Vertebral bone quality multi-modal fusion prediction method and system

Abstract

According to the vertebral body bone quality multi-modal fusion prediction method and system, firstly, biological characteristic data (gender, age, height, weight and BMI) and vertebral body image data (T11-L2 vertebral body three-dimensional bone mineral density parameters, T1, T2, T2-STIR and T1-FLAIR sequence multi-modal magnetic resonance images) of a to-be-detected object are obtained; obtaining a bone mineral density grouping result according to the three-dimensional bone mineral density parameter and a preset bone mineral density threshold value, then setting ROI in a specific region of the multi-modal magnetic resonance image, extracting signal intensity to calculate different types of VBQ values, carrying out sequence optimization on the VBQ values, and finally, combining VBQ fat pressure optimization and biological characteristic data by using binary logistic regression to obtain a multi-modal magnetic resonance image. Screening grouping and diagnosis grouping prediction models are constructed, after the prediction models integrate multiple factors, the diagnosis grouping prediction model AUC is further improved to 0.934, and the screening grouping prediction model AUC reaches 0.917.

Description

Technical Field

[0001] The present invention relates to the technical field of vertebral bone mass prediction, and particularly to a multi-modal fusion prediction method and system for vertebral bone mass. Background Art

[0002] In the field of osteoporosis (OP) diagnosis technology, current mainstream methods face many difficulties. Dual - emission X - ray Absorptiometry (DXA), as the "gold standard" recommended by the World Health Organization, measures bone mineral density (BMD) by means of two - dimensional bone area projection, and cannot accurately show the true volume density of the vertebra. Once the patient has conditions such as vertebral degeneration, osteophyte formation or vascular calcification, the measurement error will increase significantly. Since degenerative lesions of the lower lumbar spine are extremely common, the BMD measurement values in the area near the lower lumbar spine are prone to being too large, making it difficult to truly reflect the vertebral bone strength. Moreover, although the single - time radiation exposure of DXA is lower than that of ordinary X - ray examinations, there are still potential risks for patients undergoing long - term follow - up reexaminations. Coupled with the high cost of the equipment, it seriously hinders its wide application in large - scale physical examination screenings.

[0003] The Hounsfield unit (HU) method evaluates local BMD by means of CT scanning, which to a certain extent avoids the interference of vertebral degeneration, and has a good correlation with the results of traditional DXA and the T - Score (T). It can be used as a supplementary means for bone mass evaluation. However, this method is significantly affected by scanning parameters (especially tube voltage), resulting in poor repeatability in experiments and clinics, and at the same time, it cannot get rid of the problem of X - ray radiation exposure.

[0004] Quantitative Computed Tomography (QCT), as a three - dimensional bone density measurement technology, can make up for some deficiencies of DXA and provide more accurate BMD information. However, its radiation dose is 5 - 30 times that of DXA, and it also requires regular calibration with a special body membrane, and the equipment penetration rate is relatively low. These factors limit the large - scale application of QCT in the OP screening of the physical examination population.

[0005] In recent years, the Vertebral Bone Quality Score Based on MRI (VBQ) technology has emerged, opening up a new direction for non-invasive and non-destructive bone quality evaluation. By measuring and analyzing the bone marrow signal intensity of the vertebral body (the change in T1 signal intensity is partly due to fat infiltration and bone atrophy transformation), VBQ provides a metric for BMD from a new perspective, with advantages such as no radiation and strong repeatability. However, traditional VBQ methods have problems such as a single sequence and lack of standardization. The research and elaboration on the optimization of VBQ for multiple sequences and the improvement of its diagnostic efficiency among them are not deep enough. Therefore, developing a new solution that can overcome the drawbacks of existing OP diagnostic technologies and optimize VBQ technology to improve diagnostic efficiency has become an urgent problem to be solved currently. Summary of the Invention

[0006] Problems to be solved by the present invention: Provide a multi-modal fusion prediction method for vertebral bone quality to solve the problems of inaccurate measurement, high radiation, high cost, poor repeatability, and difficult equipment popularization of DXA, HU method, and QCT. At the same time, optimize VBQ to solve the problems of single sequence and lack of standardization. The first aspect of the present invention provides a multi-modal fusion prediction method for vertebral bone quality, which includes the following steps: Step 1: Obtain the biometric data and vertebral body image data of the object to be measured; the basic data includes gender, age, height, weight, and BMI; the vertebral body image data includes three-dimensional bone density parameters of T11-L2 vertebral bodies and multi-modal magnetic resonance images, and the multi-modal magnetic resonance images cover at least T1, T2, T2-STIR, and T1-FLAIR sequences; Step 2: Based on the three-dimensional bone density parameters, generate bone density grouping results for screening grouping and diagnosis grouping through the bone density thresholds of the preset osteoporosis group, non-osteoporosis group, normal bone mass group, and non-normal bone mass group; Step 3: Set regions of interest (ROI) in the trabecular bone region of T11-L2 vertebral bodies and the cerebrospinal fluid region at the L3 level in the multi-modal magnetic resonance images, and extract the vertebral body signal intensity and cerebrospinal fluid signal intensity of each sequence; Calculate the VBQ value of a single vertebral body, the average VBQ of L1-L2 vertebral bodies, and the average VBQ of the whole T11-L2 segment based on the signal intensity; Perform sequence optimization processing on the VBQ value to generate VBQ 压脂优化 、VBQ 压水优化 and VBQ 共同优化 parameters; Step 4: Using binary logistic regression to combine the VBQ fat suppression optimization, gender, age, height, weight, and BMI to construct a screening group prediction model and a diagnostic group prediction model, and obtain the probability value and the prediction model.

[0007] Preferably, the three-dimensional bone density parameters include the independent vBMD values of each vertebra from T11 to L2, the average vBMD of the L1-L2 vertebrae vBMD L1-2 , and the average vBMD of the entire T11-L2 segment vBMD T11-L2 .

[0008] Preferably, the method for setting the bone density threshold is as follows: Osteoporosis group threshold: vBMD L1-2 < 80 mg / cm³; Non-osteoporosis group threshold: vBMD L1-2 ≥ 80 mg / cm³; Normal bone mass group threshold: vBMD T11-L2 >120 mg / cm³; Non-normal bone mass group threshold: vBMD T11-L2 ≤ 120 mg / cm³.

[0009] Preferably, the formula for the screening group prediction model is: 1 - P = 1 / [1 + exp(76.330 + 0.128 × age - 50.861 × height + 0.458 × weight - 1.250 × BMI + 0.194 × VBQ 压脂优化 + A)] where: A is the gender constant, for females: A = -0.592; for males: A = 0; The formula for the diagnostic group prediction model is: 1 - P = 1 / [1 + exp(50.651 + 0.114 × age - 39.833 × height + 0.442 × weight - 1.081 × BMI + 0.428 × VBQ fat suppression optimization + A)] where: A is the gender constant, if female, then A = 0.777, if male A = 0.

[0010] Preferably, the sequence optimization process for the VBQ value: Using the fat suppression algorithm to optimize the contrast between the vertebra and cerebrospinal fluid for the T2-STIR sequence, and calculating the VBQ 压脂优化 parameters; using the water suppression algorithm to optimize the background noise for the T1-FLAIR sequence, and calculating the VBQ 压水优化 parameters; performing weighted fusion on the VBQ values of the double suppression sequence to generate the VBQ 共同优化 parameters.

[0011] In a second aspect of the present invention, a multi-modal fusion prediction system for vertebral bone mass is proposed, which is based on the multi-modal fusion prediction method for vertebral bone mass and includes a data acquisition module, a bone density grouping module, a VBQ value processing module, and a prediction model construction module. The data acquisition module is configured to: acquire the biometric data and vertebral imaging data of the object to be measured; the basic data includes gender, age, height, weight, and BMI; the vertebral imaging data includes three-dimensional bone density parameters of the T11-L2 vertebrae and multi-modal magnetic resonance images, and the multi-modal magnetic resonance images cover at least the T1, T2, T2-STIR, and T1-FLAIR sequences. The bone density grouping module is configured to: based on the three-dimensional bone density parameters, generate the bone density grouping results for screening grouping and diagnosis grouping through the bone density thresholds of the preset osteoporosis group, non-osteoporosis group, normal bone mass group, and non-normal bone mass group. The VBQ value processing module is configured to: set regions of interest (ROIs) in the trabecular bone region of the T11-L2 vertebrae and the cerebrospinal fluid region at the L3 level in the multi-modal magnetic resonance images, and extract the vertebral signal intensity and cerebrospinal fluid signal intensity of each sequence. Calculate the VBQ value of a single vertebra, the average VBQ of the L1-L2 vertebrae, and the average VBQ of the entire T11-L2 segment based on the signal intensity. Perform sequence optimization processing on the VBQ value to generate VBQ fat-suppressed optimization, VBQ water-suppressed optimization, and VBQ common optimization parameters. The prediction model construction module is configured to: use binary logistic regression to combine the VBQ fat-suppressed optimization, gender, age, height, weight, and BMI to construct a screening grouping prediction model and a diagnosis grouping prediction model, and obtain the probability value and the prediction model.

[0012] Advantages of the present invention: Through the multi-sequence MRI optimization algorithm and the multi-factor prediction model, the present invention significantly improves the accuracy and clinical applicability of bone mass assessment under non-invasive conditions. Its core advantages are as follows: 1. Safety: Radiation-free, suitable for long-term monitoring of healthy people.

[0013] 2. High precision: The AUC of the fat-suppressed optimization sequence reaches 0.912, and the AUC is increased to 0.934 after model integration.

[0014] 3. Clinical practicability: It can flexibly handle body position deviations, simplify the operation process, and meet the dual needs of screening and diagnosis. Description of the Drawings

[0015] Figure 1 It is a measurement diagram of the cerebrospinal fluid signal intensity and the ROI signal intensity of the T11-L2 vertebrae of each sequence in the present invention.

[0016] Figure 2 This is the measurement diagram of vBMD of a single vertebra (T11-L2) of the present invention.

[0017] Figure 3 This is the diagram of the AUC value of the diagnostic grouping prediction model analyzed by the ROC curve of the present invention.

[0018] Figure 4 This is the diagram of the AUC value of the screening grouping prediction model analyzed by the ROC curve of the present invention.

[0019] Figure 5 This is the probability diagram of the diagnostic grouping prediction model analyzed by the ROC curve of the present invention.

[0020] Figure 6 This is the probability diagram of the screening grouping prediction model analyzed by the ROC curve of the present invention.

[0021] Figure 7 This is the flow chart of the steps of the present invention. Detailed implementation manners

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments.

[0023] The following is combined with the attached Figure 1 to Figure 6 , and a further detailed description of the specific implementation manners of the present invention will be given.

[0024] Embodiment 1, a multi-modal fusion prediction method for vertebral bone mass, refer to Figure 7 , and this method includes the following steps: Step 1, obtaining the biometric data and vertebral imaging data of the object to be measured; the basic data includes gender, age, height, weight, and BMI (BMI is the body mass index, and its calculation method is to divide the weight (in kilograms) by the square of the height (in meters)); the vertebral imaging data includes the three-dimensional bone density parameters of the T11-L2 vertebrae and multi-modal magnetic resonance images, and the multi-modal magnetic resonance images cover at least the T1, T2, T2-STIR, and T1-FLAIR sequences; In this embodiment, the three-dimensional bone density parameters include the independent vBMD values of each vertebra of T11-L2, the average vBMD of the L1-L2 vertebrae vBMD L1-2 , and the average vBMD of the entire T11-L2 segment vBMD T11-L2 ; Specifically, the scanning range covers the T11 - L2 vertebrae. A Toshiba Aquilion 64 - slice spiral CT scanner is used and equipped with an American Mindways4 QCT body phantom. Before scanning, all examinees are regularly calibrated using this QCT body phantom. The scanning parameters are set as follows: the field of view is 500mm, the slice thickness is 1mm, the matrix is 512×512, the current uses the automatic milliampere technology, the standard reconstruction algorithm is used, and the voltage is set to 120KV. After the scanning is completed, the original image data is transmitted to a quantitative computed tomography workstation. With the QCTPro 4.0v software in the QCT bone density analysis system of the American Mindways company, according to the operation guide, the three - dimensional average bone density of the cancellous bone of each single vertebra from T11 - L2 is measured, with the unit of mg / cm³. The specific measurement details are shown in Figure 2: ① Axial and mid - sagittal plane sections of the T11 vertebra; ② Axial and mid - sagittal plane sections of the T12 vertebra; ③ Axial and mid - sagittal plane sections of the L1 vertebra; ④ Axial and mid - sagittal plane sections of the L2 vertebra; On this basis, further calculate the average vBMD of the L1 and L2 vertebrae, denoted as vBMDL1 – 2. Add the vBMD values of the L1 and L2 vertebrae and divide the sum by 2 to obtain the average vBMD of the L1 and L2 vertebrae; and calculate the average vBMD of the T11 - L2 vertebrae, denoted as vBMDT11 - L2. Similarly, add the vBMD values of the four vertebrae T11, T12, L1, and L2 and divide by 4 to get the average vBMD of the T11 - L2 vertebrae; Step 2, based on the three - dimensional bone density parameters, generate the bone density grouping results for screening and diagnostic grouping through the preset bone density thresholds of the osteoporosis group, non - osteoporosis group, normal bone mass group, and non - normal bone mass group; In this embodiment, the method for setting the bone density threshold is as follows: Osteoporosis group threshold: vBMD L1-2 < 80 mg / cm³; Non - osteoporosis group threshold: vBMD L1-2 ≥ 80 mg / cm³; Normal bone mass group threshold: vBMD T11-L2 >120 mg / cm³; Non - normal bone mass group threshold: vBMD T11-L2 ≤ 120 mg / cm³; Bone density grouping results, diagnostic grouping: vBMD < 80 mg / cm3 is defined as the osteoporosis group, and vBMD ≥ 80 mg / cm3 is defined as the non-osteoporosis group; screening grouping: vBMD > 120 mg / cm3 is defined as the normal bone mass group, and vBMD ≤ 120 mg / cm3 is the non-normal bone mass group; Step 3, set regions of interest (ROIs) in the trabecular bone region of the T11-L2 vertebrae and the cerebrospinal fluid region at the L3 level in the multi-modal magnetic resonance images, and extract the vertebral signal intensity and cerebrospinal fluid signal intensity of each sequence; The method for setting the ROIs is as follows: The first ROI is an elliptical region covering the cancellous bone region of the vertebra and avoiding the cortical bone and vertebral venous plexus; the second ROI is within the cerebrospinal fluid space at the L3 level; Based on the signal intensity, calculate the VBQ value of a single vertebra, the average VBQ of the L1-L2 vertebrae, and the average VBQ of the entire T11-L2 segment; Perform sequence optimization processing on the VBQ value to generate VBQ 压脂优化 、VBQ 压水优化 and VBQ 共同优化 parameters; The sequence optimization processing of the VBQ value: Adopt a fat suppression algorithm to optimize the vertebra-cerebrospinal fluid contrast for the T2-STIR sequence, and calculate the VBQ 压脂优化 parameters; Adopt a water suppression algorithm to optimize the background noise for the T1-FLAIR sequence, and calculate the VBQ 压水优化 parameters; Perform weighted fusion on the VBQ values of the double suppression sequence to generate the VBQ 共同优化 parameters; Specifically, use the United Imaging 1.5T uMR588 magnetic resonance scanner to carry out the scanning work, and the specific scanning parameter settings are as follows: T1 sequence: slice thickness is 5 mm, slice gap is 0.5 mm, field of view is 320×240 mm; repetition time is set to 300 s, echo time is 8.52 s, acquisition matrix is 256×240, and the standard reconstruction algorithm is used; Fluid-attenuated inversion recovery (FLAIR) sequence: slice thickness, slice gap, and field of view are 5 mm, 0.5 mm, and 320×240 mm respectively; repetition time is 1892 s, echo time is 11.86 s, inversion time is 810 s, and the acquisition matrix is also 256×240, and the standard reconstruction algorithm is adopted; T2 sequence: slice thickness, slice gap, and field of view are 5 mm, 0.5 mm, and 320×240 mm respectively; repetition time is 3100 s, echo time is 100 s, acquisition matrix is 256×240, and the standard reconstruction algorithm is executed; Short Tau Inversion Recovery (STIR) sequence: slice thickness, slice gap, and field of view are 5 mm, 0.5 mm, and 320×240 mm respectively; repetition time is 20.48 s, echo time is 48.48 s, inversion time is 160 s, acquisition matrix is 240×216, and standard reconstruction algorithm is used; The scope of this MR scan covers the T11 - L2 vertebral bodies. Under the MRI non - contrast T1, T2, T2 - STIR, and T1 - FLAIR sequences, the mid - sagittal plane images of the lumbar spine of healthy subjects are obtained. By placing an elliptical region of interest (ROI) in the trabecular bone area of the T11 - L2 vertebral bodies (Vertebral Body, VB), the average signal intensity (Signal Intensity, SI) of the corresponding region can be automatically generated with the help of the post - processing software of United Imaging 1.5T MR. At the same time, an ROI is also placed in the cerebrospinal fluid (Cerebrospinal Fluid, CSF) circulation space behind the vertebral body at the L3 level to obtain the SI (CSF) of the cerebrospinal fluid. It should be noted that the ROI of the cerebrospinal fluid signal should be as small as possible and of consistent size. Subsequently, the VBQ is standardized according to the signal intensity of the cerebrospinal fluid in the spinal canal behind the L3 level. Since the cauda equina and filum terminale are below L1 and the cerebrospinal fluid circulation space is relatively large, and the L1 segment is partially affected by the conus medullaris signal, equal - area ROIs are selected in the cerebrospinal fluid below the L1 plane to effectively control variables. The measurement details are shown in Figure 1; Regarding the placement of the vertebral body ROI, the following requirements should be followed: Avoid the cortical bone and posterior vertebral venous plexus as much as possible; Given the differences in the shape and size of the vertebral bodies, without strict requirements for the consistency of the ROI size, as much of the trabecular area as possible should be covered; Considering the influence of the subject's position or common and unavoidable conditions in clinical practice such as mild lumbar scoliosis, slippage, and rotation, the measurement should be carried out on the mid - sagittal plane that includes most of the vertebral bodies as much as possible; When the subject's position affects the examination or there are mild lumbar scoliosis, slippage, or rotation, the mid - sagittal plane section selected for measurement should be as close as possible to most of the vertebral bodies; at the same time, be sure to pay attention to the position of the CSF ROI, that is, the ROI should be placed directly behind the L3 cerebrospinal fluid; The calculation method of the VBQ of a single vertebral body is as follows: The VBQ is calculated based on the signal intensity of a single vertebral body and the signal intensity of the cerebrospinal fluid ROI in the T1, T2, T2-STIR, and T1-FLAIR sequences. The calculation formula for the VBQ of a single vertebral body is VBQ 椎体序号 = SI (椎体序号)平均值 / SI CSF(L3) , which are respectively named VBQ T11 , VBQ T12 , VBQ L1 , VBQ L2 (the above four sequences need to be measured for each vertebral body), and the vertebral body number is indicated by a subscript; The calculation of the average VBQ of the L1-L2 and T11-L2 vertebral bodies is based on the T1, T2, T2-STIR, and T1-FLAIR sequences. The calculation formula for the VBQ of the L1-2 vertebral bodies is: VBQ L1-2 = SI (L1-2)平均值 / SI CSF(L3) , which are respectively named VBQT1 L1-2 , VBQT2 L1-2 , VBQT2-STIR L1-2 , VBQT1-FLAIR L1-2 ; The calculation formula for the VBQ of the T11-L2 vertebral body is obtained by replacing L1-2 with T11-L2 (the VBQ of the L1-2 and T11-L2 vertebral bodies is calculated separately for the above four sequences, and the vertebral body number is indicated by a subscript); Based on the above, the general calculation formula for the multi-sequence optimized VBQ calculation method is as follows: The calculation formula for the fat-suppressed sequence optimized VBQ is VBQ （椎体序号）压脂优化 = VBQT1 （椎体序号） × VBQT2 （椎体序号） / VBQT2-STIR （椎体序号） ; The calculation formula for the water-suppressed sequence optimized VBQ is: VBQ （椎体序号）压水优化 = VBQT1 （椎体序号） × VBQT2 （椎体序号） / VBQT1-FLAIR （椎体序号） ; The VBQ formula jointly optimized by the fat-suppressed and water-suppressed sequences is: VBQ （椎体序号）共同优化 = VBQT1 （椎体序号） × VBQT2 （椎体序号） / VBQT1-FLAIR （椎体序号） × VBQT2-STIR （椎体序号） , and the vertebral body number is indicated by a subscript; Step 4, adopt binary logistic regression to combine with the said VBQ 压脂优化, gender, age, height, weight and BMI were used to construct a screening group prediction model and a diagnostic group prediction model; The formula of the screening group prediction model is: 1 - P = 1 / [1 + exp(76.330 + 0.128×age - 50.861×height + 0.458×weight - 1.250×BMI + 0.194×VBQ 压脂优化 + A)], where: A is the gender constant, for females: A = -0.592; for males: A = 0; The formula of the diagnostic group prediction model is: 1 - P = 1 / [1 + exp(50.651 + 0.114×age - 39.833×height + 0.442×weight - 1.081×BMI + 0.428×VBQ 压脂优化 + A)], where: A is the gender constant, if female, then A = 0.777, if male A = 0; Taking the average value of the L1 - 2 vertebrae as the grouping criteria for screening and diagnosis respectively, ROC curve analysis was carried out. The results showed that in the screening group and the diagnostic group, neither the VBQT1 - FLAIRL1 - 2 nor the VBQT2 - STIRL1 - 2 sequences had statistical significance (P > 0.05), while in the screening group and the diagnostic group, the sequence with the highest area under the curve (AUC) was VBQ (L1 - 2) 压脂优化 , and the specific data were: in the screening group, the AUC of VBQ (L1 - 2) 压脂优化 was 0.756, the maximum Youden index reached 0.492, the sensitivity was 63.2%, the specificity was 86.0%, and the optimal cut - off value was 8.4059; in the diagnostic group, the AUC of VBQ (L1 - 2) 压脂优化 was 0.912, the maximum Youden index was 0.82, the sensitivity was 86.0%, the specificity was 82.0%, and the optimal cut - off value was also 8.4059. It can be seen that the sensitivity of the diagnostic group is relatively higher and the specificity is relatively lower. It should be noted that the cut - off values (CUT - OFF values) of the screening group and the diagnostic group are both 8.4059, which indicates that this model performs poorly in distinguishing the cut - off values for screening and diagnosis. It is speculated that this may be related to the selection of the physical examination population and the relatively small sample size. The detailed data are shown in Table 1; Table 1 In summary, VBQ 压脂优化(Multiple sequences) The AUC in the diagnostic group was above 0.912, with a sensitivity ≥86% and a specificity ≥82%, significantly better than other sequences (e.g., the AUC of a single T1 sequence was 0.879); the AUC value of the prediction model for screening those with non-normal bone mass was = 0.917 (P < 0.01) (the AUC-VBQ(L1-2) before constructing the prediction model 压脂优化 = 0.756); the AUC for diagnosing osteoporosis was 0.934 (P < 0.01) (the AUC-VBQ(L1-2) for diagnosis before constructing the prediction model 压脂优化 = 0.912), as Figure 3 、 4 shown. Since the prediction model is used for probability prediction rather than classification, no specific cut-off value was set. The AUC of the fat-suppressed optimized sequence in the diagnostic group reached 0.912 (sensitivity 86%, specificity 82%), significantly better than other sequences (e.g., the AUC of a single T1 sequence was 0.879).

[0025] After the prediction model integrated multiple factors (age, gender, BMI, VBQ 压脂优化 ), the AUC of the prediction model in the diagnostic group was further increased to 0.934, and the AUC of the prediction model in the screening group reached 0.917 (the original single VBQ 压脂优化 was 0.756); In summary, this prediction model can improve the significance of screening and diagnosis, but further external validation is still needed to evaluate its generalization performance and further increase the sample size to improve accuracy, as Figure 5 、 6 shown; Using binary logistic regression combined with VBQ(L1-2) 压脂优化 (SI), age (years), gender (male or female), weight (kg), height (m), BMI (kg / m²) to obtain the probability value and the prediction model; The following are application examples of the prediction model in the screening group and the diagnostic group: The formula for the prediction model in the screening group is: 1 - P = 1 / [1 + exp(76.330 + 0.128×age - 50.861×height + 0.458×weight - 1.250×BMI + 0.194×VBQ 压脂优化 + A)] where: A is the gender constant (female: A = -0.592; male: A = 0); The formula for the prediction model in the diagnostic group is: 1 - P = 1 / [1 + exp(50.651 + 0.114×age - 39.833×height + 0.442×weight - 1.081×BMI + 0.428×VBQ 压脂优化 + A)] where: A is the gender constant, if female, then A = 0.777, if male A = 0; Example data: Assume Age: 60 years old Gender: Female Height: 1.60 meters Weight: 55 kilograms BMI: 21.48 kg / m² (Calculation formula: 55 / (1.6²) ≈ 21.48) VBQ fat-suppressed optimization (L1-2): 9.0 1. Screening group calculation (judging whether it is the "non-normal bone mass group") formula: 1 - P = 1 / [1 + exp(76.330 + 0.128×Age - 50.861×Height + 0.458×Weight - 1.250×BMI + 0.194×VBQ fat-suppressed optimization + A)]. Substitute data: A (female) = -0.592. Calculation process: Z = 76.330 + 0.128×60 - 50.861×1.6 + 0.458×55 - 1.250×21.48 + 0.194×9 + (-0.592) = 76.330 + 7.68 - 81.378 + 25.19 - 26.85 + 1.746 - 0.592 = 2.126. Probability calculation: 1 - P = 1 / (1 + exp(2.126)) ≈ 1 / (1 + 8.38) ≈ 0.107. P (probability of non-normal bone mass group) = 1 - 0.107 = 0.893 (89.3%) 2. Diagnostic group calculation (judging whether it is the "osteoporosis group") formula: 1 - P = 1 / [1 + exp(50.651 + 0.114×Age - 39.833×Height + 0.442×Weight - 1.081×BMI + 0.428×VBQ 压脂优化 + A)]. Substitute data: A (female) = 0.777. Calculation process: Z = 50.651 + 0.114×60 - 39.833×1.6 + 0.442×55 - 1.081×21.48 + 0.428×9 + 0.777 = 50.651 + 6.84 - 63.733 + 24.31 - 23.23 + 3.852 + 0.777 = -0.533. Probability calculation: 1 - P = 1 / (1 + exp(-0.533)) ≈ 1 / (1 + 0.587) ≈ 0.630. P (probability of osteoporosis group) = 1 - 0.630 = 0.370 (37.0%); Result interpretation Screening group (non-normal bone mass group): 89.3% probability, indicating a relatively high risk of abnormal bone mass.

[0026] Diagnosis group (osteoporosis group): 37.0% probability, currently with a relatively low risk of osteoporosis, but regular reexamination is required in combination with age.

[0027] Key instructions 1 Gender constant: For females, A = -0.592 is used in the screening model and A = 0.777 is used in the diagnosis model.

[0028] 2 Clinical threshold: A probability > 50% (P > 0.5) indicates a high risk.

[0029] 3 Conclusion of this case: The subject needs to pay attention to bone health, but there is no need to be diagnosed with osteoporosis for the time being.

[0030] Example 2, a multi-modal fusion prediction system for vertebral bone mass, based on the multi-modal fusion prediction method for vertebral bone mass, includes a data acquisition module, a bone density grouping module, a VBQ value processing module, and a prediction model construction module. The data acquisition module is configured to: acquire the biometric data and vertebral image data of the object to be measured; the basic data includes gender, age, height, weight, and BMI; the vertebral image data includes the three-dimensional bone density parameters of the T11-L2 vertebrae and multi-modal magnetic resonance images, and the multi-modal magnetic resonance images cover at least the T1, T2, T2-STIR, and T1-FLAIR sequences; The bone density grouping module is configured to: based on the three-dimensional bone density parameters, generate the bone density grouping results of the screening group and the diagnosis group through the preset bone density thresholds of the osteoporosis group, non-osteoporosis group, normal bone mass group, and non-normal bone mass group; The VBQ value processing module is configured to: set regions of interest ROI in the trabecular bone region of the T11-L2 vertebrae and the cerebrospinal fluid region at the L3 level in the multi-modal magnetic resonance images, and extract the vertebral signal intensity and cerebrospinal fluid signal intensity of each sequence; Calculate the single vertebral VBQ value, the average VBQ of the L1-L2 vertebrae, and the average VBQ of the entire T11-L2 segment based on the signal intensity; Perform sequence optimization processing on the VBQ value to generate VBQ fat-suppressed optimization, VBQ water-suppressed optimization, and VBQ common optimization parameters; The prediction model construction module is configured to: use binary logistic regression combined with the VBQ fat-suppressed optimization, gender, age, height, weight, and BMI to construct a screening group prediction model and a diagnosis group prediction model, and obtain the probability value and the prediction model.

[0031] The above is a further detailed description of the present invention in conjunction with specific embodiments, and it cannot be determined that the specific implementation of the present invention is only limited thereto; for those skilled in the art of the present invention and related technical fields, based on the technical solution idea of the present invention, the expansions made, as well as the replacement of operation methods and data, should all fall within the protection scope of the present invention.

Claims

1. A multi-modal fusion prediction method for vertebral bone mass, characterized in that: The method includes the following steps: Step 1: Obtain the biometric data and vertebral imaging data of the object to be measured; the basic data includes gender, age, height, weight, and BMI; the vertebral imaging data includes the three-dimensional bone density parameters of the T11-L2 vertebrae and multimodal magnetic resonance images, and the multimodal magnetic resonance images cover at least the T1, T2, T2-STIR, and T1-FLAIR sequences; Step 2: Based on the three-dimensional bone density parameters, generate the bone density grouping results for the screening group and the diagnosis group through the preset bone density thresholds of the osteoporosis group, non-osteoporosis group, normal bone mass group, and non-normal bone mass group; Step 3: Set regions of interest (ROIs) in the trabecular bone region of the T11-L2 vertebrae and the cerebrospinal fluid region at the L3 level in the multimodal magnetic resonance images, and extract the vertebral signal intensity and cerebrospinal fluid signal intensity of each sequence; Calculate the VBQ value of a single vertebra, the average VBQ of the L1-L2 vertebrae, and the average VBQ of the entire T11-L2 segment based on the signal intensity; Perform sequence optimization processing on the VBQ value to generate VBQ 压脂优化 、VBQ 压水优化 and VBQ 共同优化 parameters; Step 4: Use binary logistic regression combined with the VBQ fat-suppression optimization, gender, age, height, weight, and BMI to construct a screening group prediction model and a diagnosis group prediction model, and obtain the probability value and the prediction model.

2. The multi-modal fusion prediction method for vertebral bone mass according to claim 1, wherein: The three-dimensional bone density parameters include the independent vBMD values of each vertebral body from T11 to L2, the average vBMD of the vertebral bodies from L1 to L2, vBMD L1-2 , and the average vBMD of the entire segment from T11 to L2, vBMD T11-L2 .

3. The vertebral bone mass multimodal fusion prediction method according to claim 1, wherein: The method for setting the bone density threshold is as follows: Osteoporosis group threshold: vBMD L1-2 < 80 mg / cm³; Threshold for non-osteoporosis group: vBMD L1-2 ≥ 80 mg / cm³; Threshold for the normal bone mass group: vBMD T11-L2 > 120 mg / cm³; Threshold for the non-normal bone mass group: vBMD T11-L2 ≤ 120 mg / cm³.

4. The vertebral bone mass multimodal fusion prediction method according to claim 1, characterized in that: The formula of the screening grouping prediction model is: 1 - P = 1 / [1 + exp(76.330 + 0.128 × age - 50.861 × height + 0.458 × weight - 1.250 × BMI + 0.194 × VBQ 压脂优化 + A)], where: A is the gender constant, for female: A = -0.592; for male: A = 0; The formula for the diagnosis group prediction model is: 1-P = 1 / [1+exp(50.651 + 0.114×age - 39.833×height + 0.442×weight - 1.081×BMI + 0.428×VBQ fat-suppression optimization + A)], where: A is the gender constant, if it is female, then A = 0.777, if it is male, A = 0.

5. The vertebral bone mass multimodal fusion prediction method according to claim 1, characterized in that: The sequence optimization process for the VBQ value: Optimize the vertebral body-cerebrospinal fluid contrast using a fat suppression algorithm for the T2-STIR sequence, and calculate the VBQ 压脂优化 parameters; optimize the background noise using a water suppression algorithm for the T1-FLAIR sequence, and calculate the VBQ 压水优化 parameters; perform weighted fusion on the VBQ values of the dual suppression sequence to generate the VBQ 共同优化 parameters.

6. The vertebral bone mass multimodal fusion prediction method according to claim 1, characterized in that: The AUC values of the diagnostic grouping prediction model and the screening grouping prediction model using VBQ were analyzed through the ROC curve. 压脂优化 after 7. The vertebral bone mass multimodal fusion prediction method according to claim 1, wherein: The method for setting the ROI is as follows: The first ROI is an elliptical region covering the cancellous bone region of the vertebra and avoiding the cortical bone and vertebral venous plexus; the second ROI is within the cerebrospinal fluid space at the L3 level.

8. Vertebral bone mass multimodal fusion prediction system, based on the vertebral bone mass multimodal fusion prediction method according to any one of claims 1-6, characterized in that: It includes a data acquisition module, a bone density grouping module, a VBQ value processing module, and a prediction model construction module. The data acquisition module is configured to: Obtain the biometric data and vertebral imaging data of the object to be measured; the basic data includes gender, age, height, weight, and BMI; the vertebral imaging data includes the three-dimensional bone density parameters of the T11-L2 vertebrae and multimodal magnetic resonance images, and the multimodal magnetic resonance images cover at least the T1, T2, T2-STIR, and T1-FLAIR sequences; The bone density grouping module is configured to: Based on the three-dimensional bone density parameters, generate the bone density grouping results for the screening group and the diagnosis group through the preset bone density thresholds of the osteoporosis group, non-osteoporosis group, normal bone mass group, and non-normal bone mass group; The VBQ value processing module is configured to: Set regions of interest (ROIs) in the trabecular bone region of the T11-L2 vertebrae and the cerebrospinal fluid region at the L3 level in the multimodal magnetic resonance images, and extract the vertebral signal intensity and cerebrospinal fluid signal intensity of each sequence; Calculate the VBQ value of a single vertebra, the average VBQ of the L1-L2 vertebrae, and the average VBQ of the entire T11-L2 segment based on the signal intensity; Perform sequence optimization processing on the VBQ value to generate VBQ fat suppression optimization, VBQ water suppression optimization, and VBQ common optimization parameters; The prediction model construction module is configured to: use binary logistic regression to combine the VBQ fat suppression optimization, gender, age, height, weight, and BMI to construct a screening group prediction model and a diagnostic group prediction model, and obtain probability values and prediction models.

Images