Lumbar multifidus muscle rehabilitation exercise method based on multi-slice multi-frequency ultrasonic image fusion

Abstract

A lumbar multifidus muscle rehabilitation exercise evaluation method based on multi-slice multi-frequency ultrasonic image fusion, comprising the following steps: step S1: multi-slice, multi-frequency ultrasonic image acquisition; step S2: image preprocessing and target region extraction; step S3: fusion input construction and feature extraction; step S4, model inference and evaluation output; step S5: rehabilitation related state evaluation; step S6: rehabilitation training scheme generation; step S7: reevaluation after training execution and scheme adjustment. The method breaks through the physical trade-off bottleneck of single frequency band imaging: through high and low frequency cooperation, high resolution texture information of shallow fascia and complete ultrasonic image of deep muscle belly are obtained at the same time, solving the technical contradiction of "clear but not transparent" or "transparent but not clear" in the evaluation of obese patients or deep (L5-S1) multifidus muscle.

Description

Technical Field

[0001] This invention relates to the field of sports rehabilitation technology, specifically to a method for rehabilitation training of the lumbar multifidus muscle based on multi-slice and multi-frequency ultrasound image fusion. Background Technology

[0002] The lumbar multifidus (LM) is a core muscle group maintaining segmental spinal stability, and its injury is closely related to chronic low back pain (CLBP). Clinical studies have shown that targeted rehabilitation exercises can effectively promote the recovery of multifidus function, but the objective assessment of exercise effects remains a key bottleneck restricting precise rehabilitation. Currently, ultrasound imaging technology has become the preferred imaging method for musculoskeletal rehabilitation assessment due to its portability, real-time capability, and lack of ionizing radiation; however, existing technologies have significant limitations in rehabilitation exercise assessment scenarios. Summary of the Invention

[0003] The existing imaging techniques for assessing the multifidus muscle are as follows: (1) Two-dimensional ultrasound single-section assessment technique: Existing techniques typically use transverse or parasagittal sections to acquire images of the multifidus muscle, and then manually record and measure muscle thickness, cross-sectional area, etc. For example... Figure 1 As shown, a single transverse section image can display the anatomical relationship between the multifidus muscle and the lamina and spinous process, but it cannot fully display the long axis course of the muscle and the continuity of the fascia; a single parasagittal section can display the direction of muscle fibers, but the display of deep muscle bellies and vertebral attachment points is limited, making it impossible to assess muscle changes before and after rehabilitation exercises. Existing technologies are limited by the frequency characteristics of the probe, and usually use a single center frequency (such as 10 MHz or 12 MHz) for scanning. High-frequency ultrasound (≥12 MHz) can provide high-resolution images of the superficial fascia and perifascia (such as... Figure 2 As shown), its penetration depth is insufficient (usually <4 cm), making it difficult to clearly display the complete structure of multifidus muscle in obese patients or deep layers (L4-L5, L5-S1 segments); low-frequency ultrasound (5-8 MHz), although capable of penetrating to deep muscle bellies (such as...), is not sufficient. Figure 2 (As shown in the figure), but the spatial resolution is reduced, making it difficult to identify subtle morphological changes in structures such as fascia and muscles before and after rehabilitation exercises.

[0004] (2) Multi-parameter ultrasound acquisition technology: Some existing technologies attempt to combine two-dimensional grayscale ultrasound, shear wave elastography and color Doppler blood flow imaging to obtain morphological, stiffness and blood flow perfusion parameters respectively. However, multi-parameter examination places higher demands on ultrasound instruments and the operation process is more complicated, which limits its clinical application.

[0005] (3) Three-dimensional ultrasound volumetric imaging technology: Recently, there are technologies that use three-dimensional volumetric probes to collect data on the multifidus muscle of the lumbar region, and display the overall morphology of the muscle through three-dimensional reconstruction (e.g., Figure 3 (As shown). However, three-dimensional ultrasound still relies on single-band imaging (usually medium frequency 6-9 MHz), which cannot simultaneously obtain high-resolution surface structure images and deep tissue acoustic images in a single scan, making it difficult to meet the needs of rehabilitation exercise assessment.

[0006] (4) Magnetic Resonance Imaging (MRI) Assessment Techniques: MRI is widely recognized as the gold standard for assessing chronic multifidus muscle injury in the lumbar region (especially fatty infiltration, muscle fiber atrophy, and fibrotic changes). Through T1-weighted images, T2-weighted images, and Dixon fat quantification sequences, MRI can clearly show the signal differences between muscle and adipose tissue, quantitatively measure the rate of fat infiltration and relative cross-sectional area, and has extremely high detection sensitivity for occult intramuscular edema. However, MRI examinations are time-consuming and expensive, and are contraindicated for patients with internal metal implants (such as shrapnel, fixation plates) and claustrophobia, limiting its universality in assessment.

[0007] The main drawbacks of existing technologies are: (1) the isolation of information in single-band imaging. The anatomical layers of the multifidus muscle in the lumbar region are complex, and the physical requirements of ultrasound frequency vary between its superficial fascia, intermediate muscle belly, and deep vertebral attachment areas. (2) the dimensional limitations of single-section assessment. Multifidus muscle injuries often manifest as discontinuous lesions in three-dimensional space (such as focal muscle bundle tears and fatty infiltration bands). (3) Existing technologies have not yet established spatial registration and information fusion algorithms between ultrasound images of different frequency bands, and cannot perform pixel-level or feature-level fusion of the texture details of high-frequency images with the deep structural information of low-frequency images. This limits the comprehensive assessment ability of ultrasound for multifidus muscle rehabilitation exercises.

[0008] Therefore, there is an urgent need for an ultrasound image analysis method that can integrate multi-sectional anatomical information and multi-frequency physical information to achieve objective and accurate assessment of rehabilitation exercises for the multifidus muscle of the lumbar region. The multifidus muscle rehabilitation exercise method for the lumbar region based on multi-sectional and multi-frequency ultrasound image fusion proposed in this invention has the following technical advantages: (1) Breakthrough in the physical trade-off bottleneck of single-band imaging: Through high and low frequency synergy, high-resolution texture information of superficial fascia (axial resolution ≤0.3 mm) and complete acoustic images of deep muscle belly (penetration depth ≥6 cm) are obtained simultaneously, solving the technical contradiction of "seeing clearly but not penetrating" or "seeing clearly but not seeing clearly" in the assessment of obese patients or deep (L5-S1) multifidus muscle.

[0009] (2) By mapping the transverse section image and the sagittal parasection image to the same target segment for joint evaluation, information such as muscle cross-sectional area, contour morphology, muscle fiber orientation and local echo distribution can be characterized simultaneously, thereby improving the ability to identify complex lesions such as fat infiltration, focal damage and structural continuity abnormalities.

[0010] (3) Improved information utilization and reduced bias caused by a single view. The present invention constructs images acquired by different probes, different frequency bands and different cross sections into single-input, dual-input or multi-input fusion modes. By feature-level fusion, it makes full use of the complementary information of various images, reduces the information loss caused by a single view or a single frequency band, and improves the stability of the evaluation results.

[0011] (4) Improved the standardization of operation: By using the preset cross-section acquisition order and fusion algorithm, the dependence on the operator's personal experience is reduced; by using a pre-trained convolutional neural network model for transfer learning, the medical ultrasound images are fine-tuned to automatically extract multi-scale and multi-level discriminative features, reducing the dependence on manual feature design and improving the model's generalization ability in small sample medical data.

[0012] (5) Objective quantitative assessment of the degree of injury is achieved. Based on the trained image analysis model, this invention can output the predicted probability, classification results and corresponding confidence scores of each injury category, transforming the traditional subjective interpretation that relies on human experience into repeatable and comparable numerical output, thereby improving the consistency of assessment results and providing quantitative basis for clinical follow-up, efficacy observation and auxiliary diagnosis.

[0013] (6) It has good clinical application potential and scalability: It is suitable for conventional two-dimensional ultrasound equipment, does not require additional expensive hardware, can be embedded in existing medical image analysis systems, and supports further expansion to other muscle or soft tissue injury assessment tasks, thus having good clinical promotion value. Attached Figure Description

[0014] Figure 1 This is a schematic diagram of multi-section, multi-frequency band image acquisition in an embodiment of the present invention.

[0015] Figure 2 This is a schematic diagram of image preprocessing and target region extraction in an embodiment of the present invention.

[0016] Figure 3 This is a schematic diagram of multimodal input construction in an embodiment of the present invention.

[0017] Figure 4 This is a schematic diagram of auxiliary diagnostic analysis in an embodiment of the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments and the accompanying drawings. However, those skilled in the art will understand that this invention is not limited to the accompanying drawings and the following embodiments.

[0019] (1) Overall plan and technical route This invention proposes a rehabilitation exercise assessment scheme for the lumbar multifidus muscle based on multi-slice, multi-band ultrasound image fusion. This scheme not only acquires structural and functional information of the lumbar multifidus muscle but also generates corresponding rehabilitation training plans based on the assessment results and dynamically adjusts them during the training process. The method includes the following steps: Step S1: Acquisition of multi-section, multi-frequency ultrasound images Multi-sectional, multi-frequency ultrasound image acquisition was performed on the target lumbar spine segment of the target subject to obtain at least one of the following images: - Cross-sectional image of a high-frequency linear array probe; - Sagittal side-section image from a high-frequency linear array probe; - Cross-sectional image of a low-frequency convex array probe; - Sagittal side section image of a low-frequency convex array probe.

[0020] Step S2: Image preprocessing and target region extraction The acquired ultrasound images were preprocessed using standardization, and the target region of the lumbar multifidus muscle was located, segmented, or extracted to form standardized input data for model analysis.

[0021] Step S3: Input Fusion Construction and Feature Extraction Images from different cross sections and frequency bands are constructed as single-input, dual-input, or multi-input fused data. Images from different sources are input into a feature extraction network. The feature extraction network extracts feature information such as superficial fascia texture, muscle belly morphology, deep attachment zone structure, muscle bundle course continuity and echo distribution, and thickness change rate before and after dynamic contraction, and then performs fusion analysis.

[0022] Step S4, Model Inference and Evaluation Output The model is used to train or infer on the input image and output the classification result and corresponding probability information of the degree of multifidus muscle injury in the lumbar region. Based on the model output, confidence level, low confidence sample prompts or segmented statistical results are further generated to assist clinical interpretation.

[0023] Step S5: Assessment of Rehabilitation-Related Status Based on the fusion features, the system outputs rehabilitation-related status information for the target subject's lumbar multifidus muscle. In one embodiment, the subject is categorized into different training levels, such as Level 1, Level 2, and Level 3 rehabilitation. The status information includes one or more of the following: fat infiltration level; muscle symmetry level; and rehabilitation training adaptation level.

[0024] Step S6: Generation of Rehabilitation Training Plan Based on the rehabilitation-related status information obtained in step S5, a preset training mapping rule or training scheme generation model is invoked to automatically generate a corresponding rehabilitation training scheme. The rehabilitation training scheme includes one or more of the following: training objectives; recommended training movement categories; training postures; duration of a single training session; duration of holding a single movement; number of repetitions; number of training sets; and weekly training frequency.

[0025] Step S7: Post-training review and program adjustment After a pre-set training cycle, the subjects underwent another round of multi-sectional, multi-frequency ultrasound image acquisition and fusion analysis. The reassessment results were compared with the initial assessment results to obtain information on rehabilitation changes. Based on this information, the original rehabilitation training plan was maintained, enhanced, downgraded, or replaced to achieve individualized dynamic rehabilitation management.

[0026] In one embodiment, if the reassessment results show improvement in superficial texture, increased integrity of deep muscle belly contours, reduced echo abnormalities, or enhanced contraction response, then the training intensity is increased or advanced movements are introduced; if the reassessment results show insufficient improvement, increased risk of low confidence, or worsening of structural abnormalities, then the current training level is maintained, the load is reduced, or a manual review is requested.

[0027] (2) Specific implementation scheme description (2.1) Acquisition of multi-section, multi-frequency ultrasound images In this embodiment, different types of ultrasound probes are used to scan the multifidus muscle of the lumbar region to obtain image information at different cross sections and frequency bands. The images include at least one of the following: a cross-sectional image obtained by a high-frequency linear array probe (linear array A), a sagittal para-sectional image obtained by a high-frequency linear array probe (linear array S), a cross-sectional image obtained by a low-frequency convex array probe (convex array A), and a sagittal para-sectional image obtained by a low-frequency convex array probe (convex array S).

[0028] To ensure comparability between different images, the images were acquired corresponding to the same lumbar vertebral segment or the same anatomical region. In actual acquisition, the acquisition location was standardized using surface landmarks, bony landmarks, fascial boundaries, or a preset scanning protocol. This acquisition method can simultaneously obtain information reflecting superficial tissue details and information reflecting the overall morphology of deep tissues, forming a complementary relationship between lateral structural information and longitudinal course information.

[0029] (2.2) Image preprocessing and target region extraction The acquired raw ultrasound images are preprocessed, including at least one of grayscale normalization, size uniformity, noise reduction, and contrast enhancement.

[0030] Subsequently, the target region of the multifidus muscle in the lumbar region is determined. The determination of the target region is achieved through manual delineation, semi-automatic interactive selection, or automatic recognition. When automatic recognition is used, the multifidus muscle region is located based on fascial boundaries, bony echo interfaces, muscle contours, or pre-trained segmentation models.

[0031] Image preprocessing and target region extraction can reduce data differences caused by different devices, operators, and acquisition conditions, thereby improving the consistency of subsequent feature extraction and classification analysis.

[0032] (2.3) Construction of multimodal input (corresponding to step S3 above) Depending on the experimental design or clinical application requirements, input data can be constructed into different patterns. These patterns include, but are not limited to: single-input pattern, dual-input fusion pattern, and multi-input fusion pattern. Single input mode: Linear array A, Linear array S, Convex array A, Convex array S Dual-input fusion mode: Linear array A + Linear array S Convex array A + Convex array S Multi-input fusion mode: All four types of images are input together. For dual or multiple input scenarios, multiple feature extraction branches can be set up to extract features from each input image and then perform splicing, weighted fusion, attention fusion or other feature-level fusion, and finally output the final result through the classification head.

[0033] By using the above construction method, complementary information between images from different sources can be fully utilized to improve the characterization ability of the spatial structure and damage degree of the lumbar multifidus muscle.

[0034] (2.4) Classification model construction, training and result output (corresponding to step S4 above) In a preferred embodiment, a convolutional neural network is used as the classification model. The convolutional neural network is VGG16, ResNet, DenseNet, EfficientNet, or other network structures with image classification capabilities; among them, VGG16 is a preferred implementation.

[0035] During training, a transfer learning strategy was employed. Model parameters pre-trained on a general image dataset were used as initial parameters, and the model was fine-tuned using ultrasound image data according to this invention to adapt to the grading task of lumbar multifidus muscle injury. The severity of lumbar multifidus muscle injury was divided into three grades, and the evaluation efficacy for different modes and frequency bands is as follows:

[0036] The model outputs probability values ​​corresponding to each injury category. The category with the highest probability is taken as the prediction result, and this highest probability value is used as the prediction confidence level. In one embodiment, if the model outputs a probability vector of P=(p0, p1,p2), then the predicted category is argmax(P), and the prediction confidence level is max(P). Compared to relying solely on human experience for interpretation, the above model can provide stable numerical results, facilitating comparison of evaluation results at different time points, for different subjects, and under different input combinations.

[0037] (2.5) Auxiliary diagnostic analysis (corresponding to step S4 above) After the model obtains the classification results, further auxiliary diagnostic analysis can be performed on the model output. This auxiliary diagnostic analysis not only outputs the predicted category but also combines confidence distribution, threshold segmentation statistics, and category distribution characteristics to provide quantitative indicators of the reliability of the results.

[0038] ① Screening of Low-Confidence Samples: In one specific embodiment, statistics were performed on 957 samples. When the confidence threshold was set to 0.6, 367 samples had a confidence level ≤ 0.6, accounting for 38.35% of all samples. Among them, 192 samples were correctly predicted, accounting for 52.32% of the low-confidence samples; and 175 samples were incorrectly predicted, accounting for 47.68% of the low-confidence samples. Therefore, samples with a confidence level ≤ 0.6 can be considered as key review targets for manual review, supplementary interpretation, or re-collection.

[0039] ② Segmented Statistical Analysis: The number of samples at different confidence thresholds was statistically analyzed. The results showed that: 143 samples had a confidence level ≤ 0.5, accounting for 14.94%; 367 samples had a confidence level ≤ 0.6, accounting for 38.35%; 511 samples had a confidence level ≤ 0.7, accounting for 53.40%; 665 samples had a confidence level ≤ 0.8, accounting for 69.49%; and 827 samples had a confidence level ≤ 0.9, accounting for 86.42%. This segmented statistical analysis can provide a basis for setting different levels of warning thresholds and review strategies in clinical settings.

[0040] ③ Category Distribution Analysis: Further analysis of the proportion of low-confidence samples by predicted category revealed the following: Category 0 had 92 / 434 low-confidence samples, accounting for 21.20%; Category 1 had 207 / 343 low-confidence samples, accounting for 60.35%; and Category 2 had 68 / 180 low-confidence samples, accounting for 37.78%. This result indicates that Category 1 has a relatively high proportion of low-confidence samples, representing a category requiring significant model optimization or thorough review in clinical applications.

[0041] ④ Error Correlation Analysis: Combining the confidence level distribution chart, the overall accuracy versus confidence level curve, and the categorical accuracy versus confidence level curve, it can be seen that the overall prediction accuracy of the model increases with increasing confidence level. However, in the low confidence level range, the number of correct and incorrect samples is close, indicating a strong correlation between low confidence output and the risk of misjudgment. Therefore, the confidence level of the model output can be used as a reliability indicator in auxiliary diagnosis to identify high-risk samples and assist doctors in interpretation.

[0042] Through the above-mentioned auxiliary diagnostic analysis, the reliability information of the results can be given at the same time as the classification results are output, and the model performance can be explained with specific statistical values ​​and graphical results, thereby improving the interpretability, traceability and practicality of the present invention in clinical applications.

[0043] (2.6) Rehabilitation training stratification and program generation based on assessment results (corresponding to steps S5-S6 above) In this embodiment, the results output by the classification model or fusion analysis model are not only used to characterize the current structural and functional state of the lumbar multifidus muscle, but also to generate a rehabilitation training program corresponding to this state.

[0044] 1. Construction of Rehabilitation Stratification Rules Based on the classification results, probability distribution, confidence information, and one or more imaging feature parameters output by the fusion model, the subjects are stratified for rehabilitation. These imaging feature parameters may include, but are not limited to: multifidus muscle cross-sectional area; muscle thickness; left-right symmetry index; degree of local echo enhancement; fat infiltration grade; fascial boundary clarity; integrity of deep attachment areas; and rate of change in thickness before and after dynamic contraction.

[0045] In one embodiment, the training layers are defined as follows: Level 1 rehabilitation training: Suitable for subjects with intact structures, low fat infiltration, relatively uniform echo distribution, and good contractile response, classified as Level 1. The training goals at this level are to enhance deep stability, improve muscle endurance, and gradually restore load capacity.

[0046] Level 2 rehabilitation training is suitable for subjects with mild to moderate structural abnormalities, uneven local echogenicity, decreased continuity of muscle fiber course, or weakened contractile response, classified as Level 2. The training goal at this level is to re-establish multifidus muscle contraction ability, improve muscle control, control compensation, and gradually restore functional stability.

[0047] Level 3 rehabilitation training: Suitable for subjects with significant fat infiltration, abnormal deep structural visualization, incomplete muscle belly contour, obvious left-right asymmetry, and model indications of Level 3. This level of training carries a higher risk, and the training goals are low-load activation, avoidance of excessive stress stimulation, and a focus on protective and basic stabilization training.

[0048] 2. Rehabilitation plan mapping generation For different rehabilitation levels, preset training action libraries and parameter libraries can be called to automatically generate corresponding training programs.

[0049] In one embodiment, the training exercise library includes at least one of the following: transverse abdominis-multifidus synergistic activation training; pelvic neutral position control training; four-point kneeling unilateral upper / lower limb control training; bridge stability training; bird-dog exercise; supine or prone low-load isometric contraction training; glute bridge exercise; side-lying leg raise; wall-assisted superman exercise; dead bug exercise; anti-rotation control training; functional posture maintenance training; and respiratory synergistic stability training.

[0050] Training principles: Keep the lumbar spine neutral in all movements, without collapsing or hunching the back; the intensity should be soreness and slight fatigue, and stop immediately if stinging pain occurs; 4-6 days a week, 10-15 minutes each time; avoid prolonged sitting, bending over to lift heavy objects, and slouching.

[0051] For Level 1 rehabilitation training, a relatively high training progression rate should be set. For example: single movement holding time of 8-15 seconds; repetitions of 10-15 times; 2-4 sets; training frequency of 4-6 times per week; and gradually adding unstable plane training, resistance training, or compound movement training.

[0052] For the second level of rehabilitation training, a moderate-intensity training program should be set. For example, the holding time for a single movement should be 5-10 seconds; The number of repetitions is 8 to 12; the number of training sets is 2 to 3; the training frequency is 3 to 5 times per week; low-compensation-risk movements are preferred, and false recruitment is reduced through mirror feedback, ultrasound feedback or posture guidance.

[0053] For the third level of rehabilitation training, a protective low-load training program can be set up. For example: the holding time of a single movement is 3 to 5 seconds; the number of repetitions is 5 to 8; the number of training sets is 1 to 2; the training frequency is 3 to 5 times per week; avoid large-amplitude trunk rotation, hyperextension or high-load resistance movements; when necessary, only perform basic activation training, breathing coordination training and small-range stability control training.

[0054] 3. Correspondence based on specific evaluation results In one embodiment, the correspondence between assessment results and rehabilitation plans can be established as follows: Scenario A: The model outputs "no obvious fat infiltration or mild abnormality," and the confidence level is higher than the first threshold. In this case, the subject is deemed to have good basic structural conditions, and a training program aimed at strengthening deep stability, improving endurance, and restoring functional activity can be generated. Recommended exercises include bridge pose, bird-dog pose, and anti-rotation control training, and the holding time and movement complexity can be gradually increased.

[0055] Scenario B: The model outputs "mild fatty infiltration / mild to moderate damage," with uneven local echogenicity or decreased muscle fiber continuity. In this case, the subject is diagnosed with some structural degeneration or functional abnormality, and a training program targeting deep activation and reconstruction, motor control correction, and symmetry restoration can be generated. It is recommended to start with low-load isometric activation and pelvic neutral position control training, and then gradually transition to four-point kneeling stability training and mild functional movement training.

[0056] Scenario C: The model output is "moderate to severe fat infiltration or significant structural abnormalities", or the training risk warning is high.

[0057] At this point, the subject was deemed unsuitable to directly enter the high-intensity training phase, and a training plan was generated targeting protective activation, low-load stability training, and movement restriction. Large-amplitude flexion, extension, and rotation were restricted during training, with a focus on short-duration, low-intensity, and monitorable basic activation training, and the reassessment cycle was shortened.

[0058] Scenario D: The maximum probability of the model's output category is lower than a preset threshold. In this case, the current result is deemed unreliable. The system can output a "recommend review" flag and generate a conservative training plan, or prompt the user to re-acquire images before formulating a training plan, in order to avoid making inappropriate training arrangements based on low-reliability results.

[0059] 4. Dynamic adjustment mechanism In one embodiment, the system re-acquires ultrasound images after completing a preset training cycle and calculates the change between the re-evaluation result and the initial result. The changes include: changes in fat infiltration level; changes in cross-sectional area; changes in muscle thickness; changes in the rate of change of contraction thickness; changes in left-right symmetry; changes in echo uniformity; changes in muscle bundle clarity; and changes in training adaptation level.

[0060] If the change meets the preset improvement conditions, the training program will be further adjusted; if the change does not reach the preset threshold, the original program will be maintained or the type of movement will be adjusted; if the change indicates a deteriorating trend, the training load will be reduced and a risk warning will be output.

[0061] For example, when the rate of change in shrinkage thickness increases beyond the first threshold and the structural continuity score improves, the training can be upgraded from static activation to dynamic stable training; when the left-right symmetry improves and the low-confidence risk decreases, the resistance level can be appropriately increased. When the range of echo abnormalities expands or the post-training score decreases, reduce the number of repetitions and the range of motion, and shorten the reassessment interval.

[0062] Through the aforementioned dynamic adjustment mechanism, individualized closed-loop optimization of rehabilitation training programs can be achieved.

Claims

1. A method for assessing rehabilitation exercises of the lumbar multifidus muscle based on multi-slice, multi-frequency ultrasound image fusion, characterized in that, The method includes the following steps: Step S1: Acquisition of multi-section, multi-frequency ultrasound images; Step S2: Image preprocessing and target region extraction; Step S3: Fusion of input construction and feature extraction; Step S4, Model Inference and Evaluation Output; Step S5: Assessment of rehabilitation-related status; Step S6: Generate a rehabilitation training plan; Step S7: Post-training review and program adjustment.

2. The method according to claim 1, characterized in that, Step S1 includes: acquiring multi-section, multi-frequency ultrasound images of the target lumbar vertebral segment of the target subject to obtain at least one of the following images: - Cross-sectional image of a high-frequency linear array probe; - Sagittal side-section image from a high-frequency linear array probe; - Cross-sectional image of a low-frequency convex array probe; - Sagittal side section image of a low-frequency convex array probe.

3. The method according to claim 1, characterized in that, Step S2 includes: The acquired ultrasound images were preprocessed using standardization, and the target region of the lumbar multifidus muscle was located, segmented, or extracted to form standardized input data for model analysis.

4. The method according to claim 1, characterized in that, Step S3 includes: Images from different cross sections and frequency bands are constructed as single-input, dual-input, or multi-input fused data. Images from different sources are input into a feature extraction network. The feature extraction network extracts feature information such as superficial fascia texture, muscle belly morphology, deep attachment zone structure, muscle bundle course continuity and echo distribution, and thickness change rate before and after dynamic contraction, and then performs fusion analysis.

5. The method according to claim 1, characterized in that, Step S4 includes: The model is used to train or infer on the input image and output the classification result and corresponding probability information of the degree of multifidus muscle injury in the lumbar region. Based on the model output, confidence level, low confidence sample prompts or segmented statistical results are further generated to assist clinical interpretation.

6. The method according to claim 1, characterized in that, Step S5 includes: Based on the fusion features, the system outputs the lumbar multifidus muscle rehabilitation-related status information of the target subject; the subject is divided into different training levels such as Level 1 rehabilitation, Level 2 rehabilitation, and Level 3 rehabilitation, and the status information includes one or more of the following: fat infiltration level; muscle symmetry level; rehabilitation training adaptation level.

7. The method according to claim 1, characterized in that, Step S6 includes: Based on the rehabilitation-related status information obtained in step S5, a preset training mapping rule or training scheme generation model is invoked to automatically generate a corresponding rehabilitation training scheme. The rehabilitation training scheme includes one or more of the following: training objectives; recommended training movement categories; training postures; duration of a single training session; duration of holding a single movement; number of repetitions; number of training sets; and weekly training frequency.

8. The method according to claim 1, characterized in that, Step S7 includes: After a preset training cycle, the subjects undergo multi-slice, multi-frequency ultrasound image acquisition and fusion analysis again, and the re-evaluation results are compared with the initial evaluation results to obtain rehabilitation change information; based on the change information, the original rehabilitation training plan is maintained, enhanced, downgraded, or replaced to achieve individualized dynamic rehabilitation management; If the reassessment results show improvement in superficial texture, increased integrity of deep muscle belly contours, reduced echo abnormalities, or enhanced contraction response, then increase the training intensity or introduce advanced movements; if the reassessment results show insufficient improvement, increased risk of low confidence, or worsening of structural abnormalities, then maintain the current training level, reduce the load, or request manual review.

9. The method according to claim 1 or 4, characterized in that, S3 includes: constructing input data into different modes according to experimental design or clinical application needs; the modes include, but are not limited to: single-input mode, dual-input fusion mode, and multi-input fusion mode. Single input mode: Linear array A, Linear array S, Convex array A, Convex array S Dual-input fusion mode: Linear array A + Linear array S Convex array A + Convex array S Multi-input fusion mode: All four types of images are input together. For dual and multiple input scenarios, multiple feature extraction branches are set up to extract features from each input image and then perform splicing, weighted fusion, attention fusion or other feature-level fusion, and finally output the final result through the classification head.

10. The method according to claim 1 or 5, characterized in that, S4 includes: using a convolutional neural network as a classification model, wherein the convolutional neural network is VGG16, ResNet, DenseNet, EfficientNet or other network structures with image classification capabilities; preferably VGG16; During the training process, a transfer learning strategy is adopted, using the model parameters pre-trained on a general image dataset as initial parameters, and fine-tuning the model with the ultrasound image data according to the present invention to adapt to the grading task of lumbar multifidus muscle injury. The model output is the probability value corresponding to each damage category. The category with the highest probability is taken as the prediction result, and the highest probability value is taken as the prediction confidence. Preferably, if the model output probability vector is P=(p0, p1, p2), then the predicted category is argmax(P) and the prediction confidence is max(P). Preferably, S5 and S6 further include: (1). Construction of rehabilitation stratification rules Based on the classification results, probability distribution, confidence information, and one or more imaging feature parameters output by the fusion model, the subjects are stratified for rehabilitation. The imaging feature parameters include, but are not limited to: multifidus muscle cross-sectional area; muscle thickness; left-right symmetry index; degree of local echo enhancement; fat infiltration grade; fascial boundary clarity; integrity of deep attachment area; and thickness change rate before and after dynamic contraction. The training layers are defined as follows: Level 1 rehabilitation training: suitable for subjects with intact structures, low degree of fat infiltration, relatively uniform echo distribution, and good contractile response; the training goal of this level is to enhance deep stability, improve muscle endurance, and gradually restore load capacity; Level 2 rehabilitation training: suitable for subjects with mild to moderate structural abnormalities, uneven local echoes, decreased continuity of muscle fiber course, or weakened contractile response; the training goal at this level is to rebuild multifidus muscle contraction ability, improve muscle control, control compensation, and gradually restore functional stability. Level 3 rehabilitation training is suitable for subjects with significant fat infiltration, abnormal deep structure, incomplete muscle belly contour, obvious left-right asymmetry, or whose models indicate a high training risk. The training goal at this level is low-load activation, avoiding excessive stress stimulation, and focusing on protective training and basic stability training. (2). Rehabilitation program mapping generation For different rehabilitation levels, the system calls up the preset training action library and parameter library to automatically generate the corresponding training plan; The training exercise library includes at least one of the following: transverse abdominis-multifidus synergistic activation training; pelvic neutral position control training; four-point kneeling unilateral upper / lower limb control training; bridge stability training; bird-dog exercise; supine or prone low-load isometric contraction training; glute bridge exercise; side-lying leg raise; wall-assisted superman exercise; dead bug exercise; anti-rotation control training; functional posture maintenance training; and respiratory synergistic stability training. Training principles: Keep the lumbar spine neutral in all movements, without slouching or arching the back; the intensity should be soreness and slight fatigue, and stop immediately if stinging pain occurs; 4-6 days a week, 10-15 minutes each time; avoid prolonged sitting, bending over to lift heavy objects, and slouching. For Level 1 rehabilitation training, a relatively high training progression rate is set, for example: single movement holding time of 8-15 seconds; repetitions of 10-15 times; training sets of 2-4 sets; training frequency of 4-6 times per week; and gradually adding unstable plane training, resistance training, or compound movement training. For the second level of rehabilitation training, a moderate-intensity training program should be set up, for example, holding a single movement for 5 to 10 seconds; The number of repetitions is 8-12; the number of training sets is 2-3; the training frequency is 3-5 times per week; low-compensation-risk movements should be prioritized, and false recruitment should be reduced through mirror feedback, ultrasound feedback or posture guidance; For the third level of rehabilitation training, a protective low-load training program should be set up; for example: the holding time of a single movement is 3 to 5 seconds; the number of repetitions is 5 to 8 times; the number of training sets is 1 to 2 sets; the training frequency is 3 to 5 times per week; avoid large-amplitude trunk rotation, hyperextension or high-load resistance movements; when necessary, only basic activation training, breathing coordination training and small-range stability control training should be performed.

Images