Method and system for predicting the risk of recurrence after inlay attention-enhancing pledget procedure
By constructing a LiteNet18 network model and combining lightweight residual blocks and MAFM attention modules, the problems of high computational cost and low prediction accuracy in the prediction of postoperative recurrence risk in PELD are solved, and efficient and accurate prediction is achieved in primary healthcare institutions.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- SHANDONG UNIV
- Filing Date
- 2026-02-13
- Publication Date
- 2026-05-12
AI Technical Summary
Existing technologies for predicting recurrence risk after PELD surgery suffer from high computational costs, low prediction accuracy, and poor adaptability. They are particularly difficult to deploy on low-computing-power devices in primary healthcare institutions, and existing deep learning models cannot effectively focus on subtle pathological features in lumbar MRI.
A LiteNet18 network model was constructed, employing an embedded enhanced attention mechanism, including lightweight residual blocks and MAFM attention modules. Through medical channel feature statistical calibration, lesion perception adaptive sampling, and cross-scale constraints, the network structure was optimized to adapt to lumbar MRI images, achieving accurate prediction of recurrence risk.
The model has reduced the number of parameters, improved prediction accuracy and computational efficiency, and can be deployed in primary healthcare institutions, reducing the rate of missed diagnoses and improving the accuracy of identifying recurrence risks.
Smart Images
Figure CN121709272B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of medical image processing and artificial intelligence technology, and in particular relates to a method and system for predicting the risk of recurrence after PELD with embedded attention enhancement. Background Technology
[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.
[0003] Percutaneous endoscopic lumbar discectomy (PELD) is the mainstream minimally invasive surgical procedure for treating lumbar disc herniation (LDH), offering advantages such as less muscle damage and less intraoperative bleeding. However, postoperative recurrence (r-LDH) still significantly impacts treatment outcomes. Current research shows that the incidence of r-LDH is 5%-18%, and patients with postoperative recurrence often have poor prognoses. Therefore, accurate preoperative assessment of recurrence risk and development of personalized treatment plans are of significant clinical importance.
[0004] Currently, r-LDH prediction mainly relies on physicians' subjective judgment of MRI images (such as the size of the herniated disc, Modic changes, etc.) and basic clinical information (such as age, BMI), which suffers from strong subjectivity in assessment and a single predictive dimension. While existing deep learning models (such as ResNet18) and traditional attention-enhancing models have been applied in image classification tasks, they still have significant limitations in r-LDH prediction scenarios.
[0005] (1) Model redundancy: The ResNet18 deep learning model, which performs well in the existing technology, has tens of millions of parameters, resulting in high computational costs when processing such medical images and making it difficult to deploy on low-computing-power equipment in primary medical institutions.
[0006] (2) Low prediction accuracy: To improve the focus on key areas in medical images, existing technologies generally adopt the traditional CBAM paradigm that relies on "pooling + MLP + convolution" to enhance attention. However, when processing small sample data of lumbar MRI, this method is prone to loss of subtle pathological features (such as potential signals of intervertebral disc degeneration) due to feature compression, and it cannot adapt to the characteristics of low contrast and fixed lesion scale in medical images. Summary of the Invention
[0007] To overcome the shortcomings of the prior art, the present invention provides a method and system for predicting the risk of recurrence after PELD with embedded enhanced attention. By embedding enhanced attention, the risk of recurrence after PELD can be accurately and efficiently predicted.
[0008] To achieve the above objectives, one or more embodiments of the present invention provide the following technical solutions:
[0009] The first aspect of this invention provides a method for predicting the risk of recurrence after PELD with embedded attention enhancement.
[0010] Embedded attention-enhancing methods for predicting postoperative recurrence risk in PELD include:
[0011] Acquire historical data of PELD patients and divide the training set by score matching;
[0012] A LiteNet18 network model is constructed, which includes an input layer, an initial convolutional module, a residual block group, and an output processing module. The residual block group contains four consecutive residual block stages, each of which consists of two lightweight residual blocks connected in series. Each lightweight residual block embeds a MAFM attention enhancement module to enhance the morphology and features of key recurrence features based on a triple collaborative mechanism.
[0013] The established LiteNet18 network model is trained based on the training set, and the trained LiteNet18 network model is used to analyze the lumbar spine medical images to be predicted, outputting the prediction results of the risk of recurrence after PELD surgery.
[0014] Furthermore, the initial convolution module includes a convolutional layer, a batch normalization layer, a GELU activation function, and a max pooling layer for preliminary extraction and downsampling of image features; the output processing module includes an adaptive pooling layer, a dropout layer, and a fully connected output layer.
[0015] Furthermore, the aforementioned triple collaborative mechanism includes medical channel feature statistical calibration, lesion perception adaptive sampling, and cross-scale constraints.
[0016] Furthermore, the medical channel feature statistical calibration includes: calculating the texture complexity of each channel based on the channel variance; calculating the pixel distribution asymmetry of each channel based on the channel skewness to adapt to lesion features under low contrast in medical images; and finally, weighting and fusing the channel variance and channel skewness using learnable parameters, and generating the final channel weights by Sigmoid activation.
[0017] Furthermore, the lesion perception adaptive sampling includes: locating candidate lesion areas based on the feature map after channel attention enhancement; calculating sampling weights by interpolating to simulate the scale perception of the candidate lesion areas; and generating final spatial weights by accumulating multi-scale sampling weights, followed by normalization and Sigmoid activation.
[0018] Furthermore, the cross-scale constraint includes: constraining the scale perception range of the model by fixing the prior medical lesion scale and the learnable scale weights.
[0019] Furthermore, the implementation of collaborative optimization of the lightweight residual block includes: main path optimization, first convolution sequence optimization, second convolution sequence optimization, intelligent adaptation of skip connections, and injection of fusion mechanism.
[0020] A second aspect of the present invention provides an embedded attention-enhancing postoperative recurrence risk prediction system for PELD.
[0021] An embedded attention-enhancing PELD postoperative recurrence risk prediction system includes:
[0022] The data acquisition module is configured to: acquire historical data of PELD patients and divide the training set by score matching;
[0023] The model building module is configured to build a LiteNet18 network model, which includes an input layer, an initial convolutional module, a residual block group, and an output processing module. The residual block group contains four consecutive residual block stages, each of which consists of two lightweight residual blocks connected in series. Each lightweight residual block contains an embedded MAFM attention enhancement module, which is used to enhance the morphology and features of key recurrence features based on a triple collaborative mechanism.
[0024] The model training module is configured to train the constructed LiteNet18 network model based on the training set.
[0025] The results prediction module is configured to analyze the lumbar spine medical images to be predicted using a trained LiteNet18 network model and output the prediction results of the risk of recurrence after PELD surgery.
[0026] A third aspect of the present invention provides a computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements the steps of the embedded attention-enhanced PELD postoperative recurrence risk prediction method as described in the first aspect of the present invention.
[0027] A fourth aspect of the present invention provides an electronic device including a memory, a processor, and a program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the embedded attention-enhanced PELD postoperative recurrence risk prediction method as described in the first aspect of the present invention.
[0028] The above one or more technical solutions have the following beneficial effects:
[0029] (1) This invention constructs a novel LiteNet18 network model, which includes an input layer, an initial convolutional module, a residual block group, and an output processing module. The residual block group contains four consecutive residual block stages, each of which consists of two lightweight residual blocks connected in series. Each lightweight residual block contains an embedded MAFM attention enhancement module. The lightweight residual block LiteBlock replaces traditional convolution with depthwise separable convolution, which reduces the number of parameters by 47.0% compared to ResNet18 and by 8.8% compared to traditional CBAM-enhanced LiteNet18. The inference speed is also improved accordingly, making it suitable for deployment on low-computing devices in primary healthcare institutions to meet the real-time clinical prediction needs.
[0030] (2) The triple collaborative mechanism of the MAFM module embedded in this invention includes medical channel feature statistical calibration, lesion perception adaptive sampling, and cross-scale constraints. Therefore, compared with the prior art, this invention can more accurately focus on subtle pathological features in lumbar MRI, making the identification of recurrence risk more accurate and the missed diagnosis rate lower.
[0031] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description
[0032] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.
[0033] Figure 1 This is a comparison chart of the ROC curves of the training and validation sets of the LiteNet-18 model in Embodiment 1 of the present invention.
[0034] Figure 2 This is a comparison of the ROC curves of the training and validation sets of the ResNet-18 model in the existing technology.
[0035] Figure 3 This is a heatmap comparing the confusion matrices of the LiteNet-18 model and the ResNet-18 model in Embodiment 1 of the present invention; wherein, Figure 3 In the diagram, (a) represents the confusion matrix heatmap of the LiteNet-18 model in the binary classification task of "relapsed (r-LDH) vs. non-relapsed (n-LDH)". Figure 3 In the diagram, (b) represents the confusion matrix heatmap of the ResNet-18 model in the binary classification task of "relapsed (r-LDH) vs. non-relapsed (n-LDH)".
[0036] Figure 4This is a structural diagram of the LiteNet-18 model in Embodiment 1 of the present invention.
[0037] Figure 5 This is a structural diagram of the LiteBlock module in Embodiment 1 of the present invention.
[0038] Figure 6 This is a structural diagram of the MAFM module in Embodiment 1 of the present invention. Detailed Implementation
[0039] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.
[0040] It should be noted that the terminology used herein is for the purpose of describing particular implementations only and is not intended to limit the exemplary implementations of the present invention.
[0041] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.
[0042] Overall concept: This invention provides a method for predicting the risk of recurrence after PELD (Percutaneous Pediatric Lung Disease) with embedded attention enhancement. The method includes:
[0043] Firstly, by embedding the LiteBlock lightweight residual block, it organically integrates depthwise separable convolution, MAFM attention module and residual connection for the first time. This not only solves the problem of redundant parameters in traditional ResNet, but also achieves precise focusing of relapse-related features through a triple collaborative mechanism, so as to better adapt to medical small sample scenarios.
[0044] Secondly, based on the pathological feature distribution of lumbar spine MRI T2WI images, an 18-layer lightweight network structure was optimized and designed. This was achieved through four-stage feature extraction and channel dimension gradient enhancement (64...). 128 256 512), to better balance feature representation capability and computational efficiency.
[0045] Thirdly, by combining PSM propensity score matching with multi-device image standardization processing, the problem of uneven distribution of small clinical sample data and equipment differences can be solved, thereby improving the model's generalization ability.
[0046] Fourth, a combination of GELU activation function, AdamW optimizer, cosine annealing scheduler and label smoothing loss is adopted, along with gradient clipping, to solve problems such as gradient vanishing and overfitting during small sample training, so as to improve the model convergence speed and prediction stability.
[0047] Example 1
[0048] This embodiment discloses a method for predicting the risk of recurrence after PELD with embedded attention enhancement.
[0049] Embedded attention-enhancing methods for predicting postoperative recurrence risk in PELD include:
[0050] Step S1: Obtain historical data of PELD patients and divide the training set by score matching;
[0051] Step S2: Construct a LiteNet18 network model, which includes an input layer, an initial convolutional module, a residual block group, and an output processing module. The residual block group contains four consecutive residual block stages, each of which consists of two lightweight residual blocks connected in series. Each lightweight residual block contains an embedded MAFM attention enhancement module, which is used to enhance the morphology and features of key recurrence features based on a triple collaborative mechanism.
[0052] Step S3: Train the established LiteNet18 network model based on the training set, and use the trained LiteNet18 network model to analyze the lumbar spine medical images to be predicted, and output the prediction results of the recurrence risk after PELD surgery.
[0053] Based on the above process, this invention, through embedded enhanced attention, can achieve accurate and efficient prediction of the risk of recurrence after PELD surgery. To facilitate understanding of the technical solution of this invention, the specific implementation methods of this invention will be further explained and described below.
[0054] In step S1, historical data of PELD patients are obtained, and the training set is divided by score matching.
[0055] In this embodiment, Python 3.9 was used for data processing. The dataset used was derived from 573 patients who received PELD treatment at a hospital from May 2018 to June 2023. After screening according to inclusion and exclusion criteria, a 1:2 ratio of propensity score matching (PSM) was used to balance the baseline differences between the two groups based on confounding factors such as age, BMI, history of diabetes, and smoking history, in order to address the problem of imbalanced distribution of small sample data. Finally, data from 325 patients were obtained (122 in the r-LDH group and 203 in the non-relapse group), which were divided into a training set (260 cases) and a test set (65 cases) in an 8:2 ratio.
[0056] Subsequently, image processing was performed: an experienced orthopedic surgeon selected the image with the largest cross-section of the herniated disc and labeled the region of interest (ROI). The ROI included the affected disc and the two adjacent discs above and below it. Another experienced radiologist then reviewed the contour image; any disagreements were resolved through discussion. All images were resized to 384×384 pixels, tensor transformed, and then standardized (mean=0.5, std=0.5) to eliminate systematic errors caused by differences between 1.5T and 3.0T MRI equipment, improving the model's cross-equipment compatibility.
[0057] In step S2, a LiteNet18 network model is constructed. The LiteNet18 network model includes an input layer, an initial convolutional module, a residual block group, and an output processing module. The residual block group contains four consecutive residual block stages, each of which consists of two lightweight residual blocks connected in series. Each lightweight residual block embeds a MAFM attention enhancement module, which is used to enhance the morphology and features of key recurrence features based on a triple collaborative mechanism.
[0058] like Figure 4 As shown, the specific architecture of the LiteNet18 network model includes:
[0059] 1) Input layer: used to receive 3-channel lumbar spine MRI T2WI sequence image data (preprocessed size is 384×384).
[0060] 2) Initial convolutional module: It consists of a 7×7 convolutional layer (stride 2, padding 3), a batch normalization (BatchNorm2d) layer, a GELU activation function, and a 3×3 max pooling layer (stride 2, padding 1), which is used to perform preliminary extraction and downsampling of image features.
[0061] 3) Residual Block Group: Contains 4 consecutive residual block stages (layer1-layer4), each stage consisting of 2 LiteBlocks connected in series, with 64 channels in each stage. 128 256 512, the feature map size is gradually halved by step size control (384). 192 96 48 twenty four).
[0062] like Figure 5As shown, the LiteBlock structure consists of a main path and residual branches, also known as shortcuts. The main path contains two convolutional layers, conv1 and conv2, each using 3×3 depthwise separable convolutions instead of traditional 3×3 convolutions. conv1 performs convolution within a single channel, outputting the same number of channels as the input. conv2 performs a weighted summation of all channel information, such as changing 64 channels to 128. The pathological information is then enhanced by the MAFM attention mechanism and fused with the features of the residual branches (shortcut(x)). After activation by the GELU function, LiteBlock(x) is obtained.
[0063] 4) MAFM Attention Enhancement Module: such as Figure 6 As shown, each LiteBlock embeds an original MAFM module, which enhances key features related to recurrence (such as intervertebral disc herniation morphology and paraspinal muscle signal features) and suppresses irrelevant background noise through the triple synergy of medical channel feature statistical calibration, lesion perception adaptive sampling and cross-scale constraints.
[0064] 5) Output processing module: The output goes through an adaptive average pooling layer (AdaptiveAvgPool2d, output size 1×1), a flattening operation, a dropout layer (dropout probability 0.3, adapted to small sample feature preservation) and a fully connected layer in sequence, and finally outputs two probability distributions, "recurrent (r-LDH)" and "non-recurrent", through the softmax function.
[0065] Furthermore, the MAFM attention module is an original design of this invention specifically addressing the characteristics of low contrast, strong concealment of recurrence-related features, and fixed lesion scale in lumbar spine MRI images. Specifically, the MAFM module completely abandons the traditional CBAM paradigm of "pooling + MLP + convolution," instead achieving precise focusing on subtle pathological features through a triple mechanism of "channel feature statistical calibration + lesion-aware adaptive sampling + cross-scale constraints." The specific implementation method is as follows:
[0066] 1) Channel attention branch: Medical channel feature statistical calibration.
[0067] Based on the statistical characteristics of lumbar MRI images (channel variance reflects texture complexity, and channel skewness reflects pixel distribution asymmetry), channel weights are dynamically generated without compression or restoration operations, which can avoid the loss of features in small samples.
[0068] Channel variance calculation: This reflects the texture complexity of each channel. The variance of lesion areas (such as degenerated intervertebral discs) is significantly higher than that of normal tissue. The calculation formula is as follows:
[0069] ;
[0070] in, For the first Each channel is in The eigenvalue at the location, For the first The average of each channel, and These are the height and width of the feature map, respectively; Indicates the first The spatial dimension variance of each channel feature map is used to reflect the degree of fluctuation in eigenvalues. Specifically... The larger the value, the higher the value. The more significant the difference in pixel values between channels, the more uniform the pixel values will be.
[0071] Channel skewness calculation: This reflects the asymmetry of pixel distribution in each channel, adapting to lesion features in low-contrast medical images. The calculation formula is as follows:
[0072] ;
[0073] in, To avoid smooth terms with a denominator of zero; Indicates the first The spatial skewness of each channel feature map reflects the "asymmetry" of pixel value distribution and complements the channel variance. Specifically, The larger the absolute value, the more likely it is to be the first. The pixel distribution of the i-th channel is severely asymmetrical, and vice versa. The pixel distribution of each channel is approximately symmetrical.
[0074] Channel attention weight generation: The channel weights are generated by weighting and fusing "channel variance + channel skewness" using learnable parameters and then activating with a Sigmoid function.
[0075] ;
[0076] in, These are channel-scale learnable parameters. This is the skewness weighting coefficient. For channel bias, learnable parameters. This is the Sigmoid activation function.
[0077] 2) Spatial attention branch: adaptive sampling for lesion perception.
[0078] Based on the prior size of lumbar MRI lesions (2~8px), an adaptive sampling mechanism of "downsampling-upsampling-difference calculation" is used to focus on high-response candidate lesion areas. There is no convolution operation, which can be adapted to small sample training.
[0079] High-response mask generation: Locating candidate lesion regions based on feature maps enhanced by channel attention.
[0080] ;
[0081] in, After enhancing channel attention Location feature value The global mean of the feature map; Indicating feature maps in the spatial attention branch High-response lesion candidate mask value at location, If the feature response at that location is higher than the global average, it is identified as a candidate lesion area. If the feature response at that location is lower than or equal to the global average level, it is determined to be a background or noise area.
[0082] Multi-scale sampling weight calculation: targeting common scales of medical lesions Scale perception is simulated through interpolation, and sampling weights are calculated:
[0083] ;
[0084] ;
[0085] ;
[0086] in, It is a scaling factor, and ; For interpolation operations, Scale-learnable weights; This is a temperature coefficient used to control the sampling focus. Indicates the first Downsampled feature maps at various scales are used to simulate the local receptive field of lesions at that scale; Indicates will Upsample the scale-adapted feature map back to its original size, maintaining consistency with... Consistent spatial dimensions are used to calculate feature differences; Indicates the first Cross-scale constraint sampling weights at each scale; This represents the channel-enhanced feature map output by the channel attention branch after channel recalibration, and serves as the initial input to the spatial attention branch.
[0087] Spatial attention weight generation: Accumulate multi-scale sampling weights, normalize and apply sigmoid activation to generate the final spatial weights.
[0088] .
[0089] 3) Cross-scale constraint mechanism.
[0090] By using a fixed medical lesion scale prior (2~8px) and learnable scale weights This constrains the scale perception range of the model to avoid scale learning chaos under small samples and ensures adaptability to lesions of different sizes.
[0091] Furthermore, LiteBlock's lightweight residual block (core innovation module) is designed to adapt to the low-computing-power equipment of primary healthcare institutions while ensuring feature extraction capabilities under small sample sizes. Its core innovation lies in the organic integration of depthwise separable convolution, the original MAFM attention module, and intelligent residual connections to achieve synergistic optimization of "lightweight, strong expression, and precise focusing." The specific implementation path includes:
[0092] 1) Main path optimization: A depthwise separable convolution structure of "3×3 depthwise convolution + 1×1 pointwise convolution" is adopted to decompose the traditional convolution operation and significantly reduce the number of parameters.
[0093] 2) First convolution sequence: 3×3 depthwise convolution (groups = number of input channels) → 1×1 pointwise convolution → batch normalization → GELU activation, to achieve preliminary feature extraction and channel fusion.
[0094] 3) Second convolution sequence: 3×3 depthwise convolution → 1×1 pointwise convolution → batch normalization, further refine the feature representation and avoid information loss.
[0095] 4) Shortcut Intelligent Adaptation: When the number of input and output channels or their dimensions are inconsistent, dimensionality matching is performed through 1×1 convolutional layers and batch normalization layers to ensure effective propagation of residual gradients; when the size and number of channels are consistent, direct identity mapping is used to reduce redundant computation.
[0096] ;
[0097] in, Represents the input feature map The number of channels, This represents the number of channels in the output feature map of the main path of the residual block. This indicates the stride of the convolutional layer in the main path of the residual block; This indicates the input feature map Perform batch normalization. This indicates that a 1×1 convolution is performed on the feature map after BN processing; and When, the output is At this point, the residual branch uses an identity mapping; otherwise, the output is... .
[0098] 5) Innovative Fusion Mechanism: Features extracted from the main path via depthwise separable convolution are first weighted and enhanced with key pathological information using the MAFM attention module, then added to the residual features of skip connections, and finally output through the GELU activation function.
[0099] ;
[0100] in, This represents a 3×3 depthwise convolution of the LiteBlock main path. This represents a 1×1 pointwise convolution of the LiteBlock main path; This indicates that lesion-related features are enhanced through channel attention and spatial attention. This represents the Gaussian error linear unit activation function.
[0101] In step S3, the established LiteNet18 network model is trained based on the training set, and the trained LiteNet18 network model is used to analyze the lumbar spine medical images to be predicted, outputting the prediction result of the risk of recurrence after PELD surgery. This can be achieved specifically through the following methods:
[0102] The first step is to instantiate the MAFM-enhanced LiteNet18 model, set the number of output classes to 2 (r-LDH group, non-relapse group), and load it onto the GPU (with support for CPU fallback).
[0103] The second step is to use the SimpleLabelSmooth loss function to adapt to small sample label noise. The calculation formula is as follows:
[0104] ;
[0105] in, For the batch sample size, For the number of categories, For one-hot tags, Predict probabilities for the model.
[0106] Based on this, AdamW is used as the optimizer, cosine annealing learning rate scheduler, and gradient clipping, which work together to improve the stability of training with few samples. The initial learning rate of the optimizer is set to 5e-5, and the weight decay is set to 0.0001 to suppress overfitting. The maximum period of the cosine annealing learning rate scheduler is set to T_max=50 to dynamically adjust the learning rate. The maximum L2 norm of gradient clipping is set to max_norm=1.0.
[0107] The third step involves setting the model to training mode, reading training data in batches (batch size=8, adapted for small sample gradient stability), obtaining prediction results through forward propagation, and calculating the loss value; then updating the model parameters through backpropagation. Weak data augmentation (random horizontal flipping only, p=0.2) is used on the training set to avoid distortion of lesion morphology; performance is evaluated on the test set after each training round, and the weights of the model with the highest validation accuracy are saved.
[0108] Step 4, training termination: Stop training after 50 iterations or when the validation loss does not decrease for 3 consecutive iterations, and output the best model.
[0109] The fifth step involves inference based on the trained model, specifically including:
[0110] 1) Model loading: Load the best trained model weights and set the model to evaluation mode (turn off dropout layer).
[0111] 2) Image preprocessing: The input lumbar spine MRI T2WI image is resized (384×384), tensor transformed, and standardized (mean=0.5, std=0.5) to generate tensors that meet the model input requirements.
[0112] 3) Forward propagation: Tensor input model, through initial convolution, residual block group, attention module and output processing module to calculate two probability distributions.
[0113] 4) Output results: Output the probability value of “recurrence (r-LDH)”, combined with clinical risk stratification (low risk: probability <30%, medium risk: 30%≤probability <60%, high risk: probability ≥60%), to provide clinicians with a reference for preoperative recurrence risk assessment.
[0114] To further demonstrate the superiority of this invention, the following experiments were conducted to compare it with the existing ResNet18 model. The hardware environment used in the experiments was: Intel Core i7-12700H CPU, NVIDIA RTX 3090 GPU (24GB VRAM), and 32GB of RAM; the software environment was: Ubuntu 20.04 operating system, PyTorch 1.12.1 deep learning framework, Python version 3.9, and dependent libraries including torchvision, numpy, matplotlib, SimpleITK (for MRI image reading), and PyRadiomics (for feature extraction and verification). The comparison results are shown in Table 1.
[0115] Table 1. Performance comparison between LiteNet-18 and ResNet-18 models of this invention.
[0116]
[0117] As shown in Table 1, the prediction method of the present invention reduces the number of parameters by 49%, while improving the test set AUC by 5 percentage points and the Sensitivity by 4 percentage points, resulting in better overall performance and better meeting the needs of clinical r-LDH preoperative prediction.
[0118] like Figure 1 The ROC curves of LiteNet18 training and validation sets are shown in the figure to evaluate the model's classification performance and generalization ability. The horizontal axis (false positive rate) represents the probability that the model will incorrectly classify a negative sample as a positive sample. The smaller the value, the lower the misclassification rate of the model. The vertical axis (true positive rate) represents the probability that the model will correctly classify a positive sample as a positive sample. The larger the value, the higher the recognition accuracy of the model. Figure 1 The dashed line represents the reference line for random guessing (AUC=0.5). If the model performance is the same as random guessing, the curve will coincide with this line; the farther the curve is from this line, the better the model performance. The blue curve is the ROC curve of the training set, labeled AUC=0.91; the red curve is the ROC curve of the validation set, labeled AUC=0.88; AUC (area under the ROC curve) is the core indicator for measuring the model's classification ability.
[0119] Figure 2 Similarly, the blue curve represents the ROC curve of the training set, labeled as AUC=0.89; the red curve represents the ROC curve of the validation set, labeled as AUC=0.83.
[0120] like Figure 3 The image shows a heatmap comparing the confusion matrices of the LiteNet-18 and ResNet-18 models; where, Figure 3 In the diagram, (a) represents the confusion matrix heatmap of the LiteNet-18 model in the binary classification task of "relapsed (r-LDH) vs. non-relapsed (n-LDH)". Figure 3 (b) in the diagram represents a heatmap of the confusion matrix of the ResNet-18 model in the binary classification task of "relapse (r-LDH) and non-relapse (n-LDH)," used to show the model's classification results (the matching between the true and predicted labels). The core of the confusion matrix is the four combinations of "true label" and "predicted label" (with "non-relapse" as negative samples and "relapse" as positive samples), specifically:
[0121] TN (True Negative): True label = no relapse, predictive label = no relapse (correctly identifying "no relapse");
[0122] FP (False Positive): True label = non-relapse, predictive label = relapse (mistakenly classifying "non-relapse" as "relapse");
[0123] FN (False Negative): True label = relapse, predictive label = non-relapse (mistakenly classifying "relapse" as "non-relapse");
[0124] TP (True Positive): True label = relapse, Predictive label = relapse (correctly identifying "relapse");
[0125] It should be noted that, Figure 3 In figures (a) and (b), the larger the value of the color bar on the right, the darker the color.
[0126] Example 2
[0127] This embodiment discloses a PELD postoperative recurrence risk prediction system with embedded attention enhancement.
[0128] An embedded attention-enhancing PELD postoperative recurrence risk prediction system includes:
[0129] The data acquisition module is configured to: acquire historical data of PELD patients and divide the training set by score matching;
[0130] The model building module is configured to build a LiteNet18 network model, which includes an input layer, an initial convolutional module, a residual block group, and an output processing module. The residual block group contains four consecutive residual block stages, each of which consists of two lightweight residual blocks connected in series. Each lightweight residual block contains an embedded MAFM attention enhancement module, which is used to enhance the morphology and features of key recurrence features based on a triple collaborative mechanism.
[0131] The model training module is configured to train the constructed LiteNet18 network model based on the training set.
[0132] The results prediction module is configured to analyze the lumbar spine medical images to be predicted using a trained LiteNet18 network model and output the prediction results of the risk of recurrence after PELD surgery.
[0133] Example 3
[0134] The purpose of this embodiment is to provide a computer-readable storage medium.
[0135] A computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the steps in the embedded attention-enhanced PELD postoperative recurrence risk prediction method as described in Embodiment 1 of this disclosure.
[0136] Example 4
[0137] The purpose of this embodiment is to provide an electronic device.
[0138] An electronic device includes a memory, a processor, and a program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps in the embedded attention-enhanced PELD postoperative recurrence risk prediction method as described in Embodiment 1 of this disclosure.
[0139] The steps and methods involved in the apparatuses of Embodiments 2, 3, and 4 above correspond to those in Embodiment 1. For specific implementation details, please refer to the relevant description section of Embodiment 1. The term "computer-readable storage medium" should be understood as a single medium or multiple media including one or more instruction sets; it should also be understood as including any medium capable of storing, encoding, or carrying an instruction set for execution by a processor and enabling the processor to perform any of the methods in this invention.
[0140] Those skilled in the art will understand that the modules or steps of the present invention described above can be implemented using general-purpose computer devices. Optionally, they can be implemented using computer-executable program code, thereby allowing them to be stored in a storage device for execution by a computer device, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. The present invention is not limited to any particular combination of hardware and software.
[0141] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.
Claims
1. A method for predicting the risk of recurrence after PELD (Postoperative Pediatric Endoscopic Disorder) with embedded attention enhancement, characterized in that, include: Acquire historical data of PELD patients and divide the training set by score matching; A LiteNet18 network model is constructed, comprising an input layer, an initial convolutional module, a residual block group, and an output processing module. The residual block group contains four consecutive residual block stages, each consisting of two lightweight residual blocks connected in series. Each lightweight residual block embeds a MAFM attention enhancement module to enhance the morphology and features of key recurrence characteristics based on a triple collaborative mechanism. This triple collaborative mechanism includes medical channel feature statistical calibration, lesion perception adaptive sampling, and cross-scale constraints. The implementation of collaborative optimization of the lightweight residual blocks includes: main path optimization, first convolutional sequence optimization, second convolutional sequence optimization, intelligent adaptation of skip connections, and injection of a fusion mechanism. The main path optimization employs a depthwise separable convolutional structure of 3×3 depthwise convolution + 1×1 pointwise convolution. The established LiteNet18 network model is trained based on the training set, and the trained LiteNet18 network model is used to analyze the lumbar spine medical images to be predicted, outputting the prediction results of the risk of recurrence after PELD surgery.
2. The method for predicting the risk of recurrence after PELD surgery with embedded attention enhancement as described in claim 1, characterized in that, The initial convolution module includes a convolutional layer, a batch normalization layer, a GELU activation function, and a max pooling layer, used for preliminary extraction and downsampling of image features; the output processing module includes an adaptive pooling layer, a dropout layer, and a fully connected output layer.
3. The method for predicting the risk of recurrence after PELD surgery with embedded attention enhancement as described in claim 1, characterized in that, The medical channel feature statistical calibration includes: calculating the texture complexity of each channel based on channel variance; calculating the pixel distribution asymmetry of each channel based on channel skewness to adapt to lesion features under low contrast in medical images; and finally, weighting and fusing the channel variance and channel skewness using learnable parameters, and generating the final channel weights by Sigmoid activation.
4. The method for predicting the risk of recurrence after PELD surgery with embedded attention enhancement as described in claim 1, characterized in that, The lesion perception adaptive sampling includes: locating candidate lesion areas based on feature maps enhanced by channel attention; calculating sampling weights by interpolating to simulate scale perception of the candidate lesion areas; and generating final spatial weights by accumulating multi-scale sampling weights, followed by normalization and Sigmoid activation.
5. The method for predicting the risk of recurrence after PELD surgery with embedded attention enhancement as described in claim 1, characterized in that, The cross-scale constraint includes: constraining the scale perception range of the model by fixing the prior scale of medical lesions and the learnable scale weights.
6. An embedded attention-enhancing PELD postoperative recurrence risk prediction system, characterized in that, include: The data acquisition module is configured to: acquire historical data of PELD patients and divide the training set by score matching; The model building module is configured to: build a LiteNet18 network model, which includes an input layer, an initial convolutional module, a residual block group, and an output processing module; wherein, the residual block group contains four consecutive residual block stages, each of which consists of two lightweight residual blocks connected in series, and each lightweight residual block embeds an MAFM attention enhancement module to enhance the morphology and features of key recurrence features based on a triple collaborative mechanism; the triple collaborative mechanism includes medical channel feature statistical calibration, lesion perception adaptive sampling, and cross-scale constraints; the implementation of the collaborative optimization of the lightweight residual blocks includes: main path optimization, first convolutional sequence optimization, second convolutional sequence optimization, skip connection intelligent adaptation, and fusion mechanism injection, wherein the main path optimization adopts a depthwise separable convolutional structure of 3×3 depthwise convolution + 1×1 pointwise convolution; The model training module is configured to train the constructed LiteNet18 network model based on the training set. The results prediction module is configured to analyze the lumbar spine medical images to be predicted using a trained LiteNet18 network model and output the prediction results of the risk of recurrence after PELD surgery.
7. A computer-readable storage medium having a program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps in the embedded attention-enhanced PELD postoperative recurrence risk prediction method as described in any one of claims 1-5.
8. An electronic device, comprising a memory, a processor, and a program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps in the method for predicting the risk of recurrence after PELD with embedded enhanced attention as described in any one of claims 1-5.