Detection device and method for lumbar degenerative diseases
By improving the computer-aided diagnosis network for lumbar degenerative diseases, combined with multimodal time series data and feature extractors, the problems of dataset imbalance and high computational complexity are solved, and efficient and accurate detection of lumbar degenerative diseases is achieved.
Patent Information
- Application Number
- CN202511178565.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-08-21
- Publication Date
- 2025-10-24
AI Technical Summary
Existing technologies in lumbar magnetic resonance imaging have problems such as data set imbalance, single-modality single-task architecture that is difficult to meet complex clinical diagnostic needs, and high dependence on data labeling. In addition, the three-dimensional convolutional neural network has high computational complexity, while the two-dimensional network ignores time series features.
A computer-aided diagnosis network for lumbar degenerative diseases is adopted, including sub-networks for detecting spinal stenosis, neural foraminal stenosis, and subarticular stenosis. Multimodal time series data is utilized, combined with feature extractors for localization tasks and classification tasks, and an improved EfficientNet-B0 module, channel-spatial fusion attention module, and gated recurrent unit are used to achieve joint multimodal and multi-task processing.
It improves the accuracy and efficiency of lumbar degenerative disease detection, can effectively detect multiple diseases, reduce computational complexity, and provide efficient and robust auxiliary support.
Smart Images
Figure CN120833904A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of image processing, in particular to a lumbar degenerative disease detection device and method. BACKGROUND
[0002] Related research shows that although deep learning has made significant progress in lumbar magnetic resonance imaging (MRI) analysis, there are still three core problems to be solved: (1) Data: In the face of the problem of data set imbalance, most research methods do not respond to algorithm design or network architecture; (2) Task: The previous single-modal, single-task architecture is difficult to meet the complex clinical diagnosis requirements; (3) Network level: The network architecture based on instance segmentation or target detection relies heavily on data annotation, which exacerbates the data bottleneck problem. At the same time, although the three-dimensional convolutional neural network (3D CNN) can integrate multi-frame information, the calculation complexity is high, and the 2D network usually directly ignores the time sequence features of MRI. SUMMARY
[0003] The purpose of the present application is to provide a lumbar degenerative disease detection device and method, which can assist in detecting lumbar degenerative diseases and improve the detection efficiency of lumbar degenerative diseases.
[0004] To achieve the above purpose, the present application provides the following solutions:
[0005] In a first aspect, the present application provides a lumbar degenerative disease detection device, comprising:
[0006] An acquisition module is configured to acquire a lumbar magnetic resonance image to be processed;
[0007] A preprocessing module is connected to the acquisition module and is configured to preprocess the lumbar magnetic resonance image to be processed to obtain different image sequences;
[0008] A detection module is connected to the preprocessing module and is configured to input the different image sequences into a trained lumbar degenerative disease computer-aided diagnosis network for detection to determine the type, lesion degree and lesion position of the lumbar degenerative disease; wherein the type of the lumbar degenerative disease includes spinal stenosis, foraminal stenosis and subarticular stenosis, and the lesion position is represented by an image index and an intervertebral disc coordinate; the lumbar degenerative disease computer-aided diagnosis network includes a spinal stenosis detection subnetwork, a foraminal stenosis detection subnetwork and a subarticular stenosis detection subnetwork; each subnetwork includes a plurality of feature extractors.
[0009] In a second aspect, the present application provides a lumbar degenerative disease detection method, comprising:
[0010] Obtaining a lumbar magnetic resonance image to be processed;
[0011] Pretreating the lumbar magnetic resonance image to be processed to obtain different image sequences;
[0012] Inputting the different image sequences into a trained lumbar degenerative disease computer-aided diagnosis network for detection to determine the type, lesion degree and lesion position of the lumbar degenerative disease; wherein the type of the lumbar degenerative disease includes spinal canal stenosis, neural foramen stenosis and subarticular stenosis, and the lesion position is represented by an image index and an intervertebral disc coordinate; the lumbar degenerative disease computer-aided diagnosis network comprises a spinal canal stenosis detection subnetwork, a neural foramen stenosis detection subnetwork and a subarticular stenosis detection subnetwork; each subnetwork comprises a plurality of feature extractors.
[0013] According to the specific embodiments provided by the present application, the following technical effects are disclosed:
[0014] The present application provides a lumbar degenerative disease detection device and method. The detection module in the device uses a lumbar degenerative disease computer-aided diagnosis network to detect different image sequences for lumbar degenerative disease, obtaining the type, lesion degree and lesion position of the lumbar degenerative disease. The different image sequences include multi-modal time series data, indicating that the device can process multi-modal time series data. In addition, the lumbar degenerative disease computer-aided diagnosis network comprises a spinal canal stenosis detection subnetwork, a neural foramen stenosis detection subnetwork and a subarticular stenosis detection subnetwork. Based on this, the device can detect multiple conditions of spinal canal stenosis, neural foramen stenosis and subarticular stenosis, and realize the joint processing of multi-modal and multi-task. In each subnetwork, the feature extractor is used to fully mine the inter-frame feature correlation, which can improve the accuracy of disease detection and thus improve the detection efficiency of lumbar degenerative disease, providing efficient and robust auxiliary support for complex clinical scenarios. BRIEF DESCRIPTION OF DRAWINGS
[0015] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed in the embodiments. Obviously, the drawings described below are only some embodiments of the present application, and other drawings can also be obtained by those skilled in the art without creative labor.
[0016] Figure 1 A functional module diagram of a lumbar degenerative disease detection device in an embodiment of the present application;
[0017] Figure 2A structural diagram of a positioning task feature extractor and a classification task feature extractor provided by an embodiment of the present application is shown in FIG. 1;
[0018] Figure 3 A structural diagram of an MBConv module provided by an embodiment of the present application is shown in FIG. 2;
[0019] Figure 4 A structural diagram of a high-efficiency channel attention module provided by an embodiment of the present application is shown in FIG. 3;
[0020] Figure 5 A structural diagram of a channel-space fusion attention module provided by an embodiment of the present application is shown in FIG. 4;
[0021] Figure 6 A structural diagram of a gating recurrent unit provided by an embodiment of the present application is shown in FIG. 5; Figure 5
[0022] Figure 7 A structural diagram of a spatial attention module provided by an embodiment of the present application is shown in FIG. 6; Figure 5
[0023] Figure 8 A structural diagram of a gating recurrent unit provided by an embodiment of the present application is shown in FIG. 5;
[0024] Figure 9 A training process diagram of a positioning task feature extractor and a classification task feature extractor provided by an embodiment of the present application is shown in FIG. 7;
[0025] Figure 10 An image cutting diagram of a region of interest provided by an embodiment of the present application is shown in FIG. 8;
[0026] Figure 11 A comparison diagram of training curves of a positioning task extractor and three existing convolutional neural networks for first-stage coordinate regression and key frame index prediction provided by an embodiment of the present application is shown in FIG. 9;
[0027] Figure 12 A comparison diagram of visualization results of first-stage key frame index prediction and coordinate regression of EfficientNet-B0 provided by an embodiment of the present application is shown in FIG. 10;
[0028] Figure 13 A comparison diagram of visualization results of first-stage key frame index prediction and coordinate regression of SCANet2.5D provided by an embodiment of the present application is shown in FIG. 11;
[0029] Figure 14 A heat map of a sagittal T2-STIR weighted lumbar magnetic resonance image after adding an Improved EfficientNet-B0 module, an FAM module and a GRU module in the second stage respectively provided by an embodiment of the present application is shown in FIG. 12;
[0030] Figure 15 A heat map of a sagittal T1-weighted lumbar magnetic resonance image after adding an Improved EfficientNet-B0 module, an FAM module, and a GRU module in the second stage, provided by an embodiment of the present application;
[0031] Figure 16 A heat map of an axial T2-weighted lumbar magnetic resonance image after adding an Improved EfficientNet-B0 module, an FAM module, and a GRU module in the second stage, provided by an embodiment of the present application;
[0032] Figure 17 A flowchart of a detection method for a vertebral degenerative disease in another embodiment of the present application. DETAILED DESCRIPTION
[0033] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.
[0034] In order to make the objects, features, and advantages of the present application more obvious and easy to understand, the present application will be further described in detail below with reference to the drawings and specific embodiments.
[0035] As shown in Figure 1 The present application provides a detection device for a lumbar degenerative disease, which comprises:
[0036] An acquisition module 11 is configured to acquire a lumbar magnetic resonance image to be processed.
[0037] A preprocessing module 12 is connected to the acquisition module 11 and configured to pre-process the lumbar magnetic resonance image to be processed to obtain different image sequences.
[0038] A detection module 13 is connected to the preprocessing module 12 and configured to input the different image sequences into a trained computer-aided diagnosis network for a lumbar degenerative disease to perform detection, so as to determine the type of the lumbar degenerative disease and the lesion position. The type of the lumbar degenerative disease includes spinal stenosis, foraminal stenosis, and subarticular stenosis, and the lesion position is represented by an intervertebral disc coordinate. The computer-aided diagnosis network for the lumbar degenerative disease comprises a spinal stenosis detection subnetwork, a foraminal stenosis detection subnetwork, and a subarticular stenosis detection subnetwork. Each subnetwork comprises a plurality of feature extractors.
[0039] The detection module 13 comprises a detection unit 121 and a lumbar degenerative disease determination unit 122; the detection unit 121 is connected with the preprocessing module 12, and the detection unit 121 is configured to input different image sequences into the trained spinal stenosis detection sub-network, the neural foramen stenosis detection sub-network and the subarticular stenosis detection sub-network respectively for detection, so as to determine the lesion degree and lesion position of spinal stenosis, neural foramen stenosis and subarticular stenosis; the lumbar degenerative disease determination unit 122 is connected with the detection unit 121, and the lumbar degenerative disease determination unit 122 is configured to determine the type, lesion degree and lesion position of the lumbar degenerative disease according to the lesion degree and lesion position of spinal stenosis, neural foramen stenosis and subarticular stenosis, that is, to summarize the output of the detection unit 121. The neural foramen stenosis detection sub-network, the subarticular stenosis detection sub-network and the spinal stenosis detection sub-network each comprise a positioning task feature extractor and a classification task feature extractor.
[0040] As shown in Figure 2 , the positioning task feature extractor and the classification task feature extractor each comprise an Improved EfficientNet-B0 module, a channel-spatial fusion attention module and a gated recurrent unit connected in sequence; wherein the Improved EfficientNet-B0 module is obtained by replacing the squeeze-and-excitation attention module in the EfficientNet-B0 with an efficient channel attention module.
[0041] In one exemplary embodiment, the architecture of the EfficientNet-B0 is composed of an initial convolution module, a sequence of 16 core MBConv modules, a global average pooling layer, a random dropout layer and a classification layer, and the core innovation lies in the stacking structure of the MBConv module. Figure 3 is a structural diagram of the MBConv module. The MBConv module realizes efficient feature learning through the synergistic effect of depthwise separable convolution (two-dimensional convolution) and attention mechanism. The squeeze-and-excitation attention module in the MBConv module has two defects: (1) the channel compression operation will lose part of the inter-channel correlation information, weakening the network's learning ability for multi-channel feature interaction; (2) the dense matrix operation of the fully connected layer brings high computational overhead, which conflicts with the lightweight design goal. Therefore, in the present embodiment, the squeeze-and-excitation attention module in the EfficientNet-B0 is replaced with an efficient channel attention module to obtain the Improved EfficientNet-B0. Figure 4 is a structural diagram of the efficient channel attention module.
[0042] Specifically, the processing steps of the high-efficiency channel attention module for the image include: firstly, performing a global average pooling operation on the input feature map, then performing deformation (i.e., unfolding into one dimension), and applying a one-dimensional convolution operation to adaptively capture the local dependency between channels. After the convolution operation, the weight of each channel is limited between 0 and 1 using a Sigmoid activation function to generate a channel attention weight vector. Finally, the obtained channel attention weight is multiplied back to the original feature map channel by channel to generate a weighted feature map.
[0043] As shown in Figure 5 , the channel-spatial fusion attention module (FAM) is composed of a channel attention module (CAM) and a spatial attention module (SAM) in cascade, which respectively weights and enhances the features from the channel dimension and the spatial dimension to obtain the enhanced features. Specifically, the CAM module first aggregates the global statistical information of each channel, assigns higher weights to the channels related to the lesion through the self-attention mechanism, and thus strengthens the semantic expression thereof; the SAM module constructs a self-attention relationship in the spatial dimension, focuses on the region with diagnostic value, and improves the positioning accuracy of the lesion position. After the two modules are connected in series, the integrity of the original features can be maintained, and the anatomical structure and the small lesion can be enhanced, which can effectively cope with the challenges of low contrast and complex structure in medical images, and provide more discriminative feature representation for subsequent classification and segmentation tasks.
[0044] The core goal of the CAM module is to model the correlation between different channels and dynamically adjust the channel weights using these relationships to enhance the feature expression capability. As shown in Figure 6 , given an input feature matrix where C is the number of channels, H and W are the height and width of the feature map, respectively. First, A is regarded as a three-dimensional tensor, where each channel is an HxW feature map. In order to calculate the correlation between channels, A needs to be reshaped (also known as deformed) into N = HxW. Next, A' and A T are multiplied by matrix, and after Softmax nonlinear transformation, the channel attention matrix
[0045]
[0046] where x ji represents the influence of the i-th channel on the j-th channel. Then, X and A' are multiplied by matrix to obtain a feature enhancement matrix, which is rearranged into of 3D shape, so that it can be fused with the original feature. Finally, the feature enhancement matrix is connected with the original feature matrix by residual connection to obtain the final output matrix so that the output E j both contains global enhancement information and retains the original feature.
[0047]
[0048] where β is a hyperparameter for controlling the influence of attention. The final feature of each channel is the weighted sum of all channel features plus the original feature. The CAM module models the long-term semantic dependency between feature maps, which helps to improve the discriminability of features.
[0049] The core goal of the SAM module is to model the global dependency between pixels in the image, and enhance the semantic expression of key regions by dynamically adjusting the spatial weight. As shown in Figure 7 , first input the feature matrix where C is the number of channels, H and W are the height and width of the feature map respectively. A can be regarded as a three-dimensional tensor, and each pixel position has a C-dimensional feature vector. Then, linear mapping is performed on A to obtain new feature maps B and C, and Then, reshape (also known as warp) B and C into N = H x W. Next, perform matrix multiplication on B T and C, and obtain the spatial attention matrix
[0050]
[0051] where s ji represents the influence of pixel i on pixel j. At the same time, linearly map the feature A to the feature map and reshape (also known as warp) it into Then, perform matrix multiplication on D and S T to obtain the feature enhancement matrix and rearrange it into the 3D shape of so that it can be fused with the original feature. Finally, by residual connection, add the enhanced feature to the original feature A to obtain the final output matrix
[0052]
[0053] where α is a hyperparameter for controlling the influence of attention. The feature E jAny position of the two is the weighted sum of the corresponding enhanced feature and the original feature, a global context view is obtained, and mutual gain of similar semantic features is realized.
[0054] To break through the bottleneck of traditional 2D network in multi-slice information fusion, the embodiment innovatively embeds a Gated Recurrent Unit (GRU) in the classic 2D CNN skeleton, the structure of which is shown in Figure 8 , which is upgraded to an efficient "2.5D" hybrid architecture. Specifically, first, the 2D feature extraction module is used to extract the features of each MRI slice, and then these MRI slice features are sequentially input into the Gated Recurrent Unit module to capture the temporal correlation between adjacent slices. Both the advantages of 2D feature extraction in fine-grained local structure and the efficient integration of cross-slice information through the gating mechanism of the Gated Recurrent Unit are realized, thereby greatly reducing the computational and parameter overheads brought by 3D convolution while maintaining the timing sensitivity. The logical derivation process of the Gated Recurrent Unit is given below.
[0055] Update gate z t The Sigmoid function σ dynamically adjusts the retention ratio of the previous hidden state when z t is close to 1, the Gated Recurrent Unit tends to adopt the new candidate state; when z t is close to 0, the Gated Recurrent Unit tends to retain the hidden state of the previous step.
[0056] z t =σ(W z ·[H t-1 ,x t ]);
[0057] where W z is the learnable weight matrix of the update gate.
[0058] Reset gate r t decides which parts of the previous hidden state H should be retained when calculating the new candidate hidden state t-1 , that is, which historical information needs to be retained.
[0059] r t =σ(W r ·[H t-1 ,x t ]);
[0060] Candidate state The previous hidden state is modulated by the reset gate to obtain a "filtered" historical information, and then the current input feature x t is used to generate a new candidate hidden state.
[0061]
[0062] State update H t : Utilize update gate z t Fuse the hidden state H t-1 from the previous step with the new candidate hidden state to achieve adaptive memory and forgetting of information.
[0063]
[0064] The detection unit 121 includes a positioning subunit, a cutting subunit, and a classification subunit. Specifically as follows:
[0065] The positioning subunit is connected with the preprocessing module 12, and is used for inputting different image sequences into the positioning task feature extractors of the spinal stenosis detection subnetwork, the foraminal stenosis detection subnetwork, and the subarticular stenosis detection subnetwork respectively to perform lesion positioning, so as to obtain the lesion positions of the foraminal stenosis, the spinal stenosis, and the subarticular stenosis. The positioning subunit includes a first positioning layer, a second positioning layer, and a third positioning layer.
[0066] The first positioning layer is connected with the preprocessing module 12; the first positioning layer is used for inputting the sagittal T2-STIR weighted image sequence in the image sequence into the positioning task feature extractor in the spinal stenosis detection subnetwork to perform lesion positioning, so as to obtain the lesion position of the spinal stenosis; the lesion position is represented by the image index and the intervertebral disc coordinate.
[0067] The second positioning layer is connected with the preprocessing module 12 and the first positioning layer respectively, and is used for selecting the target axial T2 weighted image sequence from the axial T2 weighted image sequence in the image sequence according to the image index of the spinal stenosis, and inputting the target axial T2 weighted image sequence into the positioning task feature extractor of the subarticular stenosis detection subnetwork to perform lesion positioning, so as to obtain the lesion position of the subarticular stenosis. Specifically, the target axial T2 weighted image sequence is selected from the axial T2 weighted image sequence in the image sequence according to the image index in the output result of the first positioning layer.
[0068] The third positioning layer is connected with the preprocessing module 12, and is used for inputting the sagittal T1 weighted image sequence in the image sequence into the positioning task feature extractor of the foraminal stenosis detection subnetwork to perform lesion positioning, so as to obtain the lesion position of the foraminal stenosis.
[0069] The cutting sub-unit is connected with the positioning sub-unit, and is configured to cut images in each image sequence according to a lesion position of the spinal stenosis, the foraminal stenosis or the subarticular stenosis, so as to obtain a region of interest image of the spinal stenosis, the foraminal stenosis or the subarticular stenosis.
[0070] The classification sub-unit is connected with the cutting sub-unit, and is configured to input the region of interest image of the spinal stenosis, the foraminal stenosis or the subarticular stenosis into a classification task feature extractor of the spinal stenosis detection sub-network, the foraminal stenosis detection sub-network or the subarticular stenosis detection sub-network, so as to classify a disease type of the spinal stenosis, the foraminal stenosis or the subarticular stenosis, and obtain a lesion degree of the spinal stenosis, the foraminal stenosis or the subarticular stenosis.
[0071] In an example embodiment, the lumbar degenerative disease detection device further comprises a construction module and a training module.
[0072] The construction module is configured to construct the lumbar degenerative disease computer-aided diagnosis network.
[0073] The training module is connected with the construction module, and is configured to train the foraminal stenosis detection sub-network, the spinal stenosis detection sub-network and the subarticular stenosis detection sub-network by using the training sample of the foraminal stenosis, the training sample of the spinal stenosis and the training sample of the subarticular stenosis, so as to obtain the trained lumbar degenerative disease computer-aided diagnosis network.
[0074] The training process of the lumbar degenerative disease computer-aided diagnosis network is as follows:
[0075] Step 1: Select a suitable lumbar degenerative disease data set and perform uniform data preprocessing operations.
[0076] (1) Convert the original DICOM (Digital Imaging and Communications in Medicine) format magnetic resonance imaging data set into a JPG format that can be used for deep learning model training and human eye recognition, and divide it into sagittal T1 weighted images, sagittal T2-STIR weighted images and axial T2 weighted images according to the slice direction (sagittal plane) and the weighting method. Three types of image sequences. The magnetic resonance imaging data set in this embodiment is derived from the 2024 lumbar degenerative disease classification challenge.
[0077] (2) Random data augmentation is performed on the image sequence in (1).
[0078] (3) The image sequence in (2) is subjected to a zero channel completion strategy to make the data channel number consistent, so as to obtain the training sample for the lumbar degenerative disease computer-aided diagnosis network.
[0079] Step 2: Label the training samples.
[0080] Training a computer-aided diagnosis network for lumbar degenerative disease consists of two phases. The first phase trains the feature extractors for localization in the neuroforaminal stenosis, spinal canal stenosis, and subarticular stenosis detection subnetworks. The second phase trains the feature extractors for classification in the neuroforaminal stenosis, spinal canal stenosis, and subarticular stenosis detection subnetworks. Both phases employ triplet annotation for training examples. Phase one labels multiple normalized triplets, representing (image index, intervertebral disc horizontal coordinate, intervertebral disc vertical coordinate), to produce training examples for the localization task. The image index indicates the position of the image in the image sequence. Phase two labels also represent multiple normalized triplets, representing the severity of the lesion (mild, moderate, and severe), to produce training examples for the classification task. Even the absence of lesions is considered mild. For example, (0, 1, 0) represents moderate severity, and (0, 0, 1) represents severe severity. The first triplet represents the condition of the first intervertebral disc, the second triplet represents the condition of the second intervertebral disc, and so on, all intervertebral discs are labeled.
[0081] Step 3: Train the localization task feature extractors and classification task feature extractors in the neuroforaminal stenosis detection subnetwork, spinal canal stenosis detection subnetwork, and subarticular stenosis detection subnetwork respectively.
[0082] like Figure 9 As shown, the first stage: image index regression, intervertebral disc localization and cropping.
[0083] This stage combines coordinate regression with key frame index prediction strategy. The loss function of the positioning task feature extractor in this stage adopts the mean squared error loss function (MSELOSS).
[0084] Specifically, in the first stage, the localization task training samples, including sagittal T2-STIR modality training samples, sagittal T1 modality training samples, and axial T2 modality training samples, are used as input. To account for differences in the number of slices between patients, the output of the localization task feature extractor varies depending on the input modality training samples.
[0085] Specifically, sagittal T2-STIR modality training samples output five target image indices (supporting multiple discs with the same index) and five normalized disc coordinates. Sagittal T1 modality training samples and axial T2 modality training samples each output 10 target image indices and 10 normalized disc coordinates for bilateral spinal lesion localization. Figure 9 x2 in the formula (1) represents that for the subarticular stenosis and the foraminal stenosis, it is necessary to distinguish whether the left side is diseased or the right side is diseased, so the positioning task feature extractor outputs 10 triplets, corresponding to 10 images; x5 represents that there are five images in total, and the five images correspond to five intervertebral discs, which are the first intervertebral disc L1 / L2, the second intervertebral disc L2 / L3, the third intervertebral disc L3 / L4, the fourth intervertebral disc L4 / L5 and the fifth intervertebral disc L5 / S1, that is, one positioning task feature extractor outputs 5 triplets, corresponding to five intervertebral discs, also corresponding to five images, wherein L5 / S1 refers to the intervertebral disc between the 5th lumbar vertebra (Lumbar 5) and the first sacral vertebra (Sacral 1).
[0086] In the actual training process, in order to realize accurate identification of the position of a specific intervertebral disc in an axial surface magnetic resonance image, the application provides an axial surface slice selection method based on multi-modal image guidance, that is, to realize the axial position slice selection in the formula (1). Figure 9 The core of the method is to use the intervertebral disc rough positioning information which is easy to realize on the sagittal surface modality to guide the accurate positioning on the axial surface modality, so as to overcome the problems of regression model training difficulty and positioning accuracy decline caused by the high similarity between axial surface slices. For example, if the first triplet obtained by the sagittal surface T2 modality training sample is (0.2, 0.35, 0.5), and if there are 100 axial surface T2 weighted image sequences, then the 15th-25th axial surface T2 weighted image is selected as the target axial surface T2 modality training sample input into the positioning task feature extractor, that is, it is considered that the image index position corresponding to the first intervertebral disc should be between 15 and 25. The slice selection of the subsequent four intervertebral discs is also carried out in this way.
[0087] According to the coordinate center of the positioning task feature extractor output, the standardization of the region of interest is cut, and the region of interest images of the neural foramen stenosis, the spinal canal stenosis and the subarticular stenosis are obtained, and are used as the input of the second stage classification task feature extractor. This operation greatly removes the irrelevant tissue interference in the MRI image, and compared with the traditional segmentation and classification method, the demand for pixel-level data labeling is significantly reduced. The specific method of the region of interest cutting is: a mask with controllable size is created on the original image according to the predicted output coordinates. By performing an "and" operation between the mask and the original image, the region of interest image can be obtained, as shown in Figure 10 .
[0088] Second stage: classification.
[0089] The second stage training is completed by using the classification task feature extractor and the region of interest image obtained in the first stage. Specifically, the region of interest image extracted from the sagittal T2-STIR and the region of interest image extracted from the axial T2 are used together for the rating classification of spinal stenosis; the region of interest image extracted from the sagittal T1 is used for the rating classification of foraminal stenosis; and the region of interest image extracted from the axial T2 is used for the rating classification of foraminal stenosis. Among them, the spinal stenosis, the foraminal stenosis and the foraminal stenosis are all divided into mild degree, moderate degree and severe degree.
[0090] The loss function of this stage is Focal-G-Loss, and the expression of the Focal-G-Loss loss function is:
[0091]
[0092] Among them, is the classification loss, is the focal weighted loss, is the intra-group constraint loss, λ bce and λ G are weight coefficients.
[0093] The calculation formula of the intra-group constraint loss is:
[0094]
[0095] The calculation formula of the focal weighted loss is:
[0096]
[0097] Among them, is the intra-group constraint loss, G is the number of triplets, is the sum of the output probabilities of the feature extractor for the jth triplet classification task, is the focal weighted loss, is the binary cross-entropy loss, p t is the vector constructed according to the output probability of the classification task feature extractor and the real label, α is an optional balance factor, γ is a hyperparameter, N is the total number of classification tasks of the foraminal stenosis, the spinal stenosis or the subarticular stenosis, for example: the classification task of the foraminal stenosis has 5 intervertebral discs. Each intervertebral disc has 3 grades, so there are 5*3=15 classification tasks, so N is 15 at this time, and it is divided into 5 triplets. The classification task of the subarticular stenosis has five intervertebral discs, three grades, but is divided into left and right; so N is 30 at this time, and it is divided into 10 triplets. iis the output probability of the training sample of the i-th spinal stenosis classification task, the training sample of the spinal canal stenosis classification task or the training sample of the subarticular stenosis classification task in the feature extractor of the classification task, y i is the true label of the training sample of the i-th spinal stenosis classification task, the training sample of the spinal canal stenosis classification task or the training sample of the subarticular stenosis classification task.
[0098] The effectiveness of the lumbar degenerative disease computer-aided diagnosis network is verified, and the server under the Windows 11 operating system is used in this embodiment, and the server is configured as an Intel Core i9-12700K CPU and an NVIDIA 2080Ti GPU (11GB of video memory). The example uses the Python programming language and is implemented based on the PyTorch deep learning framework.
[0099] In the coordinate regression and key frame index prediction network of the first stage, the network uses the Adam optimizer, the initial learning rate is set to 1e-4, the batch size is set to 8, and the number of training iterations is 150. The loss function uses the mean square error loss (MSELoss), and the early stopping strategy is combined (when the validation set loss does not decrease for 10 consecutive iterations, the training is terminated in advance).
[0100] In the rating classification network of the second stage, the initial learning rate is set to 5e-5, the ReduceLROnPlateau learning rate decay strategy is used (when the validation set loss does not decrease for 5 consecutive iterations, the learning rate is adjusted), the batch size is set to 32, the number of iterations is 150, and the optimizer is still selected as the Adam optimizer. The loss function selects the Focal-G-Loss, and the early stopping mechanism is introduced.
[0101] Figure 11 is the training curve comparison diagram of the positioning task extractor and three existing convolutional neural networks for the first stage coordinate regression and key frame index prediction in this embodiment. Figure 12 and Figure 13 are the visualization result comparison diagrams of the positioning task extractor and EfficientNet-B0, SCANet2.5D for the first stage key frame index prediction and coordinate regression in this embodiment. By observing Figure 11-13 it can be known that the SCANet-2.5D of the application is designed through multi-dimensional feature enhancement and time sequence modeling, which not only significantly speeds up the convergence process, but also greatly improves the final coordinate regression accuracy.
[0102] By observing Figure 14 , Figure 15 and Figure 16It can be seen that the synergistic effect of the Improved EfficientNet-B0 module, the FAM module and the GRU module significantly enhances the ability of the network to focus on the lesion area under different image sequences, and improves the diagnostic accuracy of complex and subtle lesions. This multi-modal heat map comparison experiment highlights the universal advantage of the FAM and GRU modules and verifies the learning ability of SCANet-2.5D for multi-task and multi-modal medical image features.
[0103] From the observation of Table 1, Table 2 and Table 3, it can be seen that the present application, SCANet-2.5D, significantly outperforms existing methods in the rating tasks of three types of core lesions: the accuracy of spinal stenosis is 90.3%, the F1-score is 85.4%; the accuracy of nerve hole stenosis is 84.3%, the F1-score is 76.8%; the accuracy of subarticular stenosis is 83.3%, the F1-score is 83.1%.
[0104] Table 1
[0105] Research Multimodal data Multi-classification method Accuracy (%) F1-score (%) A × × Detection + Classification 82.7 84.8 B × √ Segmentation + Classification 72.7 / C × √ Detection + Classification 77.9 74.9 This program √ √ Positioning + Classification 90.3 85.4
[0106] Table 2
[0107] Research Multimodal data Multi-classification method Accuracy (%) F1-score (%) D × × Detection + Classification 88.0 71.8 E √ √ Segmentation + Classification 78.1 / F × × Detection + Classification 75.9 65.8 This program √ √ Positioning + Classification 84.3 76.8
[0108] Table 3
[0109] Research Multimodal data Multi-classification method Accuracy (%) F1-score (%) F × × Detection + Classification 71.0-75.0 / J √ √ Detection + Classification 77.9-79.3 78.1-80.3 This program √ √ Positioning + Classification 83.3 83.1
[0110] This advantage comes from the special design of SCANet-2.5D: (1) Fusion of sagittal T1 weighted image, sagittal T2-STIR weighted image and axial T2 weighted image, providing complementary contrast information. (2) Positioning-classification joint optimization, by regressing the intervertebral disc index and cropping the region of interest, effectively suppressing background interference and reducing labeling requirements. (3) Based on spatial-channel dual attention and GRU 2.5D time sequence modeling, which not only enhances the sensitivity to micro lesions, but also captures the temporal correlation of adjacent slices. (4) Multi-dimensional random data augmentation and self-developed Focal-G-Loss focus on difficult classification samples, which improves the network robustness.
[0111] These designs make SCANet-2.5D perform well in solving the problems of data set imbalance, reducing the dependence on data labeling and meeting the needs of complex clinical diagnosis. SCANet-2.5D overcomes the limitations of traditional methods and provides a reliable and efficient technical solution for the automatic diagnosis of lumbar degenerative diseases.
[0112] Based on the same inventive concept, the embodiment of the present application also provides a method for implementing the lumbar degenerative disease detection device. The implementation scheme for solving the problem provided by the method is similar to the implementation scheme described in the device, so the specific limitations in one or more lumbar degenerative disease detection method embodiments provided below can refer to the limitations of the lumbar degenerative disease detection device described above, which will not be repeated here.
[0113] In one exemplary embodiment, as shown in Figure 17 A lumbar degenerative disease detection device method is provided, which includes steps 101-103:
[0114] Step 101, acquiring a lumbar magnetic resonance image to be processed.
[0115] Step 102, preprocessing the lumbar magnetic resonance image to be processed to obtain different image sequences.
[0116] Step 103, inputting the different image sequences into the trained lumbar degenerative disease computer-aided diagnosis network for detection to determine the type, lesion degree and lesion position of the lumbar degenerative disease; wherein the type of lumbar degenerative disease includes spinal stenosis, nerve hole stenosis and joint under stenosis, and the lesion position is represented by intervertebral disc coordinates; the lumbar degenerative disease computer-aided diagnosis network includes a spinal stenosis detection sub-network, a nerve hole stenosis detection sub-network and a joint under stenosis detection sub-network; each sub-network includes a plurality of feature extractors.
[0117] The technical features of the above embodiments can be combined in any way. To make the description concise, not all possible combinations of the technical features in the above embodiments are described, but as long as the combinations of the technical features do not exist contradictions, they should be considered as the scope of the present application.
[0118] The principles and implementation modes of the present application are described by applying specific examples in this paper, and the above embodiment descriptions are only used to help understand the method and its core idea of the present application; at the same time, for those skilled in the art, according to the idea of the present application, the specific implementation mode and application range will be changed. In conclusion, the content of the present application should not be understood as a limitation.
Claims
1. A device for detecting lumbar degenerative disease, characterized in that: The method comprises: an acquisition module configured to acquire a lumbar magnetic resonance image to be processed; a preprocessing module connected to the acquisition module and configured to preprocess the lumbar magnetic resonance image to be processed to obtain different image sequences; a detection module connected to the preprocessing module and configured to input the different image sequences into a trained lumbar degenerative disease computer-aided diagnosis network to detect the type, lesion degree and lesion position of the lumbar degenerative disease, wherein the type of the lumbar degenerative disease includes spinal stenosis, foraminal stenosis and lower joint stenosis, the lesion position is represented by an image index and an intervertebral disc coordinate, and the lumbar degenerative disease computer-aided diagnosis network comprises a spinal stenosis detection subnetwork, a foraminal stenosis detection subnetwork and a lower joint stenosis detection subnetwork, each of which comprises a plurality of feature extractors.
2. The lumbar degenerative disease detecting apparatus according to claim 1, wherein The detection module comprises: a detection unit connected to the preprocessing module and configured to input the different image sequences into the trained spinal stenosis detection subnetwork, the foraminal stenosis detection subnetwork and the lower joint stenosis detection subnetwork respectively to detect the lesion degree and the lesion position of the spinal stenosis, the foraminal stenosis and the lower joint stenosis; a lumbar degenerative disease determination unit connected to the detection unit and configured to determine the type, lesion degree and lesion position of the lumbar degenerative disease based on the lesion degree and the lesion position of the spinal stenosis, the foraminal stenosis and the lower joint stenosis.
3. The apparatus for detecting lumbar degenerative disease according to claim 2, wherein The foraminal stenosis detection subnetwork, the lower joint stenosis detection subnetwork and the spinal stenosis detection subnetwork each comprise a positioning task feature extractor and a classification task feature extractor, the positioning task feature extractor and the classification task feature extractor each comprise an Improved EfficientNet-B0 module, a channel-spatial fusion attention module and a gated recurrent unit connected in sequence, and the Improved EfficientNet-B0 module is obtained by replacing the squeeze-and-excitation attention module in the EfficientNet-B0 with an efficient channel attention module.
4. The apparatus for detecting lumbar degenerative disease according to claim 3, wherein The detection unit comprises: a positioning subunit connected to the preprocessing module and configured to input the different image sequences into the positioning task feature extractors of the spinal stenosis detection subnetwork, the foraminal stenosis detection subnetwork and the lower joint stenosis detection subnetwork respectively to locate the lesions and obtain the lesion positions of the foraminal stenosis, the spinal stenosis and the lower joint stenosis; a cutting subunit connected to the positioning subunit and configured to cut the images in each image sequence according to the lesion positions of the foraminal stenosis, the spinal stenosis and the lower joint stenosis to obtain the region of interest images of the foraminal stenosis, the spinal stenosis and the lower joint stenosis; The classification subunit, connected with the cutting subunit, is used for inputting the images of the regions of interest of the spinal stenosis, the foraminal stenosis and the subarticular stenosis into the classification task feature extractors of the spinal stenosis detection subnetwork, the foraminal stenosis detection subnetwork and the subarticular stenosis detection subnetwork respectively to classify the types of diseases, and obtain the lesion degrees of the spinal stenosis, the foraminal stenosis and the subarticular stenosis.
5. The apparatus for detecting lumbar degenerative disease according to claim 4, wherein The positioning subunit comprises: The first positioning layer is connected with the preprocessing module and is used for inputting the sagittal T2-STIR weighted image sequence in the image sequence into the positioning task feature extractor in the spinal stenosis detection subnetwork to locate the lesion, and obtain the lesion position of the spinal stenosis; the lesion position is represented by the image index and the intervertebral disc coordinate; The second positioning layer is connected with the preprocessing module and the first positioning layer respectively, and is used for selecting the target axial T2 weighted image sequence from the axial T2 weighted image sequence in the image sequence according to the image index of the spinal stenosis, and inputting the target axial T2 weighted image sequence into the positioning task feature extractor of the subarticular stenosis detection subnetwork to locate the lesion, and obtain the lesion position of the subarticular stenosis; The third positioning layer is connected with the preprocessing module and is used for inputting the sagittal T1 weighted image sequence in the image sequence into the positioning task feature extractor of the foraminal stenosis detection subnetwork to locate the lesion, and obtain the lesion position of the foraminal stenosis.
6. The apparatus for detecting lumbar degenerative disease according to claim 3, wherein The detection device for the lumbar degenerative disease further comprises: The construction module is used for constructing the computer-aided diagnosis network for the lumbar degenerative disease; The training module is connected with the construction module and is used for training the foraminal stenosis detection subnetwork, the spinal stenosis detection subnetwork and the subarticular stenosis detection subnetwork respectively by using the training sample of the foraminal stenosis, the training sample of the spinal stenosis and the training sample of the subarticular stenosis, and obtaining the trained computer-aided diagnosis network for the lumbar degenerative disease.
7. The apparatus for detecting lumbar degenerative disease according to claim 6, wherein The loss functions in the training processes of the foraminal stenosis detection subnetwork, the spinal stenosis detection subnetwork and the subarticular stenosis detection subnetwork all comprise the mean square error loss function and the Focal-G-Loss loss function.
8. The apparatus for detecting lumbar degenerative disease according to claim 7, wherein The Focal-G-Loss loss function comprises the focal weighting loss and the intra-group constraint loss.
9. The lumbar degenerative disease detecting apparatus according to claim 8, wherein The expression of the Focal-G-Loss loss function is: wherein, is a classification loss, is a focal loss, is an intra-group constraint loss, λ bce and λ G are weight coefficients; The calculation formula of the intra-group constraint loss is: The calculation formula of the focal weighting loss is: wherein, is the intra-group constraint loss, G is the number of triplets, is the sum of the output probabilities of the feature extractor for the j-th triplet classification task, is the focal weighting loss, is the binary cross-entropy loss, p t is the vector constructed from the output probabilities of the classification task feature extractor and the true labels, a is an optional balancing factor, g is a hyperparameter, N is the total number of the spinal stenosis, spinal canal stenosis, or subarticular stenosis classification tasks, p i is the output probability of the classification task feature extractor for the i-th training sample of the neurospinal stenosis classification task, the training sample of the spinal canal stenosis classification task, or the training sample of the subarticular stenosis classification task, y i is the true label of the i-th training sample of the neurospinal stenosis classification task, the training sample of the spinal canal stenosis classification task, or the training sample of the subarticular stenosis classification task.
10. A method for detecting lumbar degenerative disease, characterized by, The detection method for the lumbar degenerative disease is applied to the detection device for the lumbar degenerative disease in any one of claims 1-9, and the detection method for the lumbar degenerative disease comprises: Obtaining the lumbar magnetic resonance image to be processed; Preprocessing the lumbar magnetic resonance image to be processed to obtain different image sequences; Different image sequences are input into the trained lumbar degenerative disease computer-aided diagnosis network for detection to determine the type, lesion degree and lesion position of the lumbar degenerative disease; wherein the type of the lumbar degenerative disease includes spinal canal stenosis, neural foramen stenosis and subarticular stenosis, and the lesion position is represented by an image index and a disc coordinate; the lumbar degenerative disease computer-aided diagnosis network comprises a spinal canal stenosis detection subnetwork, a neural foramen stenosis detection subnetwork and a subarticular stenosis detection subnetwork; each subnetwork comprises a plurality of feature extractors.