, moderate ROI images, and severe ROI images to assist doctors in making disease judgments, reminding doctors of image information that needs attention, and assisting doctors in making quick judgments on the lumbar MRI images of patients to be tested.

[0085] In at least one embodiment of the present application, the step of constructing a YOLOv5 model, inputting the lumbar MRI image dataset into the YOLOv5 model, and obtaining a feature image dataset includes:

[0086] S201. The YOLOv5 model extracts the features of the central spinal canal, lateral recess, neural foramen, and intervertebral disc from the lumbar MRI image dataset to obtain the feature image dataset. The feature image dataset includes: each detected ROI bounding box in the lumbar MRI image dataset, the classification result dataset, the ROI image, and the confidence score.

[0087] In this embodiment, the YOLOv5 model processes the input MRI image through a convolutional neural network and automatically detects multiple feature regions in the image, including the central spinal canal, lateral recess, neural foramen, and intervertebral disc.

[0088] The YOLOv5 model identifies the bounding box of each ROI based on the learned features and generates the corresponding classification label and confidence score.

[0089] The bounding box of the detected ROI, classification results, ROI image and confidence score are integrated into a feature image dataset for subsequent analysis.

[0090] The YOLOv5 model was used to automatically detect targets in lumbar spine MRI images, significantly reducing the workload of doctors in manually reviewing images and improving efficiency.

[0091] Using the YOLOv5 target detection algorithm, key lumbar spine structural features can be more accurately identified and located, reducing the possibility of misdiagnosis and missed diagnosis.

[0092] The YOLOv5 model has a faster detection speed and is suitable for real-time or near real-time clinical applications, improving work efficiency.

[0093] The output feature image dataset contains not only the bounding box and classification information of the ROI, but also the corresponding confidence score, helping doctors make more scientific diagnostic decisions.

[0094] The generated ROI images and classification results provide a basis for subsequent in-depth analysis, and the Mix-inresNet model can be further used for refined classification or more complex analysis.

[0095] It should be further clarified that in the MRI-based lumbar spine region detection method, all standard MRI sequences were acquired using fractional readout fast spin echo (frFSE) sequences, with a uniform formatted output resolution of 512x512 pixels. Routine clinical lumbar spine MRI imaging studies were performed using a Signa HDx 1.5T scanner (GE Healthcare).

[0096] In parallel, routine clinical lumbar spine MRI imaging studies were performed on the external test set using an Optima MR360 1.5T scanner (GE Healthcare).

[0097] When training the YOLOv5 model, professionals first annotated lumbar spine MRI images to identify the ROIs. The annotation method was to mark the central canal and lateral recess on axial T2 MRI, the neural foramen on sagittal T1 MRI, and the lumbar intervertebral disc on sagittal T2 MRI. Subsequently, professional clinicians graded the annotated dataset to train the YOLOv5 and Mix-inresNet models.

[0098] In at least one embodiment of the present application, the steps of constructing a Mix-inresNet model, inputting the feature image dataset into the Mix-inresNet model, and outputting a classification result dataset from the Mix-inresNet model include:

[0099] S301. Input the ROI image into the Mix-inresNet model. The Mix-inresNet model classifies the ROI image into normal, mild, moderate, and severe categories to obtain a classified result data set.

[0100] Please refer to Figure 1-4 ,In this implementation, first, a Mix-inresNet model is constructed and ,configured to integrate multiple convolutional layer structures to ,improve the feature extraction and classification capabilities of ROI ,images.

[0101] Input the ROI image in the feature image dataset extracted by the YOLOv5 model into the Mix-inresNet model, ensuring that the image format and size meet the model requirements.

[0102] The Mix-inresNet model performs forward propagation on the input ROI image and extracts deep features through multi-layer convolution and pooling operations.

[0103] Applying activation functions such as ReLU introduces nonlinearity, enabling the model to learn complex feature representations.

[0104] In the output layer, the feature map is converted into category probability through the Softmax function, and the category of each ROI (normal, mild, moderate, severe) is finally determined.

[0105] The Mix-inresNet model outputs a classification result dataset, which contains the classification label of each ROI image for subsequent analysis and clinical reference.

[0106] The Mix-inresNet model uses multiple convolutional structures and deep learning algorithms to more accurately classify ROIs in lumbar spine MRI images, thereby improving classification accuracy.

[0107] The automated classification process significantly increases processing speed and reduces the time doctors spend manually analyzing images, thereby improving work efficiency.

[0108] The model can subdivide the classification results into normal, mild, moderate and severe according to the severity of the disease, providing richer information to help doctors formulate treatment plans.

[0109] By generating a classification result dataset, the Mix-inresNet model provides doctors with a reliable basis and enhances the confidence and accuracy of clinical diagnosis.

[0110] The feature image dataset extracted by YOLOv5 is input into the Mix-inresNet model, which effectively utilizes the previous detection results and improves the efficiency of the overall diagnosis process.

[0111] It should be noted that this detection method based on the lumbar spine area in MRI images uses a soft voting method to integrate the lumbar spine image scoring results of a single patient, and uses NLG technology to convert the results into a detailed reference report to assist doctors in making judgments on the patient's lumbar spine.

[0112] In at least one embodiment of the present application, the method further includes:

[0113] The MRI images of the lumbar spine of the patient to be tested are labeled to obtain a lumbar spine MRI image dataset.

[0114] Please refer to Figure 1-4 In this embodiment, an MRI image of the lumbar spine is collected from the patient to be examined.

[0115] Use medical image annotation tools to annotate the collected MRI images. The specific steps include:

[0116] Doctors identify and label key structures in each MRI image.

[0117] Draw a bounding box for each marked region and assign a corresponding classification label.

[0118] Organize the labeled images into a dataset, ensuring that each image has its corresponding bounding box and classification label for subsequent training, validation, and testing of the model.

[0119] Provide high-quality training data for subsequent deep learning models (such as YOLOv5, Mix-inresNet, etc.) to ensure that the model can effectively learn and identify different categories of lumbar diseases.

[0120] It should be noted that the lumbar spine MRI image dataset includes bounding boxes and classification labels (central spinal canal, lateral recess, neural foramen, and intervertebral disc).

[0121] In at least one embodiment of the present application, the steps of constructing a Mix-inresNet model, inputting the feature image dataset into the Mix-inresNet model, and the Mix-inresNet model outputting a classification result dataset further include:

[0122] Construct the training set, validation set, and test set of the Mix-inresNet model, and divide the images and labels of the original dataset into the training set, validation set, and test set in a ratio of 8:1:1, so that the training set accounts for 80% of the total dataset, and the validation set and test set each account for 10% of the total dataset.

[0123] Please refer to Figure 1-4 ,In this implementation, we start to build a Mix-inresNet model, which integrates multiple convolutional layers and ,attention mechanisms to optimize the image classification ability.

[0124] The feature image dataset extracted by the YOLOv5 model is input into the Mix-inresNet model for feature extraction and classification.

[0125] From the original labeled dataset, the data is divided into training set, validation set and test set in a ratio of 8:1:1.

[0126] Ensure that each dataset contains images and their corresponding labels to facilitate subsequent model training and evaluation.

[0127] Use the training set to train the Mix-inresNet model and adjust the model parameters.

[0128] After each training cycle, the validation set is used to evaluate the model performance and hyperparameters are tuned based on the results.

[0129] Use the test set to perform a final evaluation of the model to obtain the model's performance on unseen data and ensure its reliability and effectiveness.

[0130] The Mix-inresNet model improves the classification accuracy of lumbar spine MRI images and reduces the misdiagnosis rate through learning from high-quality training sets.

[0131] Reasonable division of training, validation, and test sets ensures the evaluation of the model at different stages, promotes the effective adjustment of model parameters, and improves performance.

[0132] By training and evaluating on different datasets, the model can better adapt to new data and improve its reliability in practical applications.

[0133] The output classification results provide doctors with fast and accurate diagnostic information, promoting the clinical examination and treatment of lumbar diseases.

[0134] It should be noted that the Mix-inresNet model classifies the input ROI image, and the output result dataset includes the classification label of each ROI (such as normal, mild, moderate, and severe).

[0135] The images and labels of the original dataset are divided in a ratio of 8:1:1, as follows:

[0136] Training set (80%): used for model training and parameter adjustment.

[0137] Validation set (10%): used to evaluate performance and adjust hyperparameters during model training.

[0138] Test set (10%): used to finally evaluate the generalization ability of the model and test its performance on unseen data.

[0139] In at least one embodiment of the present application, the Mix-inresNet model includes a backbone network module, which integrates a mixed convolutional layer, an average pooling layer, a dropout layer, and a softmax layer;

[0140] The hybrid convolution layer includes: depthwise separable convolution, dilated convolution and standard convolution;

[0141] The backbone network module performs multi-scale processing and parallel processing on the hybrid convolution layer, average pooling layer, and dimensionality reduction module to obtain the ROI image features after parallel processing.

[0142] Please refer to Figure 1-4 ,In this implementation, a Mix-inresNet model is constructed, the ,backbone network module is integrated, and the mixed convolution layer, ,average pooling layer, Dropout layer and Softmax layer are ,configured.

[0143] The extracted ROI image is input into the backbone network module for feature processing.

[0144] Perform convolution operations on the input ROI image and extract various image features through different types of convolution (depthwise separable, extended, standard).

[0145] The convolution feature map is downsampled to reduce the spatial dimension of the feature map and extract more representative features.

[0146] During the training process, a certain proportion of neurons are randomly discarded to enhance the generalization ability of the model.

[0147] The final feature output is converted into category probability distribution to provide a basis for subsequent classification decisions.

[0148] The ROI image features after multi-scale and parallel processing are obtained for subsequent classification and analysis.

[0149] The hybrid convolutional layer combines multiple convolution techniques to effectively extract complex image features and improve classification accuracy.

[0150] Multiple convolution and pooling operations of different scales are performed simultaneously, and the results are connected in parallel. This multi-scale parallel strategy is used to capture and integrate image features of different scales and directions.

[0151] In the feature extraction stage of the backbone network, the Mix-inresNet model uses a hybrid convolution approach, integrating depthwise separable convolution, dilated convolution, and standard convolution in parallel connections. The depthwise separable convolution module consists of two layers: depthwise convolution and pointwise convolution. Pointwise convolution is a simple 1×1 convolution used to form a linear combination of the outputs from the depthwise layer. The depthwise convolution with one filter per input channel can be expressed as:

[0152]

[0153] ; For depthwise convolution kernels of different sizes, filters are applied to the channels in F to produce channels of filtered output feature maps. The main advantage of depthwise separable convolution is that it can significantly reduce the complexity and computational cost of the model.

[0154] A further module of hybrid convolution is dilated convolution, which introduces a dilation factor that enables the convolution kernel to skip input elements when applied, achieving the purpose of detecting larger scale content. Specifically, given a dilation factor, the dilated convolution operation is defined as:

[0155] (F*lk)(p)=∑s+lt=pF(s)k(t);

[0156] F is the input, k is the convolution kernel, and s and t are the sliding variable values ​​​​in the entire operation process.

[0157] Applying dilated convolution can expand the receptive field (also called the receptive field) of the convolutional neural network, enabling the network to perceive a wider range of information to help the model acquire richer environmental information.

[0158] Hybrid convolution combines features from depthwise separable convolution, unrolled convolution, and standard convolution. This allows for deeper, wider-scale, and higher-level abstractions by mixing information at different levels, scopes, and abstractions. This, in turn, facilitates the extraction of richer and more adaptive image representations. The combination of these three convolution types allows for the extraction of richer features while effectively controlling computational resources, thus balancing computational efficiency and performance.

[0159] It should be further explained that depthwise separable convolution is mainly used to extract low-level features in images, such as edges, corners, and simple textures. Using 1x1 convolution to linearly combine the features of all channels captures the relationship between channels, significantly reducing the number of parameters and computational complexity.

[0160] Its structure contains small-scale convolution kernels, and the output is added to the input through residual connections.

[0161] The formula is as follows:

[0162] Branch_1=Conv(1x1)(x)

[0163] Branch_2=Conv(3x3)(Conv(1x1)(x))

[0164] Branch_3=Conv(3x3)(Conv(3x3)(Conv(1x1)(x)))

[0165] InceptionOutput_A=[Branch_1,Branch_2,Branch_3]

[0166] y_A=x+alpha*Conv(1x1)(InceptionOutput_A).

[0167] Dilated convolution is used to capture mid-level features, such as larger structural information and contextual information. These features help the model understand more complex patterns in the image, especially when dealing with blurred edges and object boundaries.

[0168] Dilated convolution introduces a dilation factor (dilation rate) to skip input elements when the convolution kernel is applied, thereby increasing the receptive field of the convolution kernel and enabling it to "see" a wider area of ​​the image.

[0169] Its structure contains medium-sized convolution kernels, and the output is added to the input through residual connections.

[0170] The formula is as follows:

[0171] Branch_1=Conv(1x1)(x)

[0172] Branch_2=Conv(7x1)(Conv(1x7)(Conv(1x1)(x)))

[0173] InceptionOutput_B=[Branch_1,Branch_2]

[0174] y_B=x+alpha*Conv(1x1)(InceptionOutput_B).

[0175] Standard convolution is mainly used to extract high-level features, such as the complex shapes of objects and specific visual patterns, comprehensively identify complex patterns and object shapes, and provide support for classification decisions.

[0176] This enables the Mix-inresNet model to effectively extract features at different levels, thereby improving the detection and classification capabilities of the lumbar spine region in MRI images.

[0177] The Block-C module is used to extract high-level features. Its structure contains larger-scale convolution kernels and adds the output to the input through residual connections.

[0178] The formula is as follows:

[0179] Branch_1=Conv(1x1)(x)

[0180] Branch_2=Conv(3x1)(Conv(1x3)(Conv(1x1)(x)))

[0181] InceptionOutput_C=[Branch_1,Branch_2]

[0182] y_C=x+alpha*Conv(1x1)(InceptionOutput_C).

[0183] Here, x is the input tensor, which is represented as a three-dimensional tensor of R^H x W x C, where H represents the height, W represents the width, and C represents the number of channels.

[0184] Conv(kxk) is a kxk convolution operation, where k represents the size of the convolution kernel. For example, Conv(3x3) represents a 3x3 convolution kernel.

[0185] Alpha is a scaling factor used to adjust the influence of the residual connection part and is usually set to a small value, such as alpha = 0.1.

[0186] [□]: Channel concatenation operation, which connects the outputs of parallel branches in the channel dimension.

[0187] y_A, y_B, y_C are the output tensors of the depthwise separable convolution, unfolded convolution, and standard convolution modules, and their dimensions are the same as the input tensor x.

[0188] It should be further explained that the Mix-inresNet model also includes a dimensionality reduction module, which includes ReductionA and ReductionB modules. By reducing the spatial dimension or number of channels of the feature map, the computational complexity is reduced and more representative features are provided for the subsequent feature extraction layer.

[0189] The ReductionA module is typically used to significantly reduce the spatial dimension of a feature map while retaining some channel information. It combines multiple parallel convolution and pooling operations and finally obtains the output feature map through splicing.

[0190] The formula is as follows:

[0191] Branch_1=MaxPool(3x3,stride=2)(x)

[0192] Branch_2=Conv(3x3,stride=2)(x)

[0193] Branch_3=Conv(1x1)(x)->Conv(3x3)(Branch_3)

[0194] Branch_4=Conv(1x1)(x)->Conv(3x3)(Branch_4)->Conv(3x3,stride=2)(Branch_4)

[0195] ReductionA_Output=[Branch_1, Branch_2, Branch_3, Branch_4]

[0196] Among them, x is the input tensor, ReductionA_Output is the output after the dimensionality reduction and feature extraction operations of different paths, and finally the outputs of these branches are spliced ​​together in the channel dimension.

[0197] The ReductionB module is usually used to further reduce the spatial dimension and number of channels of the feature map. It combines more complex convolution operations and pooling operations and is suitable for the middle and late stages of the network.

[0198] The formula is as follows:

[0199] Branch_1=MaxPool(3x3,stride=2)(x)

[0200] Branch_2=Conv(1x1)(x)->Conv(3x3,stride=2)(Branch_2)

[0201] Branch_3=Conv(1x1)(x)->Conv(7x1)(Branch_3)->Conv(1x7)(Branch_3)->Conv(3x3,stride=2)(Branch_3)

[0202] ReductionB_Output=[Branch_1,Branch_2,Branch_3]

[0203] Among them, x is the input tensor, ReductionB_Output is the output after the dimensionality reduction and feature extraction operations of different paths, and finally the outputs of these branches are spliced ​​together in the channel dimension.

[0204] Here, x is the input tensor, which is represented as a three-dimensional tensor of R^H x W x C, where H represents the height, W represents the width, and C represents the number of channels.

[0205] MaxPool(3x3,stride=2): Maximum pooling operation, using a 3x3 pooling kernel and a sliding window with a stride of 2, usually used to reduce the spatial dimensions (H and W) by half.

[0206] Conv(kxk, stride=s): Convolution operation, using a convolution kernel of size kxk with a stride of s. Stride=2 is often used to reduce spatial dimensions.

[0207] ReductionA_Output and ReductionB_Output: are the outputs of the ReductionA and ReductionB modules respectively. After feature extraction and dimensionality reduction operations through different paths, the outputs of the branches are spliced ​​together to form the output.

[0208] In at least one embodiment of the present application, the detection method based on the lumbar region in the MRI image further includes:

[0209] The Mix-inresNet model also includes a self-attention module, which is placed before the average pooling layer;

[0210] The output of the self-attention module is calculated as follows:

[0211] F g ∈R(C′×H×W)F h ∈R(C′×H×W)F g F h G∈R(C′×N)H∈R(C′×N)N=H×Wo∈R(C′×N)

[0212] ; O = softmax((G T )*H / sqrt(d))*H;

[0213] Get the final feature map.

[0214] C, H, and W represent the channel count, height, and width of the feature map, respectively.

[0215] OF rescales it and restores it to the original feature dimension through linear convolution, and then adds it to the original feature map to obtain the final feature map.

[0216] In this embodiment, in the Mix-inresNet model, the self-attention module is integrated as a part, and the integrated self-attention module is located before the average pooling layer.

[0217] The feature maps that have passed through convolution and other processing layers are input into the self-attention module.

[0218] The self-attention module generates attention weights by calculating the similarity or correlation between each position in the feature map.

[0219] According to the calculated weights, the input features are weighted to highlight important features and suppress background noise.

[0220] The feature map processed by the self-attention module is passed to the average pooling layer.

[0221] The final feature map used for classification contains features optimized by self-attention.

[0222] The self-attention mechanism enables the model to dynamically adjust feature weights based on the input data, thereby adapting to different types of images and task requirements.

[0223] The final output feature map provides richer and more useful features for subsequent classification or detection, promoting the accuracy and efficiency of medical image analysis.

[0224] It should be noted that the constructed self-attention module first generates two new feature mappers through convolution operation, and then we flatten them along the spatial dimension to obtain and respectively.

[0225] It is important to note that the average pooling layer in the Mix-inresNet model is used to reduce the spatial size (width and height) of the feature map to reduce computational complexity and aggregate information. The purpose of average pooling is to smooth the feature map by averaging the local regions, thereby extracting features with smooth transitions while preserving a certain amount of spatial information.

[0226] For an input feature map X, assuming its size is H×W and the size of the pooling window is k h ×k w , the step size is s, X[i×s+m,j×s+n] is the pixel value of the input feature map within the pooling window. Then each element y of the output feature map Y i,j It can be calculated by the following formula:

[0227]

[0228] In at least one embodiment of the present application, the detection method based on the lumbar region in the MRI image further includes:

[0229] Receive the result correction data of the external input Mix-inresNet model to train the Mix-inresNet model.

[0230] Please refer to Figure 1-4 In this embodiment, the result correction data of the Mix-inresNet model is received from an external source (such as a clinician or a data management system).

[0231] The corrected data is organized and integrated into the training dataset to facilitate subsequent model training.

[0232] Retrain the Mix-inresNet model using the training set containing the corrected data and update the model weights and parameters.

[0233] During training, the model will use the correction data to adjust its classification decisions so that its output is more consistent with actual clinical outcomes.

[0234] After training is complete, the model is evaluated against new data to ensure that its performance has been effectively improved.

[0235] If necessary, new correction data can be received and the above process can be repeated to continuously optimize the model.

[0236] By introducing external result correction data, the model can learn the actual classification rules more accurately, improving the detection and classification capabilities of lumbar MRI images, thereby improving classification accuracy and enhancing model robustness.

[0237] It should be noted that the activation function ReLU retains the positive gradient by directly setting the part of the input less than 0 to 0, which partially alleviates the gradient disappearance problem.

[0238] The mathematical expression of the ReLU function is: f(x) = max(0,x)

[0239] in:

[0240] x is the input value.

[0241] When x≥0, output f(x)=x.

[0242] When x<0, output f(x)=0.

[0243] The ReLU function outputs all negative values ​​as 0, and positive values ​​remain unchanged.

[0244] During model training, the loss function is the cross-entropy loss function. In addition, to improve the training process, the model also introduces an auxiliary classifier to provide additional gradient signals, thereby enhancing feature extraction capabilities. The loss function and its formula in the Mix-inresNet model are as follows:

[0245]

[0246] Where N is the number of samples, y i is the actual label, is the probability distribution predicted by the model, is the cross entropy loss of the main classifier, is the cross entropy loss of the auxiliary classifier, use In the cross entropy loss formula, α is the weight coefficient of the auxiliary loss.

[0247] A device 200 for detecting a lumbar region in an MRI image, applied to any of the above-described methods for detecting a lumbar region in an MRI image, comprises:

[0248] The data acquisition module 210 is used to acquire an MRI image dataset of the lumbar spine of the patient to be tested;

[0249] YOLOv5 model 220, used for feature extraction of lumbar spine MRI image dataset;

[0250] Mix-inresNet model 230, which processes the feature image dataset and outputs the scoring results;

[0251] The detection device based on the lumbar spine area in the MRI image performs the following steps:

[0252] Obtain an MRI image dataset of the lumbar spine of the patient to be tested;

[0253] Constructing a YOLOv5 model, inputting the lumbar MRI image dataset into the YOLOv5 model to obtain a feature image dataset;

[0254] Constructing a Mix-inresNet model, inputting the feature image dataset into the Mix-inresNet model, and the Mix-inresNet model outputting a classification result dataset;

[0255] The Mix-inresNet model performs soft voting on the classification result dataset and outputs the scoring results to assist doctors in lumbar spine examinations.

[0256] Please refer to Figure 5 In this embodiment, first, the detection device 200 based on the lumbar spine area in the MRI image obtains the lumbar spine MRI image dataset of the patient to be detected through the data acquisition module 210, and inputs the obtained MRI image dataset into the YOLOv5 model 220 for target detection. The YOLOv5 model 220 automatically extracts the ROI related to the lumbar spine and generates a feature image dataset. The feature image dataset output by the YOLOv5 model is input into the Mix-inresNet model 230, and the Mix-inresNet model 230 performs in-depth classification on the feature image dataset and outputs a classification result dataset. The classification results are soft-voted and combined with the classification results of different models. The final scoring result is output to assist doctors in diagnosing lumbar diseases. Through automated image processing and analysis, the time for doctors to manually review and analyze images is reduced, thereby improving work efficiency.

[0257] Please refer to Figure 6 A computer device includes a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the following steps:

[0258] Obtain an MRI image dataset of the lumbar spine of the patient to be tested;

[0259] Constructing a YOLOv5 model, inputting the lumbar MRI image dataset into the YOLOv5 model to obtain a feature image dataset;

[0260] Constructing a Mix-inresNet model, inputting the feature image dataset into the Mix-inresNet model, and the Mix-inresNet model outputting a classification result dataset;

[0261] The Mix-inresNet model performs soft voting on the classification result dataset and outputs the scoring results to assist doctors in lumbar spine examinations.

[0262] In this embodiment, first, a lumbar spine MRI image dataset of the patient to be tested is obtained, and the obtained MRI image dataset is input into the YOLOv5 model for target detection. YOLOv5 automatically extracts the ROI related to the lumbar spine and generates a feature image dataset. The feature image dataset output by YOLOv5 is input into the Mix-inresNet model, and the Mix-inresNet model performs in-depth classification of the feature image dataset and outputs a classification result dataset. The classification results are soft-voted and combined with the classification results of different models. The final scoring results are output to assist doctors in diagnosing lumbar diseases. Through automated image processing and analysis, the time doctors spend manually reviewing and analyzing images is reduced, thereby improving work efficiency.

[0263] Those skilled in the art will appreciate that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through a computer program. The program can be stored in a non-volatile computer-readable storage medium. When the program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0264] The technical features of the above embodiments can be combined arbitrarily. In order to make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0265] The above-described embodiments merely represent several implementation methods of the present application. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that a person of ordinary skill in the art may make various modifications and improvements without departing from the spirit of the present invention, all of which fall within the scope of protection of the present invention. Therefore, the scope of protection of the present invention shall be determined by the appended claims.

Claims

1. A method for detecting the lumbar spine region based on MRI images, characterized in that: The detection method based on the lumbar spine area in MRI images includes: Obtain an MRI image dataset of the lumbar spine of the patient to be tested; Constructing a YOLOv5 model, inputting the lumbar MRI image dataset into the YOLOv5 model to obtain a feature image dataset; The YOLOv5 model extracts features of the central spinal canal, lateral recess, neural foramen, and intervertebral disc from the lumbar MRI image dataset to obtain the feature image dataset, wherein the feature image dataset includes: each detected ROI bounding box in the lumbar MRI image dataset, a classification result dataset, an ROI image, and a confidence score; When training the YOLOv5 model, lumbar spine MRI images were annotated to identify the ROI images. The annotation method was as follows: the central canal and lateral recess were marked on the axial T2 MRI, the neural foramen on the sagittal T1 MRI, and the lumbar intervertebral disc on the sagittal T2 MRI. The annotated dataset was graded to train the YOLOv5 model and the Mix-inresNet model. Constructing a Mix-inresNet model, inputting the feature image dataset into the Mix-inresNet model, and the Mix-inresNet model outputting a classification result dataset; The Mix-inresNet model includes a backbone network module, which integrates a mixed convolutional layer, an average pooling layer, a dropout layer, and a softmax layer; The hybrid convolution layer includes: depthwise separable convolution, dilated convolution and standard convolution; The backbone network module performs multi-scale processing and parallel processing on the hybrid convolution layer, average pooling layer, and dimensionality reduction module to obtain the ROI image features after parallel processing; The detection method based on the lumbar region in the MRI image further includes: The Mix-inresNet model also includes a self-attention module, which is placed before the average pooling layer; The output of the self-attention module is calculated as follows: F g ∈R(C′×H×W)F h ∈R(C′×H×W)F g F h G∈R(C′×N)H∈R(C×N)N=H×WO∈R(C′×N); <h2 style=";text-align:left;direction:ltr">O = softmax((G<h2 style=";text-align:left;direction:ltr"> T <h2 style=";text-align:left;direction:ltr"> )*H / sqrt(d))*H; Get the final feature map; The Mix-inresNet model performs soft voting on the classification result dataset and outputs the scoring results to assist doctors in lumbar spine examinations.

2. The method for detecting the lumbar vertebrae region based on MRI images according to claim 1, characterized in that: The steps of constructing a Mix-inresNet model, inputting the feature image dataset into the Mix-inresNet model, and having the Mix-inresNet model output a classification result dataset include: The ROI image is input into the Mix-inresNet model, which classifies the ROI image into normal, mild, moderate, and severe categories to obtain a classified result dataset.

3. The method for detecting the lumbar region based on MRI images according to claim 1, characterized in that: The detection method based on the lumbar region in the MRI image further includes: The MRI images of the lumbar spine of the patient to be tested are labeled to obtain a labeled dataset.

4. The method for detecting the lumbar region based on MRI images according to claim 1, characterized in that: The steps of constructing a Mix-inresNet model, inputting the feature image dataset into the Mix-inresNet model, and having the Mix-inresNet model output a classification result dataset further include: Construct the training set, validation set, and test set of the Mix-inresNet model, and divide the images and labels of the original dataset into the training set, validation set, and test set in a ratio of 8:1:1, so that the training set accounts for 80% of the total dataset, and the validation set and test set each account for 10% of the total dataset.

5. The method for detecting the lumbar region based on MRI images according to claim 1, characterized in that: The detection method based on the lumbar region in the MRI image further includes: Receive the result correction data of the external input Mix-inresNet model to train the Mix-inresNet model.

6. A device for detecting the lumbar region based on MRI images, applied to the method for detecting the lumbar region based on MRI images as claimed in any one of claims 1 to 5, characterized in that: The detection device based on the lumbar region in the MRI image comprises: A data acquisition module is used to acquire an MRI image dataset of the lumbar spine of the patient to be tested; The YOLOv5 model is used to extract features from the lumbar spine MRI image dataset; Mix-inresNet model, which processes feature image datasets and outputs scoring results.

7. A computer device, characterized in that: The method comprises a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the method for detecting the lumbar region based on MRI images as claimed in any one of claims 1 to 5.

Images