Lumbar senescence assessment method and system based on age estimation and digital biomarkers

Abstract

The invention discloses a method for evaluating lumbar vertebra aging based on a lumbar vertebra age estimation network of slice contrast learning, mainly solving the problem that the prior art cannot meet individualized and continuous quantitative early warning of precision medical treatment, and the scheme comprises the following steps: obtaining a lumbar vertebra MRI image and dividing the image into a training set and a test set; training and testing an existing segmentation network to obtain a segmentation result of the lumbar anatomical structure, and performing grouping cutting on the segmentation result according to lumbar segments to obtain a lumbar anatomical structure image data set; dividing the cut images of the first lumbar vertebra, the second lumbar vertebra and the third lumbar vertebra into a training set and a test set for lumbar vertebra age estimation; constructing a lumbar vertebra age estimation network of slice contrast learning, and training the lumbar vertebra age estimation network by using a lumbar vertebra age estimation training set; inputting the lumbar vertebra age estimation test set into the trained lumbar vertebra age estimation model, and outputting a predicted age result and a lumbar vertebra age estimation significance heat map; and according to an age prediction result, evaluating a degenerative aging risk result of different anatomical regions of the lumbar vertebra. According to the method, automatic segmentation of the lumbar intervertebral disc anatomical structure and estimation of the lumbar vertebra age can be realized, and the method can be used for lumbar vertebra age and senescence risk assessment and early risk early warning.

Description

Technical Field

[0001] This invention belongs to the field of image processing technology, and specifically relates to a method and system for assessing lumbar spine aging, which can be used for automatic segmentation of lumbar intervertebral disc anatomy, estimation of lumbar spine age, aging assessment based on digital markers, and early risk warning. Background Technology

[0002] Lower back pain and intervertebral disc herniation caused by intravertebral disc degeneration (IVDD) are common age-related diseases closely associated with human aging. The incidence of these diseases is rising globally, affecting more than 65 million people annually. If IVDD is not detected and effectively treated early, it can eventually develop into spinal instability and symptoms of spinal cord and nerve root compression, severely impacting an individual's quality of life and imposing a significant medical and economic burden on society. Currently, IVDD diagnosis primarily relies on imaging modalities such as magnetic resonance imaging (MRI) and computed tomography (CT). While these technologies provide information about the current structural state of the intervertebral disc, they lack the ability to predict the long-term risk of IVDD progression.

[0003] Treatment strategies for IVDD fall into four tiers: conservative treatment, interventional therapy, minimally invasive surgery, and spinal fixation and fusion. However, clinical decisions regarding the optimal treatment tier still largely rely on clinicians' experience and personal judgment, rather than evidence-based personalized guidance derived from large-scale clinical data analysis. This experience-based paradigm increases the likelihood of overtreatment or suboptimal treatment outcomes. Developing prognostic tools capable of predicting the early trajectory and progression rate of IVDD is a critical unmet need. Such predictive capabilities would significantly improve the selection of personalized treatments and the optimization of treatment outcomes, thereby addressing the limitations of current clinical management models.

[0004] Lumbar spine age estimation is used to assess the growth, development, and degenerative changes of the lumbar spine, thereby understanding the current physiological stage and the potential risk of future lumbar spine diseases. This method helps physicians assess the degree of lumbar spine degeneration and predict abnormal aging conditions that may not be consistent with actual age. Traditional lumbar spine age estimation methods, including biomarker-based methods and imaging techniques, have their limitations. Biomarkers, such as serum markers, can reflect bone metabolism, but are influenced by external factors such as nutrition and disease, and lack specificity. Imaging methods, such as bone morphology analysis, are subjective and prone to error. For example, studies involving lumbar MRI and X-ray images rely on visual assessment and feature extraction, which are both time-consuming and prone to bias. Recent advances have involved the use of deep learning techniques to improve the accuracy and efficiency of lumbar spine age estimation.

[0005] Regarding qualitative grading, Yang et al. [discussed] Automatic Diagnosis and Grading of LumbarDisc Degeneration via The research on Deep Learning with a Novel Region-Feature Recalibration Network proposed a region feature recalibration network, RFR-Net, for fine-grained classification of lumbar disc degeneration. It enhances the model's ability to focus on key regions by introducing an attention mechanism. While this method achieved good performance on the Pfirrmann classification task, it remains essentially a "black box" end-to-end classification model. Its decision-making process lacks clear anatomical and radiological basis, resulting in poor interpretability. Therefore, it is difficult for clinicians to fully trust and use it for critical decision-making.

[0006] In terms of quantitative measurement, Zheng et al. made significant progress in Deep Learning-based High-precision Quantification of Lumbar. The study of intervertebral disc degeneration from MRI utilized deep learning to achieve high-precision automated measurement of key lumbar spine anatomical structures. It used semantic segmentation networks such as U-Net to perform pixel-level analysis of lumbar spine MRI images, automatically delineating key anatomical structures such as intervertebral discs and the spinal canal; subsequently, it programmed quantitative indicators such as intervertebral disc height and spinal canal diameter using geometric formulas. While it provides objective and repeatable quantitative data, the measurement results lack an automated and interpretable mapping bridge between the final clinical diagnosis, grading, and treatment plan, requiring additional comprehensive judgment from physicians, thus indicating insufficient level of intelligence.

[0007] Regarding automatic grading, patent document CN115049898B discloses an automatic grading method for lumbar intervertebral disc degeneration based on region block feature enhancement and inhibition. This method uses a target detection network such as DETR to locate the intervertebral disc and then employs a convolutional neural network to discretely grade the degree of degeneration, improving the automation level of the process. However, this method cannot achieve a continuous and refined quantitative description of the degree of degeneration, and the assessment process is decoupled from individual age and physiological background, thus making it difficult to achieve a continuous quantitative description and individualized risk stratification of lumbar degeneration.

[0008] In summary, existing medical imaging-based methods for assessing lumbar spine degeneration fail to meet the core requirements of precision medicine for personalized and continuously quantitative early warning. While age assessment technologies for organs such as the brain are relatively mature, the unique anatomy and degenerative mechanisms of the lumbar intervertebral discs make it difficult for these existing methods to accurately model and quantify the aging state of the lumbar spine at different ages and under different degenerative patterns. Therefore, developing a new paradigm that integrates three-dimensional imaging information, provides interpretable continuous quantitative indicators, and enables personalized stratified assessment is crucial to overcoming current bottlenecks in clinical management. Summary of the Invention

[0009] The purpose of this invention is to propose a method and system for assessing lumbar spine aging based on age estimation and digital biomarkers, so as to overcome the shortcomings of existing technologies that can only perform discrete grading, lack continuous quantification capabilities and individualized prediction capabilities, and realize refined assessment and stratified early warning of the degree of lumbar intervertebral disc degeneration.

[0010] The technical approach to achieving the objective of this invention is as follows: a lumbar spine age estimation network based on slice comparison learning captures cross-layer correlation features of three-dimensional magnetic resonance images, and integrates a similarity comparison learning strategy between adjacent slices to generate accurate age estimation results; by constructing two types of digital biomarkers, Age Delta and Age Selta, hierarchical early warning of physiological and pathological degeneration is achieved; by outputting significant heatmaps and risk scatter plots, the internal degeneration patterns of the intervertebral disc are visually displayed, thereby improving the interpretability of the model and providing more reliable quantitative evidence for clinical diagnosis and treatment decisions.

[0011] Based on the above ideas, the technical solution of the present invention includes:

[0012] 1. A lumbar spine age estimation network based on slice comparison learning, characterized in that it comprises:

[0013] An encoder consisting of seven layers of different convolutional modules was built to process two-dimensional slices of each lumbar vertebral anatomical structure, extract feature information, and obtain the output after encoding two branches.

[0014] An aggregation module is built, which includes vector query Q, key K and value V, to determine the attention weight at each position in the output sequence. The weighted sum is then used for age estimation.

[0015] A fully connected regression layer is constructed, which includes a weight matrix and a bias vector, to derive the predicted age.

[0016] A neighboring slice contrast learning module is established, comprising a cascaded projector and predictor, which calculates the neighboring slice contrast loss based on the outputs of the projector and predictor. This loss is then used as the regularization term and the mean squared error loss of the main regression. The weighted summation constitutes the total loss function of the network. ;

[0017] The encoding module, aggregation module, fully connected regression layer, and adjacent slice comparison learning module are cascaded in sequence to form a slice comparison learning lumbar spine age estimation network.

[0018] Furthermore, the seven convolutional layers in the encoder have the following structural layers and connection relationships:

[0019] The first five convolutional modules have the same structure. Each module includes a 3×3 two-dimensional convolutional layer, an instance normalization layer, a max pooling layer, and a ReLU activation layer, which are connected in sequence to extract and compress multi-scale local features of the slices step by step.

[0020] The sixth convolutional module includes a 1×1 two-dimensional convolutional layer, an instance normalization layer, and a ReLU activation layer, which are connected in sequence to fuse and integrate the semantic information of deep features.

[0021] The seventh convolutional module includes an average pooling layer, a culling layer, and a 1×1 two-dimensional convolutional layer, which are connected in sequence to aggregate global information and generate the final slice feature encoding.

[0022] The seven convolutional modules are cascaded in sequence to form an encoder.

[0023] Furthermore, the projector and predictor in the adjacent slice similarity loss module have the following structures and transmission relationships:

[0024] The projector includes at least one fully connected layer for mapping high-dimensional features extracted by the similarity loss encoder to a low-dimensional projection space that is more suitable for similarity comparison.

[0025] The predictor includes at least one fully connected layer for predicting the projection features of another adjacent slice from the low-dimensional projection features of one slice, and the output is used to calculate feature similarity to construct an auxiliary loss.

[0026] 2. A method for assessing lumbar spine aging using a lumbar spine age estimation network based on the above-mentioned slice comparison learning, characterized in that it includes:

[0027] (1) Obtain the dataset of lumbar spine MRI images and perform windowing processing, dividing it into training and test sets at a ratio of 1:9:

[0028] (2) The existing nnFormer segmentation network is trained using the training set and tested using the test set to obtain the segmentation results of the lumbar anatomical structure, thus forming a lumbar vertebra segmentation image dataset.

[0029] (3) The lumbar spine segmentation image dataset is grouped and cropped according to lumbar spine segments to obtain an image dataset of lumbar spine anatomy;

[0030] (4) Group the image dataset of lumbar anatomical structures by gender and assess the overall trend of lumbar degeneration in the population;

[0031] (5) The images of the first to third lumbar vertebrae in the lumbar anatomical structure image dataset are divided into a training set and a test set for lumbar age estimation in a 7:3 ratio.

[0032] (6) Input the training set of lumbar spine age estimation into the lumbar spine age estimation network of slice comparison learning, and use the Adam optimizer to iteratively update the model parameters in the manner of random single sample learning each time until the preset number of times, so as to obtain the trained lumbar spine age estimation model.

[0033] (7) Input the test set of lumbar spine age estimation into the trained lumbar spine age estimation model, and output the predicted age results and the significance heatmap of lumbar spine age estimation.

[0034] (8) Predicted age results: Define digital biomarker Age Delta and digital biomarker Age Selta to assess the risk of degenerative aging lesions in different anatomical regions of the lumbar spine.

[0035] Furthermore, in step (3), the lumbar vertebral segmentation image dataset is grouped and cropped according to lumbar vertebral segments to obtain an image dataset of lumbar vertebral anatomy, which is implemented as follows:

[0036] (3a) The lumbar spine segmentation image dataset is divided into 5 groups according to the lumbar spine anatomical region, namely, the first lumbar vertebra L1 to the third lumbar vertebra L3, the third lumbar vertebra L3 to the fourth lumbar vertebra L4, the fourth lumbar vertebra L4 to the fifth lumbar vertebra L5, the first lumbar vertebra L1 to the fifth lumbar vertebra L5, and the first lumbar vertebra L1 to the first sacral vertebra S1;

[0037] (3b) Keeping the resolution, sequence parameters and anatomical information of the lumbar MRI images unchanged, each group of lumbar regions is cropped to form an image dataset of lumbar anatomical structures.

[0038] Furthermore, the assessment of the overall lumbar degenerative trend in the population in (4) includes:

[0039] (4a) The signal intensity of the intervertebral discs and the high signal intensity of the spinal cord were normalized for different genders to obtain the intervertebral disc signal score for each lumbar anatomical structure of each individual. :

[0040] ;

[0041] in, Represents the average signal intensity of the intervertebral disc. The average signal intensity representing the spinal cord, The intervertebral disc index represents the index that is indexed downwards from the intervertebral disc between the eleventh thoracic vertebra T11 and the twelfth thoracic vertebra T12.

[0042] (4b) Based on the intervertebral disc signal scores obtained by different genders, the average values ​​of different genders in each age group are calculated at 5-year intervals. These values ​​are then used to generate aging curves to describe the imaging characteristics of intervertebral disc changes over time and to assess the overall lumbar degeneration trend of the population.

[0043] Furthermore, in (8), the digital biomarker Age Delta and the digital biomarker Age Selta are defined to assess the risk of degenerative aging lesions in different anatomical regions of the lumbar spine. This is achieved through:

[0044] (8a) Define Age Delta as the difference between the predicted age and the actual biological age of a specific segment. :

[0045] (8b) Define Age Selta as the difference between the predicted age of the three variable intervertebral discs (L3-L4, L4-L5, and L5-S1) and the predicted age of the normal aging region (L1-L3). :

[0046] (8c) Based on the Age Delta, an age estimation map of the lumbar spine is drawn to determine whether there are abnormalities in specific segments from a macroscopic perspective; based on the Age Selta, a risk map of lumbar spine aging lesions is drawn to assess the degenerative changes of vulnerable areas relative to the baseline; based on the evaluation results of these two maps, the risk results of aging lesions in different anatomical regions of the lumbar spine are finally obtained.

[0047] Compared with the prior art, the present invention has the following advantages:

[0048] First, it has enabled continuous quantitative and individualized early warning of lumbar spine degeneration.

[0049] Existing methods are mostly limited to discrete grading or isolated measurements, lacking the ability to continuously assess age-related factors. This invention, through an age estimation model and digital biomarkers (Age Delta and Age Selta), enables continuous quantitative analysis of the degree of lumbar spine degeneration and achieves individualized risk stratification and early warning based on the offset of key segments, which is more in line with the needs of precision medicine.

[0050] Second, it significantly improves the interpretability of the evaluation results.

[0051] To address the "black box" problem of existing deep learning models, this invention combines slice-level attention mechanisms with saliency heatmaps to intuitively display the anatomical regions and degeneration patterns that the model focuses on in its decision-making. This helps doctors understand the sources and trends of risk, enhancing clinical trust and practical value. Attached Figure Description

[0052] Figure 1 This is a schematic diagram of the lumbar spine age estimation network structure for slice comparison learning according to the present invention;

[0053] Figure 2 This is a flowchart illustrating the implementation of the lumbar aging assessment method based on a lumbar spine age estimation network using slice comparison learning, as described in this invention.

[0054] Figure 3 This is a significance heatmap of the lumbar spine age estimation for different age groups by gender according to the present invention;

[0055] Figure 4 This is a scatter plot of age estimation for different anatomical regions of the lumbar spine based on the present invention;

[0056] Figure 5 This invention provides a risk map of lumbar spine aging lesions for different anatomical regions of the lumbar spine. Detailed Implementation

[0057] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention and not all of them. Based on the embodiments of the present invention, other embodiments obtained by those skilled in the art without creative effort should all fall within the protection scope of the present invention.

[0058] It should be noted that the step numbers in the specification and claims of this invention are only for the purpose of clearly describing the embodiments of this invention and facilitating understanding, and their order is not limited.

[0059] Example 1: Lumbar spine age estimation network based on slice comparison learning.

[0060] Reference Figure 1 This example includes: an encoder, an aggregation module, a fully connected regression layer, and an adjacent slice contrastive learning module. The structure and function of each module are as follows:

[0061] The encoder includes seven convolutional layers connected in sequence, wherein:

[0062] The first five convolutional modules have the same structure. Each module includes a 3×3 two-dimensional convolutional layer, an instance normalization layer, a max pooling layer, and a ReLU activation layer. These layers are connected in sequence to extract and compress multi-scale local features of the slices step by step.

[0063] The sixth convolutional module includes a 1×1 two-dimensional convolutional layer, an instance normalization layer, and a ReLU activation layer, which are connected in sequence to fuse and integrate the semantic information of deep features.

[0064] The seventh convolutional module includes an average pooling layer, a culling layer, and a 1×1 two-dimensional convolutional layer, which are connected in sequence to aggregate global information and generate the final slice feature encoding.

[0065] The encoder processes two-dimensional slices of each lumbar vertebral anatomical structure, extracts feature information, and obtains two sets of encoding sequences, as specifically implemented below:

[0066] The adjacent lumbar spine slice sequences were divided into two branches: the first, third, fifth, and seventh slices of the adjacent lumbar spine were the first branch, and the second, fourth, sixth, and eighth slices of the adjacent lumbar spine were the second branch.

[0067] Feature extraction is performed on the first branch of the adjacent slice sequence of the lumbar spine to obtain the output h1 after encoding the first branch;

[0068] Feature extraction is performed on the second branch of the adjacent slice sequence of the lumbar spine to obtain the output h2 after encoding the second branch;

[0069] The aggregation module includes a query Q, a key K, and a value V, wherein:

[0070] The query Q is a vector used in the attention mechanism to calculate similarity, representing the semantic information of a specific position in the input sequence; the key K and the value V are vectors used to represent the semantic information of different positions in the input sequence, the key is used to calculate similarity, and the value V is used to generate the output; the attention mechanism determines the weight of each position in the output sequence by calculating the similarity between Q and K.

[0071] The aggregation module, which is used to determine the attention weight at each position in the output sequence, is implemented as follows:

[0072] Let the total number of slices be , Indicates the first The encoding of each slice;

[0073] set up For trainable query vectors, Converted to bonds via linear mapping Sum All key vectors form a matrix ;

[0074] Calculate the attention weights between the query vector and the key matrix. :

[0075] ;

[0076] The final output is a linear combination of weighted vectors. It is used for lumbar spine age estimation.

[0077] The fully connected regression layer module includes a weight matrix and a bias vector, which are used to derive the predicted age.

[0078] The adjacent slice comparison learning module includes a projector and a predictor, wherein:

[0079] A projector, comprising at least one fully connected layer, is used to map high-dimensional features extracted by a similarity loss encoder to a low-dimensional projection space that is more suitable for similarity comparison.

[0080] A predictor, comprising at least one fully connected layer, for predicting the projection features of another adjacent slice from the low-dimensional projection features of one slice.

[0081] The adjacent slice comparison learning module calculates the adjacent slice comparison loss based on the outputs of the projector and the predictor. This loss is then used as the regularization term and the mean squared error loss of the main regression. The weighted summation constitutes the total loss function of the network. Its implementation is as follows:

[0082] The output after encoding the first branch Input to the projector to generate the first implicit label. The label The predictor generates the first output. ;

[0083] Then encode the output of the second branch. Input to the projector to generate a second implicit label. and calculate the labels With the first branch output Negative first cosine similarity between , :

[0084] , in Let L2 be the norm of the vector;

[0085] Then label The input is fed into the predictor to obtain the second output. And calculate its relationship with the first implicit tag. Negative value of second cosine similarity :

[0086] ;

[0087] Then calculate the negative value of the first cosine similarity. , Negative value of second cosine similarity , The average value L:

[0088] L , ,

[0089] Then, negative values ​​of the first cosine similarity are calculated separately. , Negative value of second cosine similarity , Perform a gradient descent stop operation and calculate the contrast loss between adjacent slices. :

[0090] ,

[0091] in To stop the gradient operation;

[0092] Then compare the loss between adjacent slices. As a regularization term and the mean squared error loss of the main regression Weighted summation constitutes the total loss function of the lumbar spine age estimation network. :

[0093] ;

[0094] in The loss is due to the mean square error (MSE). It is an adjustable hyperparameter.

[0095] The encoder, aggregation module, fully connected regression layer module, and adjacent slice comparison learning module are cascaded in sequence to form a lumbar spine age estimation network.

[0096] Example 2: A method for assessing lumbar spine aging using the lumbar spine age estimation network learned through slice comparison learning described above.

[0097] Reference Figure 2 The implementation steps of this example include the following:

[0098] Step 1. Obtain the dataset of lumbar spine MRI images.

[0099] 1.1) A dataset of thousands of T2-weighted sagittal MRI scans of the lumbar spine was obtained from the radiology department of a tertiary-level Class A hospital. This example obtained, but is not limited to, 1597 cases.

[0100] 1.2) Windowing is performed on the images in the lumbar MRI image dataset. That is, based on the signal intensity characteristics of the lumbar anatomical structure, the window width and window level parameters are set to linearly map the original image intensity to the standard grayscale range.

[0101] 1.3) The windowed MRI images were divided into training and test sets at a ratio of 1:9.

[0102] Step 2. Train the existing nnFormer segmentation network using the training set and test it using the test set to obtain the segmentation results of the lumbar anatomical structure, thus forming a lumbar vertebra segmentation image dataset.

[0103] 2.1) Perform random rotation, random scaling, and random flipping data augmentation operations on the lumbar MRI images in the training set in sequence;

[0104] 2.2) Input the enhanced data into the existing nnFormer segmentation network;

[0105] 2.3) Optimize the existing nnFormer segmentation network through multi-round iterative training:

[0106] 2.3.1) The learning rate is set to be dynamically adjusted based on the warm-up mechanism, the momentum is 0.9, the batch size is 2, the weight decay is 0.01, and the combination of Dice loss and cross-entropy loss is used as the segmentation loss function.

[0107] 2.3.2) In each training iteration, samples are randomly selected from the training set, and the segmentation loss of the nnFormer segmentation network is calculated;

[0108] 2.3.3) Based on the segmentation loss, backpropagation is performed using the AdamW optimizer to update the model parameters of the nnFormer segmentation network;

[0109] 2.3.4) Repeat steps 2.3.2)-2.3.3) until the loss function converges and the preset 1000 iterations are reached to obtain the trained nnFormer segmentation model, which outputs the segmentation results of the lumbar anatomical structure, namely the thoracic vertebrae, lumbar vertebrae, sacral vertebrae, seven intervertebral discs, and part of the spinal cord.

[0110] 2.4) Using the lumbar MRI images from the test set in step 1.3) as input, the trained nnFormer segmentation model is used for inference to output the corresponding lumbar anatomical structure segmentation results, namely the thoracic vertebrae, lumbar vertebrae, sacral vertebrae, seven intervertebral discs, and part of the spinal cord.

[0111] 2.5) Use the lumbar anatomical structure segmentation results obtained in steps 2.3.4) and 2.4) to construct a lumbar vertebra segmentation image dataset.

[0112] Step 3. Obtain an image dataset of lumbar anatomical structures.

[0113] 3.1) The lumbar spine segmentation image dataset was divided into 5 groups according to the lumbar spine anatomical regions: the first group is from the first lumbar vertebra L1 to the third lumbar vertebra L3; the second group is from the third lumbar vertebra L3 to the fourth lumbar vertebra L4; the third group is from the fourth lumbar vertebra L4 to the fifth lumbar vertebra L5; the fourth group is from the first lumbar vertebra L1 to the fifth lumbar vertebra L5; and the fifth group is from the first lumbar vertebra L1 to the first sacral vertebra S1.

[0114] 3.2) In the sagittal, coronal and axial three-dimensional views, locate the minimum bounding box containing the required complete lumbar vertebral body region based on the lumbar anatomical structure segmentation results of step 2.5); the upper boundary of this bounding box includes the upper edge of the first vertebral body in the group, the lower boundary includes the lower edge of the last vertebral body in the group, and the left and right lateral boundaries and the anterior and posterior boundaries completely include the vertebral body, intervertebral disc, spinal canal and surrounding related soft tissues;

[0115] 3.3) Without changing the voxel intensity value distribution of the original image and retaining all key sequence parameters in the DICOM header file, each lumbar region is cropped according to the bounding box to form an image dataset of lumbar anatomical structures.

[0116] Step 4. Perform data analysis on the image dataset of lumbar anatomical structures to assess the overall trend of lumbar degeneration in the population.

[0117] 4.1) The image dataset of lumbar anatomical structures obtained in 3.2) was further grouped by gender. In this embodiment, there were 557 female samples and 1040 male samples.

[0118] 4.2) The signal intensity of the intervertebral discs and the high signal intensity of the spinal cord were normalized for different genders to obtain the intervertebral disc signal score for each lumbar anatomical structure of each individual. :

[0119] ,

[0120] in, Represents the average signal intensity of the intervertebral disc. The average signal intensity representing the spinal cord, The intervertebral disc index represents the index that is indexed downwards from the intervertebral disc between the eleventh thoracic vertebra T11 and the twelfth thoracic vertebra T12.

[0121] 4.3) After obtaining the intervertebral disc signal scores for different gender groups, the population average for each age group was calculated at 5-year intervals;

[0122] 4.4) Based on the calculated average values ​​of different age groups, aging curves are generated, plotted separately for gender, and analyzed. The horizontal axis represents age, from 10 to 80 years old, calculated every 5 years. The vertical axis represents the ratio of intervertebral disc signal to spinal cord signal, describing the imaging characteristics of intervertebral disc changes over time, and assessing the overall lumbar spine degeneration trend of the population.

[0123] Step 5. Obtain the dataset for the lumbar spine age estimation network using slice comparison learning.

[0124] 5.1) For the image dataset of lumbar anatomical structures obtained in step 3.2), perform automated degeneration assessment on large-scale lumbar MRI data based on the Pfirrman grading criteria;

[0125] 5.2) By statistically analyzing the degeneration distribution, the segments exhibiting the lowest degree of intervertebral disc degeneration were identified, and the L1-L3 segments of the lumbar spine were selected accordingly as the dataset for the lumbar spine age estimation network for slice comparison learning and as baseline data reflecting physiological age characteristics.

[0126] 5.3) The dataset of the lumbar spine age estimation network based on slice comparison learning is divided into a training set for lumbar spine age estimation and a test set for lumbar spine age estimation in a 7:3 ratio.

[0127] Step 6. Train the lumbar spine age estimation network based on slice comparison learning, and test the obtained lumbar spine age estimation model based on slice comparison learning.

[0128] 7.1) The learning rate is dynamically adjusted based on the warm-up mechanism, the momentum is 0.9, the batch size is 24, the weight decay is 0.0005, and the number of attention maps is 128. The training set of lumbar spine age estimation obtained in 5.3) is input into the lumbar spine age estimation network.

[0129] 7.2) In each training iteration, samples are randomly selected from the training set, and the total loss of the lumbar spine age estimation network is calculated. ;

[0130] 7.3) Estimating the total network loss based on the lumbar spine age. Backpropagation is performed using the Adam optimizer to update the model parameters of the lumbar spine age estimation network;

[0131] 7.4) Repeat steps 7.2)-7.3) until the set 200 iterations of training are reached, and the trained slice comparison learning lumbar spine age estimation model is obtained.

[0132] 7.5) Input the test set data of the lumbar spine age estimation obtained in step 5.3) into the lumbar spine age estimation model of slice comparison learning for testing, obtain the predicted age results, and output the significance heatmap of lumbar spine age estimation.

[0133] 7.6) In the lumbar spine age estimation significance heatmap output in step 7.5), lumbar spine age estimation significance heatmaps for different age groups are selected and displayed separately for each gender, such as... Figure 3 As shown.

[0134] Step 8. Define digital biomarkers Age Delta and Age Selta to assess the risk of degenerative changes in different anatomical regions.

[0135] 8.1) Based on comprehensive anatomical insights, the L1-L3 region of the lumbar spine is defined as the benchmark for normal aging, and early warning of degenerative lesions is provided for segments such as L3-L4, L4-L5 and L5-S1 that are more prone to degenerative lesions.

[0136] 8.2) Two new digital biomarkers, Age Delta and Age Selta, are defined to assess the risk of degenerative changes in different anatomical regions:

[0137] The Age Delta refers to the difference between the predicted age and the actual biological age of a specific segment. :

[0138] ,

[0139] in Predicting age for specific lumbar vertebral segments. The subject's actual physiological age;

[0140] The Age Selta is a more precise measurement method, defined as the difference between the predicted age of the three degenerate intervertebral discs L3-L4, L4-L5, and L5-S1 and the predicted age of the normal aging region L1-L3. :

[0141]

[0142] in, Intervertebral discs, which are prone to degeneration, predict age. Predicted age of normally aging regions.

[0143] 8.3) Assess the risk of degenerative changes in different anatomical regions based on the respective parameter standards of Age Delta and Age Selta:

[0144] When using the Age Delta increment to assess the risk of degenerative changes in different anatomical regions, if an abnormality is detected, namely the difference between the predicted age and the actual biological age of the lumbar spine... A value greater than 3 indicates a significant risk in that segment, and the area cannot be used as a baseline for degeneration prediction. In this embodiment, most tests in the L1-L3 segment were normal, meaning the difference between the predicted lumbar spine age and the actual biological age was [value missing]. If the value is less than 3, then this region is used as the baseline for degradation prediction.

[0145] When using the Age Selta to assess the risk of degenerative changes in different anatomical regions, if the absolute value of Age Selta is within 3 years, the segment is considered to be aging normally with no significant risk. However, if the value exceeds 3 years, the corresponding intervertebral disc is considered to have a potential risk of degeneration and requires further attention and monitoring.

[0146] These digital biomarkers provide insights into aging processes and degradation risks, offering valuable tools for early diagnosis and targeted intervention.

[0147] 8.4) Visual assessment of lumbar spine age difference and aging risk analysis:

[0148] Based on the Age Delta calculated in step 8.2), a lumbar spine age estimation map is drawn to macroscopically determine whether there are abnormalities in a specific segment;

[0149] Based on the Age Selta calculated in step 8.3), a risk map of lumbar spine aging lesions is plotted to centrally assess the degenerative changes of vulnerable areas relative to the baseline. The horizontal axis represents the study participants, and the vertical axis represents the predicted age. The predicted age of the L1-L3 region of the lumbar spine is used as the baseline for normal aging. Shifting upward by 3 years represents the risk curve. Points above the risk curve indicate a higher risk of degenerative aging.

[0150] The effects of this invention can be further illustrated by the following simulation experiments:

[0151] I. Simulation conditions:

[0152] The simulation test platform of this invention is a PC with an Intel Core i7-9700K CPU 3.6GHz, 128GB of memory, and an Nvidia RTX4090 graphics card. The simulation platform is based on the Ubuntu 18.04 operating system, uses the PyTorch deep learning framework, and is implemented in Python.

[0153] Data Source: The data for this experiment comes from two orthopedic hospitals. The dataset consists of T2-weighted sagittal MRI scans of the lumbar spine from 1597 adults (11-80 years old), with pixel spacing between 4 and 5 mm and slice thickness between 4.4 and 6.25 mm. Sample size and age distribution show that there are fewer samples under 20 years old and over 70 years old, while there are more samples from young and middle-aged groups, indicating an imbalance in the distribution.

[0154] Pearson correlation coefficient (PCC) and Spearman rank correlation coefficient (SRCC) were used as assessment indicators to measure the correlation between estimated age and actual age.

[0155] Mean absolute error (MAE) and root mean square error (RMSE) are used as evaluation indicators to measure the error between estimated age and actual age. Their calculation formulas are as follows:

[0156] ,

[0157] ;

[0158] Cumulative score (CS) and mean cumulative score (MCS) are used as evaluation metrics to measure accuracy within the tolerable error range. Their calculation formulas are as follows:

[0159]

[0160] ;

[0161] Where N is the total number of samples in the dataset. It is the actual age of the i-th sample. It is the model's estimated age for the i-th sample. It is a tolerable error.

[0162] The closer the MAE and RMSE are, and the higher the CS and MCS values ​​are, the more accurate the age estimation will be.

[0163] The Pearson correlation coefficient (PCC) and Spearman rank correlation coefficient (SRCC) are close to 1, indicating a strong correlation between the estimated age and the actual age, reflecting better model performance.

[0164] II. Simulation Content and Result Analysis:

[0165] Simulation 1: This invention plots a scatter plot of lumbar spine age estimation based on the Age Delta calculated from different lumbar anatomical regions. The results are as follows: Figure 4 ,in,

[0166] Figure 4 (a) is a scatter plot of lumbar spine age estimation drawn in the L1-L3 segment of the lumbar spine according to the present invention;

[0167] Figure 4 (b) is a scatter plot of lumbar spine age estimation drawn in the L1-L5 segment of the lumbar spine according to the present invention;

[0168] Figure 4 (c) is a scatter plot of lumbar spine age estimation drawn in the L4-L5 segment of the lumbar spine according to the present invention;

[0169] Figure 4 (d) is a scatter plot of lumbar spine age estimation drawn in the L1-S1 segment of the lumbar spine according to the present invention;

[0170] from Figure 4 As can be seen, this invention, by investigating age estimation associated with lumbar degenerative changes, can assess the aging status of different anatomical regions L1-L3, L1-L5, L4-L5, and L1-S1. It can also display scatter plots of different anatomical regions, visualize age differences, and show MAE values, the coefficient of determination R-squared, and the Spearman correlation coefficient. Among these:

[0171] Figure 4 (a) Pearson correlation coefficient (PCC) and coefficient of determination (R) 2 The values ​​are all close to 1, indicating a strong correlation between the estimated age and the actual age, and the model's prediction effect is good.

[0172] contrast Figure 4 (a)- Figure 4 (d) This invention provides the most accurate and stable age prediction in the L1-L3 lumbar region, with the age difference mainly within 5 years. In contrast, the predicted age difference in the L1-L5 and L1-S1 regions is a wider range of 7-10 years, while the performance in the L4-L5 region is the worst, with the predicted age difference reaching 15 years.

[0173] The age estimation evaluation indicators for different lumbar anatomical regions are shown in Table 1.

[0174] Table 1. Results of age estimation evaluation indicators for different lumbar anatomical regions.

[0175]

[0176] Table 1 shows the Pearson correlation coefficient (PCC) and Spearman rank correlation coefficient (SRCC), reflecting the correlation between estimated age and actual age. Mean absolute error (MAE) and root mean square error (RMSE) are two evaluation metrics that measure the error between estimated and actual age. MAE(11,20) represents the mean absolute error for the 11-20 age group. Cumulative score (CS) and mean cumulative score (MCS) evaluate the cumulative score of the model within the tolerable error range. These two metrics range from 0 to 100, with higher values ​​indicating better age estimation performance.

[0177] As shown in Table 1, quantitative assessment indicators (PCC, SRCC, RMSE, MAE) indicate that the L1-L3 region of the lumbar spine outperformed other regions, achieving PCC 0.97, SRCC 0.97, RMSE 3.341, and MAE 2.592. Analysis of age estimation accuracy across different age groups showed that the L1-L3 region consistently provided the closest estimates in individuals aged 30-70 years. However, due to less degenerative change, the L4-L5 region performed better in younger individuals aged 11-30 years. The better performance of the L4-L5 region in younger individuals (11-30 years) is attributed to less degeneration.

[0178] Cumulative score (CS) and mean cumulative score (MCS) results further showed that L1-L3 was the best performing region, with a CS of 87% over a 5-year period and 99% over a 10-year period. These results suggest that future studies should prioritize the L1-L3 region to obtain more reliable lumbar spine age estimates.

[0179] Simulation 2: This invention generates a risk map of lumbar spine aging-related diseases based on the Age Selta calculated from different lumbar anatomical regions. The results are as follows: Figure 5 ,in,

[0180] Figure 5 (a) A risk map of lumbar spine age-related lesions drawn in the L3-L4 region;

[0181] Figure 5 (b) A risk map of lumbar spine age-related lesions plotted in the L4-L5 region;

[0182] Figure 5 (c) A risk map of lumbar aging lesions drawn in the L5-S region of the lumbar spine;

[0183] from Figure 5As shown in (a), the blue dots represent the predicted age of the L1-L3 region, while the purple dots represent the predicted age of the L3-L4 region. The blue line serves as a reference line. When the age increment (Age Selta) equals 3 years, a red line outlines the risk curve, defining the area above it as the risk zone. Red dots falling within this risk zone indicate a warning of degenerative risk in the L3-L4 intervertebral disc. Figure 5 (b) and (c) similarly show the bias in predicted age between the L4-L5 and L5-S1 regions and the L1-L3 region. The predicted values ​​marked by red dots falling within the red risk zone indicate that these samples have a higher risk of degenerative aging.

[0184] Figure 5 (a) to Figure 5 (c) The increased density of perceived risk points indicates that the degeneration rate of L4-L5 and L5-S1 intervertebral discs is higher than that of L3-L4 intervertebral discs. Figure 5 The yellow circle in the image indicates that using the digital biomarker Age Selta value to assess the risk of degenerative changes can effectively identify and predict high-risk groups, thereby promoting early intervention and prevention of degenerative diseases.

Claims

1. A lumbar spine age estimation network based on slice comparison learning, characterized in that, include: An encoder consisting of seven layers of different convolutional modules was built to process two-dimensional slices of each lumbar vertebral anatomical structure, extract feature information, and obtain the output after encoding two branches. An aggregation module is built, which includes vector query Q, key K and value V, to determine the attention weight at each position in the output sequence. The weighted sum is then used for age estimation. A fully connected regression layer is constructed, which includes a weight matrix and a bias vector, to derive the predicted age. A neighboring slice contrast learning module is established, comprising a cascaded projector and predictor, which calculates the neighboring slice contrast loss based on the outputs of the projector and predictor. This loss is then used as the regularization term and the mean squared error loss of the main regression. Weighted summation constitutes the total loss function of the network. ; The encoding module, aggregation module, fully connected regression layer, and adjacent slice comparison learning module are cascaded in sequence to form a slice comparison learning lumbar spine age estimation network.

2. The network according to claim 1, characterized in that, The seven convolutional layers in the encoder, along with their interconnections, are as follows: The first five convolutional modules have the same structure. Each module includes a 3×3 two-dimensional convolutional layer, an instance normalization layer, a max pooling layer, and a ReLU activation layer, which are connected in sequence to extract and compress multi-scale local features of the slices step by step. The sixth convolutional module includes a 1×1 two-dimensional convolutional layer, an instance normalization layer, and a ReLU activation layer, which are connected in sequence to fuse and integrate the semantic information of deep features. The seventh convolutional module includes an average pooling layer, a culling layer, and a 1×1 two-dimensional convolutional layer, which are connected in sequence to aggregate global information and generate the final slice feature encoding. The seven convolutional modules are cascaded in sequence to form an encoder.

3. The network according to claim 1, characterized in that, The aggregation module determines the attention weight at each position in the output sequence, and its implementation is as follows: Let the total number of slices be , For the first The encoding of each slice; set up For trainable query vectors, Converted to bonds via linear mapping Sum All key vectors form a matrix ; The attention weights between the query vector and the key matrix are calculated using the Softmax function. : ， in, For trainable query vectors With all keys dot product, This is a transpose.

4. The network according to claim 1, characterized in that, The structure and transmission relationship of the projector and predictor in the adjacent slice comparison learning module are as follows: The projector includes at least one fully connected layer for mapping high-dimensional features extracted by the encoder to a low-dimensional projection space that is more suitable for similarity comparison. The predictor includes at least one fully connected layer for predicting the projection features of another adjacent slice from the low-dimensional projection features of one slice, and the output is used to calculate feature similarity to construct an auxiliary loss.

5. The network according to claim 1, characterized in that, The adjacent slice contrast loss is calculated based on the outputs of the projector and the predictor. This loss is then used as the regularization term and the mean squared error loss of the main regression. Weighted summation constitutes the total loss function of the network. The formula is as follows: ， ， in, , For the two outputs of the projector slice pair, , For the two outputs of the predictor slice pair, The negative cosine similarity value is used for gradient stopping operation. The loss is due to the mean square error (MSE). It is an adjustable hyperparameter.

6. A method for assessing lumbar spine aging using a lumbar spine age estimation network based on slice comparison learning as described in claim 1, characterized in that, include: (1) Obtain the dataset of lumbar spine MRI images and perform windowing processing, dividing it into training and test sets at a ratio of 1:9: (2) The existing nnFormer segmentation network is trained using the training set and tested using the test set to obtain the segmentation results of the lumbar anatomical structure, thus forming a lumbar vertebra segmentation image dataset. (3) The lumbar spine segmentation image dataset is grouped and cropped according to lumbar spine segments to obtain an image dataset of lumbar spine anatomy; (4) Group the image dataset of lumbar anatomical structures by gender and assess the overall trend of lumbar degeneration in the population; (5) The images of the first to third lumbar vertebrae in the lumbar anatomical structure image dataset are divided into a training set and a test set for lumbar age estimation in a 7:3 ratio. (6) Input the training set of lumbar spine age estimation into the lumbar spine age estimation network of slice comparison learning, and use the Adam optimizer to iteratively update the model parameters in the manner of random single sample learning each time until the preset number of times, so as to obtain the trained lumbar spine age estimation model of slice comparison learning. (7) Input the test set of lumbar spine age estimation into the lumbar spine age estimation model that has been trained by slice comparison learning, and output the predicted age results and the significance heatmap of lumbar spine age estimation. (8) Based on the predicted age results, define the digital biomarker Age Delta and the digital biomarker Age Selta to assess the risk of degenerative aging lesions in different anatomical regions of the lumbar spine.

7. The method according to claim 6, characterized in that, The step (2) of training the existing nnFormer segmentation network using the training set includes: (2a) Perform data augmentation operations on the lumbar MRI images in the training set, including random rotation, random scaling and random flipping, and then input the augmented data into the existing nnFormer segmentation network. (2b) The nnFormer segmentation network is trained through multiple rounds of iteration until the predetermined number of training rounds or the loss function converges, and the trained nnFormer segmentation model is obtained, which outputs the segmentation results of the lumbar anatomical structure, namely the thoracic vertebrae, lumbar vertebrae, sacral vertebrae, seven intervertebral discs and part of the spinal cord.

8. The method according to claim 6, characterized in that, The process of segmenting the lumbar spine image dataset into lumbar spine segments to obtain an image dataset of lumbar spine anatomy is described in (3). (3a) The lumbar spine segmentation image dataset is divided into 5 groups according to the lumbar spine anatomical region, namely, the first lumbar vertebra L1 to the third lumbar vertebra L3, the third lumbar vertebra L3 to the fourth lumbar vertebra L4, the fourth lumbar vertebra L4 to the fifth lumbar vertebra L5, the first lumbar vertebra L1 to the fifth lumbar vertebra L5, and the first lumbar vertebra L1 to the first sacral vertebra S1; (3b) Keeping the resolution, sequence parameters and anatomical information of the lumbar MRI images unchanged, each group of lumbar regions is cropped to form an image dataset of lumbar anatomical structures.

9. The method according to claim 6, characterized in that, The assessment of lumbar degenerative trends in the overall population in (4) includes: (4a) The signal intensity of the intervertebral discs and the high signal intensity of the spinal cord were normalized for different genders to obtain the intervertebral disc signal score for each lumbar anatomical structure of each individual. : ； in, Represents the average signal intensity of the intervertebral disc. The average signal intensity representing the spinal cord, The intervertebral disc index represents the index that is indexed downwards from the intervertebral disc between the eleventh thoracic vertebra T11 and the twelfth thoracic vertebra T12. (4b) Based on the intervertebral disc signal scores obtained for different genders, the average values ​​for each age group for different genders are calculated at 5-year intervals. These values ​​are then used to generate aging curves to describe the imaging characteristics of intervertebral disc changes over time and to assess the overall lumbar degeneration trend of the population.

10. The method according to claim 6, characterized in that, The step (6) involves inputting the training set for lumbar spine age estimation into the slice-comparison learning lumbar spine age estimation network, and using the Adam optimizer to iteratively update the model parameters in a random single-sample learning manner. This process includes: (6a) Set optimizer parameters, including using a warm-up mechanism to dynamically adjust the learning rate, setting momentum to 0.9, batch size to 24, and weight decay to 0.0005; (6b) In each training iteration, samples are randomly selected from the training set, and the total loss of the lumbar age estimation network learned by the slice comparison is calculated. (6c) Based on the total loss of the lumbar spine age estimation network learned by the slice comparison, backpropagation is performed using the Adam optimizer to update the model parameters of the lumbar spine age estimation network learned by the slice comparison. (6d) Repeat steps (6b)-(6c) until the predetermined number of training rounds is reached to obtain the lumbar spine age estimation model based on slice comparison learning after training.

11. The method according to claim 6, characterized in that, In (8), the digital biomarker Age Delta and the digital biomarker Age Selta are defined to assess the risk of degenerative aging-related diseases in different anatomical regions of the lumbar spine. Their implementation includes: (8a) Define Age Delta as the difference between the predicted age and the actual biological age of a specific segment. : ， in, Predicting age for specific lumbar vertebral segments. The subject's actual physiological age; (8b) Define Age Selta as the difference between the predicted age of the three variable intervertebral discs (L3-L4, L4-L5, and L5-S1) and the predicted age of the normal aging region (L1-L3). : ， in, Intervertebral discs, which are prone to degeneration, predict age. Predicted age for the normal aging baseline region; (8c) Based on the Age Delta, an age estimation map of the lumbar spine is drawn to determine whether there are abnormalities in specific segments from a macroscopic perspective; based on the Age Selta, a risk map of lumbar spine aging lesions is drawn to assess the degenerative changes of vulnerable areas relative to the baseline; based on the evaluation results of these two maps, the risk results of aging lesions in different anatomical regions of the lumbar spine are finally obtained.

Images