Bone tumor rehabilitation process evaluation method and system based on deep learning

Abstract

The invention discloses a bone tumor rehabilitation process evaluation method and system based on deep learning. The method comprises the following steps: S1, obtaining a preprocessed bone tumor postoperative multi-source rehabilitation data set; s2, generating a high-frequency rehabilitation feature set and a low-frequency rehabilitation feature set; s3, fusion rehabilitation feature representation is formed; s4, obtaining optimized model parameters; s5, using the optimized model parameters to replace corresponding parameters in the bone tumor postoperative rehabilitation trend discrimination model, performing supervised training on the preprocessed bone tumor postoperative multi-source rehabilitation data set, and outputting the trained bone tumor postoperative rehabilitation trend discrimination model; and S6, applying the trained bone tumor postoperative rehabilitation trend discrimination model to fusion rehabilitation feature representation of a patient to be discriminated, and generating a rehabilitation trend prediction result composed of a rehabilitation progress level, a rehabilitation trend curve and a prognosis risk level. The rehabilitation rhythm change and risk trend of the patient can be visually reflected, and a data-driven staging management basis is provided for doctors.

Description

Technical Field

[0001] The present invention relates to the field of bone tumor rehabilitation technology, and in particular to a bone tumor rehabilitation process evaluation method and system based on deep learning. Background Art

[0002] With the continuous development of medical artificial intelligence and deep learning technologies, postoperative rehabilitation assessment has gradually evolved towards an intelligent and data-driven direction. In the postoperative rehabilitation management of bone tumors, accurate identification of rehabilitation trends is of great significance for the formulation of clinical intervention plans, the assessment of patient prognostic risks, and the optimization of personalized rehabilitation pathways. However, in the existing medical system, the identification of postoperative rehabilitation trends after bone tumors still mostly relies on manual experience evaluation and rule-driven analysis models, which have significant problems such as strong subjectivity, poor timeliness, and low accuracy.

[0003] Currently, mainstream methods for identifying rehabilitation trends include manual scoring, conventional statistical regression models, and traditional convolutional neural networks. These methods have several shortcomings when dealing with multi-source, heterogeneous rehabilitation data. First, manual assessments are subject to significant subjective influences, making it difficult to maintain consistency within a standardized process. Second, conventional statistical models make strong assumptions about data distribution and struggle to adapt to the nonlinear, dynamically changing feature structures of rehabilitation data. Traditional convolutional neural networks, on the other hand, have limited feature extraction capabilities when faced with rehabilitation sequence data containing a high level of redundant information. This is particularly true when it comes to balancing changes related to the lesion area with changes in macro-trends, making it difficult to fully capture high- and low-frequency variation information.

[0004] In summary, there is an urgent need for an intelligent discrimination method that can integrate high-frequency and low-frequency features, optimize multi-objective parameters, and adapt to lesion area perception, so as to effectively improve the accuracy of rehabilitation trend identification and the overall robustness of the system. Summary of the Invention

[0005] One purpose of the present invention is to propose a bone tumor rehabilitation process assessment method and system based on deep learning. The present invention can intuitively reflect the changes in the patient's rehabilitation rhythm and risk trends, and provide doctors with a data-driven basis for staged management.

[0006] A method for evaluating bone tumor rehabilitation process based on deep learning according to an embodiment of the present invention includes the following steps:

[0007] S1. Collect multi-source rehabilitation data after bone tumor surgery, including medical imaging data, continuous physiological indicator sequence data, and rehabilitation assessment scale data, and uniformly encode them to construct a multi-source rehabilitation dataset after bone tumor surgery. Then, perform standardization, denoising, time series alignment, and missing data completion on the multi-source rehabilitation dataset after bone tumor surgery to obtain the pre-processed multi-source rehabilitation dataset after bone tumor surgery.

[0008] S2. Perform frequency domain division on the preprocessed multi-source rehabilitation dataset after bone tumor surgery based on the adaptive frequency allocation factor to generate a high-frequency rehabilitation feature set and a low-frequency rehabilitation feature set;

[0009] S3. Input the high-frequency rehabilitation feature set into the high-frequency improved Octave convolution channel, and input the low-frequency rehabilitation feature set into the low-frequency improved Octave convolution channel. Perform convolution operations on each channel while keeping the channels independent. Set cross-layer residual connections between the convolution layers of the improved Octave convolution network, and perform bidirectional compensation fusion on the convolution outputs of the high-frequency rehabilitation feature set and the low-frequency rehabilitation feature set to form a fused rehabilitation feature representation.

[0010] S4. Input the fused rehabilitation feature representation into the initial bone tumor postoperative rehabilitation trend discrimination model to complete the structural construction of the bone tumor postoperative rehabilitation trend discrimination model. Initialize the hybrid seagull search optimization module, use the bone tumor postoperative rehabilitation trend discrimination model parameters as search individuals, combine the global exploration strategy with the local perturbation strategy, establish a hybrid seagull search population, construct a multi-objective fitness function based on accuracy, verification loss, and training time, perform iterative optimization on the hybrid seagull search population, and obtain the optimized model parameters.

[0011] S5. Use the optimized model parameters to replace the corresponding parameters in the bone tumor postoperative rehabilitation trend discrimination model, perform supervised training on the preprocessed bone tumor postoperative multi-source rehabilitation dataset, and output the trained bone tumor postoperative rehabilitation trend discrimination model;

[0012] S6. Apply the trained bone tumor postoperative rehabilitation trend discrimination model to the fused rehabilitation feature representation of the patient to be discriminated, and generate a rehabilitation trend prediction result consisting of rehabilitation progress level, rehabilitation trend curve and prognostic risk level.

[0013] Optionally, the S1 includes the following steps:

[0014] S11. Collect raw multi-source rehabilitation data after bone tumor surgery and set a collection time range to construct a raw multi-source rehabilitation data set after bone tumor surgery. Each piece of raw multi-source rehabilitation data after bone tumor surgery includes a timestamp, a medical image vector, a physiological indicator vector, and a rehabilitation scale vector. The medical image vector represents the patient's corresponding medical image feature information at the time of collection, the physiological indicator vector represents the patient's continuous physiological state measurement value at that time, and the rehabilitation scale vector represents the corresponding assessment questionnaire score information.

[0015] S12. Standardize each data item in the original multi-source rehabilitation data set after bone tumor surgery. Standardization involves mapping the value of each dimension in the original data to a uniform numerical range based on its minimum and maximum values in that dimension through a linear normalization transformation, thereby obtaining a standardized multi-source rehabilitation data set after bone tumor surgery.

[0016] S13. Perform denoising on the standardized multi-source rehabilitation dataset after bone tumor surgery. A wavelet threshold denoising strategy is used to process the various dimensional elements in the medical image vectors, physiological indicator vectors, and rehabilitation scale vectors. Denoising involves weakening or zeroing out details below a preset threshold in each data vector while maintaining the original feature structure, thereby enhancing the stability of key information features. This ultimately yields a denoised multi-source rehabilitation dataset after bone tumor surgery.

[0017] S14. Time-align the denoised multi-source rehabilitation dataset after bone tumor surgery according to a unified reference time axis. Time alignment refers to generating a standard reference time axis based on a preset time start point, fixed interval, and time end point, and aligning the originally unevenly distributed data on the time axis. During the time alignment process, a spline interpolation function is used to numerically interpolate each dimensional medical image vector, physiological indicator vector, and rehabilitation scale vector at the reference time point to construct a time-consistent multi-source rehabilitation dataset after bone tumor surgery, thereby obtaining a time-aligned multi-source rehabilitation dataset after bone tumor surgery.

[0018] S15. Perform missing value filling operations on the missing elements in the time-series aligned multi-source rehabilitation dataset after bone tumor surgery. Use the k-nearest neighbor mean filling strategy to handle missing values. The k-nearest neighbor mean filling strategy refers to finding the nearest k samples for each data sample containing missing elements in the feature space, and using the average value of these k samples in the corresponding missing dimension to fill the position, thus generating a complete and non-missing preprocessed multi-source rehabilitation dataset D after bone tumor surgery. pre .

[0019] Optionally, the S2 includes the following steps:

[0020] S21. Extract each piece of postoperative bone tumor multi-source rehabilitation data from the preprocessed postoperative bone tumor multi-source rehabilitation dataset. Based on the adaptive frequency allocation factor, divide the medical image vector, physiological indicator vector, and rehabilitation scale vector in each piece of postoperative bone tumor multi-source rehabilitation data into frequency dimensions. Perform frequency domain division on each set of vectors on the feature channel according to the adaptive frequency allocation factor to obtain high-frequency and low-frequency parts of the corresponding vector. The high-frequency part is used to represent information above the fluctuation threshold or local change contained in the postoperative bone tumor rehabilitation data, and the low-frequency part is used to represent information below the change threshold or overall trend contained in the postoperative bone tumor rehabilitation data. The adaptive frequency allocation factor controls the proportion of the high-frequency dimension after each vector division, and the remaining dimensions are divided into the low-frequency part.

[0021] S22. The high-frequency parts obtained after frequency domain division of each postoperative multi-source rehabilitation data of bone tumor surgery are aggregated to form a high-frequency feature set of multi-source rehabilitation after bone tumor surgery. The low-frequency parts obtained after frequency domain division of each postoperative multi-source rehabilitation data of bone tumor surgery are aggregated to form a low-frequency feature set of multi-source rehabilitation after bone tumor surgery. The high-frequency feature set of multi-source rehabilitation after bone tumor surgery is used to reflect the high-frequency detail change pattern reflected in the postoperative rehabilitation process of bone tumor surgery. The low-frequency feature set of multi-source rehabilitation after bone tumor surgery is used to reflect the change law of the overall trend evolution in the postoperative rehabilitation process of bone tumor surgery.

[0022] Optionally, S3 includes the following steps:

[0023] S31. Obtaining high-frequency feature set D for multi-source rehabilitation after bone tumor surgery high and low-frequency feature set D of multi-source rehabilitation after bone tumor surgery low , and construct the initial input feature tensor pair

[0024]

[0025] Where b is the batch size, c is H 、c L are the number of high-frequency and low-frequency channels, respectively, and h and w are the spatial dimensions;

[0026] S32. Calculate a lesion region significance weight map, which is used to represent the importance distribution of the postoperative lesion region in the bone tumor postoperative medical image. The lesion region significance weight map is obtained by inputting the lesion mask tensor in the bone tumor postoperative medical image as input to the lesion attention feature extraction network to extract the lesion semantic representation, and then normalizing it through the normalization function Softmax;

[0027] S33. Execute an adaptive frequency redistribution operation, weighting the high-frequency input tensor and the low-frequency input tensor of bone tumor postoperative rehabilitation with the lesion region significance weight map element by element, and introduce the high-frequency adjustment factor and the low-frequency adjustment factor generated by the learnable adjustment factor network. The high-frequency adjustment factor is used to control the dynamic enhancement ratio of low-frequency information to the high-frequency path, and the low-frequency adjustment factor is used to control the dynamic enhancement ratio of high-frequency information to the low-frequency path. The element-by-element weighted results are respectively used as the high-frequency input tensor and the low-frequency input tensor after lesion guidance:

[0028]

[0029] Among them, α(·), β(·) are the high-frequency adjustment factors and low-frequency adjustment factors generated by the learnable adjustment factor network, and ⊙ is the element-by-element multiplication;

[0030] S34. Construct a channel decoupled convolution structure in the first improved Octave convolution layer, and input the high-frequency input tensor after lesion guidance and the low-frequency input tensor after lesion guidance into the channel decoupled convolution structure for spatial-aware convolution operation. The channel decoupled convolution structure performs spatial convolution operation independently in each channel to optimize the response ability of each channel to local structural changes. The results of the spatial-aware convolution operation generate high-frequency convolution output tensors respectively. Convolution with low frequency output tensor The convolution output tensor preserves the input spatial structure and has lesion sensitivity;

[0031] S35. Perform cross-path residual update on the high-frequency convolution output tensor and the low-frequency convolution output tensor, use the low-frequency convolution output tensor to extract channel perception information through the global average pooling function, and then input it into the Sigmoid activation function to generate a high-frequency complementary adjustment factor At the same time, the high-frequency convolution output tensor is used to generate the low-frequency complementary adjustment factor The high-frequency complementary adjustment factor is used to adjust the influence of the high-frequency input tensor from the previous layer on the current high-frequency output. The low-frequency complementary adjustment factor is used to adjust the influence of the low-frequency input tensor from the previous layer on the current low-frequency output. The cross-path residual update operation adds the adjusted high-frequency input tensor of the previous layer to the current high-frequency output tensor, and adds the adjusted low-frequency input tensor of the previous layer to the current low-frequency output tensor to form a high-frequency residual update tensor. Update the tensor with the low-frequency residual

[0032]

[0033] S36. Concatenate the high-frequency residual update tensor and the low-frequency residual update tensor in the channel dimension:

[0034]

[0035] And iteratively execute steps S32 to S36 until the Lth layer, and finally obtain the fused rehabilitation feature tensor The fused rehabilitation feature tensor contains the multi-scale dynamic response features of multi-source rehabilitation data after bone tumor surgery and the lesion area guidance features.

[0036] Optionally, the S4 includes the following steps:

[0037] S41. Fusion recovery feature tensor F fusionThe data is input into the bone tumor postoperative rehabilitation trend discrimination model, which is used to output the patient's rehabilitation trend prediction value. The bone tumor postoperative rehabilitation trend prediction value is generated by the fusion rehabilitation feature tensor through the multi-layer fully connected network, activation function and output mapping function in the bone tumor postoperative rehabilitation trend discrimination model. All trainable parameters of the multi-layer fully connected network, activation function and output mapping function constitute the parameter set θ;

[0038] S42. Construct a hybrid seagull search optimization module for trend perception in postoperative bone tumor rehabilitation. Use each parameter vector in the parameter set θ to construct a search individual and associate a trend feedback vector with each search individual. The trend feedback vector consists of the current predicted trend value, the true trend label, and the difference between the two. The trend feedback vector serves as a feedback signal for global direction adjustment during the optimization process.

[0039] S43. Construct a trend-weighted fitness function based on the trend feedback vector of each search individual. The trend-weighted fitness function consists of three parts: a first-order measure of the trend error value, a normalized value of the loss function value, and the time consumed by model training. The trend-weighted fitness function is obtained by weighted combination of these three parts. The trend-weighted fitness function is used to measure the degree of adaptation of the search individual to the task of identifying the postoperative rehabilitation trend of bone tumors.

[0040] S44. Adjust the search behavior of the hybrid seagull search optimization module based on the stage ratio of the current optimization process. The stage ratio is the current number of iterations divided by the maximum number of iterations. When the stage ratio is less than 0.3, the global search capability is enhanced; when the stage ratio is between 0.3 and 0.7, the global search capability and the local fine-tuning capability are balanced; when the stage ratio is greater than 0.7, the local convergence capability is enhanced;

[0041] S45. Execute a trend-bias-aware hybrid seagull search update strategy for each search individual. The hybrid seagull search update strategy consists of two parts: a periodic perturbation update based on the optimal individual direction, used to perform global search; and a fine-grained adjustment of individual parameters based on the perturbation intensity of the current stage ratio to achieve local convergence. The final individual update result is a weighted synthesis of the two parts.

[0042] S46. When it is detected that the validation error of an individual continues to decrease over multiple iterations but the training error remains unchanged, the individual is judged to have fallen into a local overfitting trap. At this time, the local perturbation jump mechanism is triggered. The local perturbation jump mechanism adds a slight Gaussian perturbation to the current parameters to achieve a jump out of the local optimal area.

[0043] S47. Repeat the trend-weighted fitness function construction, stage-by-stage weight adjustment, trend deviation perception update, and local disturbance jump mechanism until the termination condition is met. The termination condition includes that the current number of iterations reaches the maximum number of iterations, or the change in the optimal individual between two consecutive rounds is less than a preset error threshold. The optimization process stops when any of the termination conditions is met.

[0044] S48. Select the individual with the largest fitness value from all search individuals as the final optimization result. The parameter set corresponding to the final optimization result is the optimized parameter set θ. * , the optimized parameter set replaces the original parameter set in the bone tumor postoperative rehabilitation trend discrimination model to obtain the final bone tumor postoperative rehabilitation trend discrimination model.

[0045] Optionally, the S6 includes the following steps:

[0046] S61. Input the preprocessed multi-source rehabilitation dataset of the patient undergoing bone tumor surgery to the trained improved Octave convolutional network to generate a fused rehabilitation feature representation. This fused rehabilitation feature representation is then input into the trained bone tumor surgery rehabilitation trend discrimination model to obtain a series of rehabilitation trend prediction values continuously output by the model. This series of rehabilitation trend prediction values represents the changes in the patient's rehabilitation status at consecutive time points and is used to construct a rehabilitation trend curve.

[0047] S62. Calculate the rehabilitation trend slope and volatility index based on the rehabilitation trend prediction value sequence. The rehabilitation trend slope is used to indicate the speed and direction of rehabilitation progress, and the volatility index is used to measure the stability of the rehabilitation process. Generate a rehabilitation progress grade based on the final rehabilitation trend prediction value, trend slope, and volatility index.

[0048] S63. Based on the rehabilitation trend prediction value sequence and the patient's historical rehabilitation assessment reference threshold, calculate the probability of rehabilitation failure in the future prediction time period and output the prognostic risk level.

[0049] Optionally, the rehabilitation progress levels are divided according to the following rules:

[0050] Rehabilitation progress level I is the rapid recovery stage: the slope of the rehabilitation trend is greater than 0.5 and the volatility index is less than 0.2;

[0051] Rehabilitation progress level II is the stable rehabilitation stage: the slope of the rehabilitation trend is 0.20.5 and the volatility index is 0.20.4;

[0052] Rehabilitation progress level III is the slow recovery stage: the slope of the rehabilitation trend is ≤0.2 or the volatility index is >0.4.

[0053] Optionally, the prognostic risk level is divided according to the following rules:

[0054] Prognostic risk level A is a low-risk stage: all predicted values in the next 7 days are higher than the stable threshold;

[0055] Prognostic risk level B is the medium-risk stage: the predicted value will be lower than the stable threshold 1-2 times in the next 7 days;

[0056] Prognostic risk level C is a high-risk stage: the predicted value in the next 7 days will be lower than the stable threshold for 3 or more times in a row.

[0057] A deep learning-based bone tumor rehabilitation process evaluation system is used to implement a deep learning-based bone tumor rehabilitation process evaluation method, including the following modules:

[0058] Data acquisition module, used to collect original data sets of multi-source rehabilitation after bone tumor surgery;

[0059] The data preprocessing module is used to perform standardization, denoising, time series alignment, and missing value filling on the original multi-source rehabilitation data set after bone tumor surgery to generate a preprocessed multi-source rehabilitation data set after bone tumor surgery;

[0060] The feature extraction module includes a frequency domain partitioning submodule and an improved Octave convolution feature extraction submodule. The frequency domain partitioning submodule is used to divide the preprocessed rehabilitation dataset into high-frequency features and low-frequency features according to the adaptive frequency allocation factor. The improved Octave convolution feature extraction submodule is used to perform lesion-guided decoupled convolution operations, cross-channel residual fusion, and dynamic complementary updates on the high-frequency features and low-frequency features, respectively, to ultimately generate a fused rehabilitation feature representation.

[0061] The trend discrimination model construction module is used to input the fusion rehabilitation feature representation into the initial bone tumor postoperative rehabilitation trend discrimination model to complete the structure construction;

[0062] The model parameter optimization module includes a hybrid seagull search optimization submodule. This submodule iteratively optimizes the model parameters based on the trend feedback vector and the multi-objective fitness function, combining a phased search adjustment mechanism with a local perturbation mechanism to obtain the final optimal parameter set and complete the training of the bone tumor postoperative rehabilitation trend discrimination model.

[0063] The trend result output module is used to receive the output results of the trained bone tumor postoperative rehabilitation trend discrimination model, and combine the trend prediction value sequence to generate a rehabilitation trend curve, rehabilitation progress level and prognostic risk level for bone tumor rehabilitation process evaluation.

[0064] The beneficial effects of the present invention are:

[0065] (1) The present invention divides the pre-processed multi-source postoperative bone tumor rehabilitation dataset into high-frequency and low-frequency domains, introduces a lesion attention mechanism to guide the Octave convolution channel, so that the high-frequency path focuses on local structural changes and the low-frequency path captures global trend information. Through channel decoupling, complementary residual update and cross-channel fusion, the present invention realizes the joint modeling of fine-grained changes and trend evolution in postoperative rehabilitation data. Compared with the traditional single-channel convolution, the frequency feature fusion strength can be dynamically adjusted, and higher discrimination accuracy is shown in the joint representation of lesion area changes and postoperative systemic state fluctuations.

[0066] (2) The present invention embeds model prediction error, trend deviation and training efficiency into the fitness function design, and combines trend deviation feedback to guide the search direction, introduces staged behavior regulation and local disturbance jump mechanism, and effectively solves the problems of slow convergence and easy to fall into local optimality of traditional optimization algorithms in high-dimensional space, and realizes the trend-guided dynamic evolution path in discriminant model training.

[0067] (3) In the model prediction output stage, the present invention constructs a three-in-one prediction result structure of rehabilitation trend curve, rehabilitation progress level and prognosis risk level, and interprets the rehabilitation path in segments through trend slope and volatility indicators, and performs graded risk identification in combination with changes in predicted values in future time periods. It can intuitively reflect the changes in the patient's rehabilitation rhythm and risk trends, and provide doctors with a data-driven basis for staged management. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0069] Figure 1 This is a flowchart of a deep learning-based bone tumor rehabilitation process evaluation method proposed in the present invention. DETAILED DESCRIPTION

[0070] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams that illustrate the basic structure of the present invention in a schematic manner.

[0071] refer to Figure 1 , a bone tumor rehabilitation process evaluation method based on deep learning, including the following steps:

[0072] S1. Collect multi-source rehabilitation data after bone tumor surgery, including medical imaging data, continuous physiological indicator sequence data, and rehabilitation assessment scale data, and uniformly encode them to construct a multi-source rehabilitation dataset after bone tumor surgery. Then, perform standardization, denoising, time series alignment, and missing data completion on the multi-source rehabilitation dataset after bone tumor surgery to obtain the pre-processed multi-source rehabilitation dataset after bone tumor surgery.

[0073] S2. Perform frequency domain division on the preprocessed multi-source rehabilitation dataset after bone tumor surgery based on the adaptive frequency allocation factor to generate a high-frequency rehabilitation feature set and a low-frequency rehabilitation feature set;

[0074] S3. Input the high-frequency rehabilitation feature set into the high-frequency improved Octave convolution channel, and input the low-frequency rehabilitation feature set into the low-frequency improved Octave convolution channel. Perform convolution operations on each channel while keeping the channels independent. Set cross-layer residual connections between the convolution layers of the improved Octave convolution network, and perform bidirectional compensation fusion on the convolution outputs of the high-frequency rehabilitation feature set and the low-frequency rehabilitation feature set to form a fused rehabilitation feature representation.

[0075] S4. Input the fused rehabilitation feature representation into the initial bone tumor postoperative rehabilitation trend discrimination model to complete the structural construction of the bone tumor postoperative rehabilitation trend discrimination model. Initialize the hybrid seagull search optimization module, use the bone tumor postoperative rehabilitation trend discrimination model parameters as search individuals, combine the global exploration strategy with the local perturbation strategy, establish a hybrid seagull search population, construct a multi-objective fitness function based on accuracy, verification loss, and training time, perform iterative optimization on the hybrid seagull search population, and obtain the optimized model parameters.

[0076] S5. Use the optimized model parameters to replace the corresponding parameters in the bone tumor postoperative rehabilitation trend discrimination model, perform supervised training on the preprocessed bone tumor postoperative multi-source rehabilitation dataset, and output the trained bone tumor postoperative rehabilitation trend discrimination model;

[0077] S6. Apply the trained bone tumor postoperative rehabilitation trend discrimination model to the fused rehabilitation feature representation of the patient to be discriminated, and generate a rehabilitation trend prediction result consisting of rehabilitation progress level, rehabilitation trend curve and prognostic risk level.

[0078] In this embodiment, S1 includes the following steps:

[0079] S11. Collect raw multi-source rehabilitation data after bone tumor surgery and set a collection time range to construct a raw multi-source rehabilitation data set after bone tumor surgery. Each piece of raw multi-source rehabilitation data after bone tumor surgery includes a timestamp, a medical image vector, a physiological indicator vector, and a rehabilitation scale vector. The medical image vector represents the patient's corresponding medical image feature information at the time of collection, the physiological indicator vector represents the patient's continuous physiological state measurement value at that time, and the rehabilitation scale vector represents the corresponding assessment questionnaire score information.

[0080] S12. Standardize each data item in the original multi-source rehabilitation data set after bone tumor surgery. Standardization involves mapping the value of each dimension in the original data to a uniform numerical range based on its minimum and maximum values in that dimension through a linear normalization transformation, thereby obtaining a standardized multi-source rehabilitation data set after bone tumor surgery.

[0081] S13. Perform denoising on the standardized multi-source rehabilitation dataset after bone tumor surgery. A wavelet threshold denoising strategy is used to process the various dimensional elements in the medical image vectors, physiological indicator vectors, and rehabilitation scale vectors. Denoising involves weakening or zeroing out details below a preset threshold in each data vector while maintaining the original feature structure, thereby enhancing the stability of key information features. This ultimately yields a denoised multi-source rehabilitation dataset after bone tumor surgery.

[0082] S14. Time-align the denoised multi-source rehabilitation dataset after bone tumor surgery according to a unified reference time axis. Time alignment refers to generating a standard reference time axis based on a preset time start point, fixed interval, and time end point, and aligning the originally unevenly distributed data on the time axis. During the time alignment process, a spline interpolation function is used to numerically interpolate each dimensional medical image vector, physiological indicator vector, and rehabilitation scale vector at the reference time point to construct a time-consistent multi-source rehabilitation dataset after bone tumor surgery, thereby obtaining a time-aligned multi-source rehabilitation dataset after bone tumor surgery.

[0083] S15. Perform missing value filling operations on the missing elements in the time-series aligned multi-source rehabilitation dataset after bone tumor surgery. Use the k-nearest neighbor mean filling strategy to handle missing values. The k-nearest neighbor mean filling strategy refers to finding the nearest k samples for each data sample containing missing elements in the feature space, and using the average value of these k samples in the corresponding missing dimension to fill the position, thus generating a complete and non-missing preprocessed multi-source rehabilitation dataset D after bone tumor surgery. pre .

[0084] In this embodiment, S2 includes the following steps:

[0085] S21. Extract each piece of postoperative bone tumor multi-source rehabilitation data from the preprocessed postoperative bone tumor multi-source rehabilitation dataset. Based on the adaptive frequency allocation factor, divide the medical image vector, physiological indicator vector, and rehabilitation scale vector in each piece of postoperative bone tumor multi-source rehabilitation data into frequency dimensions. Perform frequency domain division on each set of vectors on the feature channel according to the adaptive frequency allocation factor to obtain high-frequency and low-frequency parts of the corresponding vector. The high-frequency part is used to represent information above the fluctuation threshold or local change contained in the postoperative bone tumor rehabilitation data, and the low-frequency part is used to represent information below the change threshold or overall trend contained in the postoperative bone tumor rehabilitation data. The adaptive frequency allocation factor controls the proportion of the high-frequency dimension after each vector division, and the remaining dimensions are divided into the low-frequency part.

[0086] S22. The high-frequency parts obtained after frequency domain division of each postoperative multi-source rehabilitation data of bone tumor surgery are aggregated to form a high-frequency feature set of multi-source rehabilitation after bone tumor surgery. The low-frequency parts obtained after frequency domain division of each postoperative multi-source rehabilitation data of bone tumor surgery are aggregated to form a low-frequency feature set of multi-source rehabilitation after bone tumor surgery. The high-frequency feature set of multi-source rehabilitation after bone tumor surgery is used to reflect the high-frequency detail change pattern reflected in the postoperative rehabilitation process of bone tumor surgery. The low-frequency feature set of multi-source rehabilitation after bone tumor surgery is used to reflect the change law of the overall trend evolution in the postoperative rehabilitation process of bone tumor surgery.

[0087] In this embodiment, S3 includes the following steps:

[0088] S31. Obtaining high-frequency feature set D for multi-source rehabilitation after bone tumor surgery high and low-frequency feature set D of multi-source rehabilitation after bone tumor surgery low , and construct the initial input feature tensor pair

[0089]

[0090] Where b is the batch size, c H 、c L are the number of high-frequency and low-frequency channels, respectively, and h and w are the spatial dimensions;

[0091] S32. Calculate a lesion region significance weight map, which is used to represent the importance distribution of the postoperative lesion region in the bone tumor postoperative medical image. The lesion region significance weight map is obtained by inputting the lesion mask tensor in the bone tumor postoperative medical image as input to the lesion attention feature extraction network to extract the lesion semantic representation, and then normalizing it through the normalization function Softmax;

[0092] S33. Execute an adaptive frequency redistribution operation, weighting the high-frequency input tensor and the low-frequency input tensor of bone tumor postoperative rehabilitation with the lesion region significance weight map element by element, and introduce the high-frequency adjustment factor and the low-frequency adjustment factor generated by the learnable adjustment factor network. The high-frequency adjustment factor is used to control the dynamic enhancement ratio of low-frequency information to the high-frequency path, and the low-frequency adjustment factor is used to control the dynamic enhancement ratio of high-frequency information to the low-frequency path. The element-by-element weighted results are respectively used as the high-frequency input tensor and the low-frequency input tensor after lesion guidance:

[0093]

[0094] Among them, α(·), β(·) are the high-frequency adjustment factors and low-frequency adjustment factors generated by the learnable adjustment factor network, and ⊙ is the element-by-element multiplication;

[0095] S34. Construct a channel decoupled convolution structure in the first improved Octave convolution layer, and input the high-frequency input tensor after lesion guidance and the low-frequency input tensor after lesion guidance into the channel decoupled convolution structure for spatial-aware convolution operation. The channel decoupled convolution structure performs spatial convolution operation independently in each channel to optimize the response ability of each channel to local structural changes. The results of the spatial-aware convolution operation generate high-frequency convolution output tensors respectively. Convolution with low frequency output tensor The convolution output tensor preserves the input spatial structure and has lesion sensitivity;

[0096] S35. Perform cross-path residual update on the high-frequency convolution output tensor and the low-frequency convolution output tensor, use the low-frequency convolution output tensor to extract channel perception information through the global average pooling function, and then input it into the Sigmoid activation function to generate a high-frequency complementary adjustment factor At the same time, the high-frequency convolution output tensor is used to generate the low-frequency complementary adjustment factor The high-frequency complementary adjustment factor is used to adjust the influence of the high-frequency input tensor from the previous layer on the current high-frequency output. The low-frequency complementary adjustment factor is used to adjust the influence of the low-frequency input tensor from the previous layer on the current low-frequency output. The cross-path residual update operation adds the adjusted high-frequency input tensor of the previous layer to the current high-frequency output tensor, and adds the adjusted low-frequency input tensor of the previous layer to the current low-frequency output tensor to form a high-frequency residual update tensor. Update the tensor with the low-frequency residual

[0097]

[0098] S36. Concatenate the high-frequency residual update tensor and the low-frequency residual update tensor in the channel dimension:

[0099]

[0100] And iteratively execute steps S32 to S36 until the Lth layer, and finally obtain the fused rehabilitation feature tensor The fused rehabilitation feature tensor contains the multi-scale dynamic response features of multi-source rehabilitation data after bone tumor surgery and the lesion area guidance features.

[0101] In this embodiment, S4 includes the following steps:

[0102] S41. Fusion recovery feature tensor F fusion The data is input into the bone tumor postoperative rehabilitation trend discrimination model, which is used to output the patient's rehabilitation trend prediction value. The bone tumor postoperative rehabilitation trend prediction value is generated by the fusion rehabilitation feature tensor through the multi-layer fully connected network, activation function and output mapping function in the bone tumor postoperative rehabilitation trend discrimination model. All trainable parameters of the multi-layer fully connected network, activation function and output mapping function constitute the parameter set θ;

[0103] The postoperative rehabilitation trend discrimination model for bone tumors receives the fusion rehabilitation feature tensor F fusion As input, the fused rehabilitation feature tensor is a structured two-dimensional tensor representing the high- and low-frequency lesion-guided features extracted in one or more time windows for each patient, with a size of (b×c f ), where b represents the number of sample batches, c f Indicates the total number of dimensions of the fused features.

[0104] The postoperative rehabilitation trend discrimination model for bone tumors adopts a multi-layer fully connected neural network structure, which performs linear transformation, nonlinear activation and output mapping operations on the fused rehabilitation feature tensor in sequence, and finally generates a continuous rehabilitation trend prediction value representing the patient's postoperative rehabilitation status. The rehabilitation trend prediction value can be understood as the rehabilitation status score or progress index estimated by the model under given feature input.

[0105] The multi-layer fully connected network consists of several linear transformation layers. Each layer contains a set of weight matrices and bias vectors, which are used to perform affine transformations on the input feature vectors. The trainable parameters of the l-th layer of the fully connected network include:

[0106] Weight Matrix Indicates the input dimension d l-1 To the output dimension d l The mapping relationship;

[0107] Bias vector Used to add translation bias to linear output;

[0108] The above two items together constitute the subset of trainable parameters θ of layer l (l) ={W (l),b (l)}.

[0109] After the output of each fully connected layer, a nonlinear activation function is connected. The activation function uses ReLU or PReLU. This activation function performs a nonlinear transformation on the linear output to increase the model's ability to express complex relationships. If the PReLU activation function is used, each layer contains an additional learnable slope parameter a. (l) , which serves as an additional trainable parameter of the activation function of this layer.

[0110] The output mapping function is configured at the end of the model to map the last hidden state to the target prediction scalar. The output mapping function is a single neuron output layer with a linear activation function. The trainable parameters of the output layer include the ultimate weight vector With bias scalar where d L Represents the output dimension of the last hidden layer.

[0111] In summary, all trainable parameters of the bone tumor postoperative recovery trend discrimination model include:

[0112] The weight matrix and bias vector set of all fully connected layers {W (l) ,b( l)};

[0113] The set of learnable parameters contained in all activation functions {a (l)}(if using PReLU);

[0114] The weight vector and bias scalar {W (out) ,b (out)}.

[0115] S42. Construct a hybrid seagull search optimization module for trend perception in postoperative bone tumor rehabilitation. Use each parameter vector in the parameter set θ to construct a search individual and associate a trend feedback vector with each search individual. The trend feedback vector consists of the current predicted trend value, the true trend label, and the difference between the two. The trend feedback vector serves as a feedback signal for global direction adjustment during the optimization process.

[0116] S43. Construct a trend-weighted fitness function based on the trend feedback vector of each search individual. The trend-weighted fitness function consists of three parts: a first-order measure of the trend error value, a normalized value of the loss function value, and the time consumed by model training. The trend-weighted fitness function is obtained by weighted combination of these three parts. The trend-weighted fitness function is used to measure the degree of adaptation of the search individual to the task of identifying the postoperative rehabilitation trend of bone tumors.

[0117] The first-order measure of the trend error value refers to the calculation of the first-order norm of the error between the predicted value of the postoperative rehabilitation trend of bone tumors predicted by the current model and the true trend label. The first-order measure of the trend error value is obtained in the following way: for each validation sample, the difference between the predicted value and the true value is calculated, and the L2 norm of the difference vector is taken as the trend error value of the sample. The mean of the trend error values of all samples is taken as the trend deviation of the search individual. The trend deviation is used to measure the accuracy of the predicted trend. The parameter is

[0118] The loss function value refers to the square error evaluation of the predicted value and the true value of the current search individual on the validation set using the mean square error loss function. The prediction errors of all validation samples are averaged to form the loss function value. The loss function value is used to measure the regression fitting ability of the search individual in the prediction task, which is recorded as

[0119] The normalized value of the model training time refers to the difference between the actual time consumed by the current search individual to complete the training and verification calculations in the tth round of iteration and the preset maximum training allowable time T max The ratio between , which is used to quantify the training efficiency of the model, is expressed as The smaller the value, the faster the convergence speed of the search individual corresponding model and the higher the resource efficiency.

[0120] S44. Adjust the search behavior of the hybrid seagull search optimization module based on the stage ratio of the current optimization process. The stage ratio is the current number of iterations divided by the maximum number of iterations. When the stage ratio is less than 0.3, the global search capability is enhanced; when the stage ratio is between 0.3 and 0.7, the global search capability and the local fine-tuning capability are balanced; when the stage ratio is greater than 0.7, the local convergence capability is enhanced;

[0121] S45. Execute a trend-bias-aware hybrid seagull search update strategy for each search individual. The hybrid seagull search update strategy consists of two parts: a periodic perturbation update based on the optimal individual direction, used to perform global search; and a fine-grained adjustment of individual parameters based on the perturbation intensity of the current stage ratio to achieve local convergence. The final individual update result is a weighted synthesis of the two parts.

[0122] S46. When it is detected that the validation error of an individual continues to decrease over multiple iterations but the training error remains unchanged, the individual is judged to have fallen into a local overfitting trap. At this time, the local perturbation jump mechanism is triggered. The local perturbation jump mechanism adds a slight Gaussian perturbation to the current parameters to achieve a jump out of the local optimal area.

[0123] S47. Repeat the trend-weighted fitness function construction, stage-by-stage weight adjustment, trend deviation perception update, and local disturbance jump mechanism until the termination condition is met. The termination condition includes that the current number of iterations reaches the maximum number of iterations, or the change in the optimal individual between two consecutive rounds is less than a preset error threshold. The optimization process stops when any of the termination conditions is met.

[0124] S48. Select the individual with the largest fitness value from all search individuals as the final optimization result. The parameter set corresponding to the final optimization result is the optimized parameter set θ. * , the optimized parameter set replaces the original parameter set in the bone tumor postoperative rehabilitation trend discrimination model to obtain the final bone tumor postoperative rehabilitation trend discrimination model.

[0125] In this embodiment, S6 includes the following steps:

[0126] S61. Input the preprocessed multi-source rehabilitation dataset of the patient undergoing bone tumor surgery to the trained improved Octave convolutional network to generate a fused rehabilitation feature representation. This fused rehabilitation feature representation is then input into the trained bone tumor surgery rehabilitation trend discrimination model to obtain a series of rehabilitation trend prediction values continuously output by the model. This series of rehabilitation trend prediction values represents the changes in the patient's rehabilitation status at consecutive time points and is used to construct a rehabilitation trend curve.

[0127] S62. Calculate the rehabilitation trend slope and volatility index based on the rehabilitation trend prediction value sequence. The rehabilitation trend slope is used to indicate the speed and direction of rehabilitation progress, and the volatility index is used to measure the stability of the rehabilitation process. Generate a rehabilitation progress grade based on the final rehabilitation trend prediction value, trend slope, and volatility index.

[0128] S63. Based on the rehabilitation trend prediction value sequence and the patient's historical rehabilitation assessment reference threshold, calculate the probability of rehabilitation failure in the future prediction time period and output the prognostic risk level.

[0129] In this implementation, the rehabilitation progress levels are divided according to the following rules:

[0130] Rehabilitation progress level I is the rapid recovery stage: the slope of the rehabilitation trend is greater than 0.5 and the volatility index is less than 0.2;

[0131] Rehabilitation progress level II is the stable rehabilitation stage: the slope of the rehabilitation trend is 0.20.5 and the volatility index is 0.20.4;

[0132] Rehabilitation progress level III is the slow recovery stage: the slope of the rehabilitation trend is ≤0.2 or the volatility index is >0.4.

[0133] In this embodiment, the prognostic risk level is divided according to the following rules:

[0134] Prognostic risk level A is a low-risk stage: all predicted values in the next 7 days are higher than the stable threshold;

[0135] Prognostic risk level B is the medium-risk stage: the predicted value will be lower than the stable threshold 1-2 times in the next 7 days;

[0136] Prognostic risk level C is a high-risk stage: the predicted value in the next 7 days will be lower than the stable threshold for 3 or more times in a row.

[0137] A deep learning-based bone tumor rehabilitation process evaluation system is used to implement a deep learning-based bone tumor rehabilitation process evaluation method, including the following modules:

[0138] Data acquisition module, used to collect original data sets of multi-source rehabilitation after bone tumor surgery;

[0139] The data preprocessing module is used to perform standardization, denoising, time series alignment, and missing value filling on the original multi-source rehabilitation data set after bone tumor surgery to generate a preprocessed multi-source rehabilitation data set after bone tumor surgery;

[0140] The feature extraction module includes a frequency domain partitioning submodule and an improved Octave convolution feature extraction submodule. The frequency domain partitioning submodule is used to divide the preprocessed rehabilitation dataset into high-frequency features and low-frequency features according to the adaptive frequency allocation factor. The improved Octave convolution feature extraction submodule is used to perform lesion-guided decoupled convolution operations, cross-channel residual fusion, and dynamic complementary updates on the high-frequency features and low-frequency features, respectively, to ultimately generate a fused rehabilitation feature representation.

[0141] The trend discrimination model construction module is used to input the fusion rehabilitation feature representation into the initial bone tumor postoperative rehabilitation trend discrimination model to complete the structure construction;

[0142] The model parameter optimization module includes a hybrid seagull search optimization submodule. This submodule iteratively optimizes the model parameters based on the trend feedback vector and the multi-objective fitness function, combining a phased search adjustment mechanism with a local perturbation mechanism to obtain the final optimal parameter set and complete the training of the bone tumor postoperative rehabilitation trend discrimination model.

[0143] The trend result output module is used to receive the output results of the trained bone tumor postoperative rehabilitation trend discrimination model, and combine the trend prediction value sequence to generate a rehabilitation trend curve, rehabilitation progress level and prognostic risk level for bone tumor rehabilitation process evaluation.

[0144] Example 1:

[0145] The Orthopedics Center at Hospital A admitted a 52-year-old male patient, surnamed Liu, 42 days after surgery. His recovery after right distal femoral osteosarcoma resection was uncertain, and the hospital needed to assess whether there was a risk of delayed recovery over the next three months. The hospital enrolled the patient in the system presented in this invention for trend prediction.

[0146] The patient initially submitted rehabilitation data covering five postoperative time points, from day 7 to day 35, during the initial system access phase. The uploaded data included five MRI sequences, continuous gait sensor recordings at corresponding time points (including 28 indicators, including cadence, left / right lower limb angle difference, and tremor coefficient), and WOMAC scores and pain visual analog scale data manually recorded by the attending physician.

[0147] The system first performed data normalization: Cadence data at time point 1 was missing, so the system filled in the missing values by referencing similar structures at time points 2, 3, and 4 based on a k-nearest-neighbor mean strategy. Subsequently, the system performed spline alignment on all data based on a unified reference timeline, ensuring strict synchronization of different data types along the postoperative time dimension.

[0148] The system completed the frequency domain division operation: the system detected that the energy of Liu's tremor coefficient and gait asymmetry rate at time points 2 and 3 was abnormally increased in the high-frequency channel (1.9 times higher than the average level), and the system automatically classified it into the high-frequency rehabilitation feature set, while the slow averaging process of the bone marrow signal in the MRI at time points 4 and 5 was classified as low-frequency channel information.

[0149] The system then activated an improved Octave convolutional network to model high- and low-frequency features. In the second-layer channel decoupling convolution, the system introduced a lesion attention map: the semantic segmentation mask from the MRI indicated that there was still mild exudative signal at the right femoral resection edge, and the system increased the convolution response factor of the high-frequency channel in this area. In the third-layer fusion stage, the system automatically generated a fused rehabilitation feature tensor of size [1×256] and passed it to the trend discrimination module.

[0150] The system activated the Hybrid Seagull Search Optimization module, dynamically adjusting the 524,312 parameters of the initial trend discrimination model. During the 21st iteration, the system detected that a particular search individual performed well on the validation set while maintaining the same error on the training set. This automatically triggered the "local perturbation jump mechanism," injecting a Gaussian perturbation factor (with a standard deviation of 0.05) on top of the original parameters, ultimately achieving successful convergence on the 37th iteration.

[0151] The system outputs trend prediction results: the overall recovery trend curve presents an "up-plateau-slow decline" pattern; the rehabilitation progress level is assessed by the system as level III (slow recovery); the prognostic risk level is level C (high risk), and it is detected that there will be three consecutive trend decline prediction points below the threshold T in the next 7 days.s =0.63.

[0152] After generating the report, the system automatically sent an electronic alert to the account of the attending physician, Dr. Zhang. The report indicated that "the patient's recovery trend has stagnated, and there is a risk of functional deterioration in the surgical area." It also recommended reviewing the imaging, performing a pathological review, and adjusting the intervention plan. The doctor arranged for an MRI review that same day, which revealed an increase in the area of effusion in the surgical area. Bone window analysis initially suggested the possibility of microinfection. The doctor then adjusted the anti-inflammatory medication and rehabilitation intensity, implementing preventive interventions.

[0153] In order to compare and verify the value of the system, the hospital also used the traditional LSTM trend prediction model to predict Liu's data at the previous four time points. The system did not detect a downward trend in the output, and the prognosis level was determined to be B (medium risk), and no alarm signal was generated.

[0154] After 90 days of follow-up, Liu's MRI examination on the 120th day after surgery showed good soft tissue recovery, controlled inflammation, and muscle strength scores restored to 80% of their preoperative level. The attending physician stated, "Had the system not provided a warning of the trend over the next seven days, we might have overlooked the early stage of effusion, leading to even more severe functional deterioration."

[0155] This patient was included as part of a summary of systems evaluation, with the following findings:

[0156] Table 1 Comparison data between the present invention and traditional LSTM

[0157]

[0158] In addition, this system has treated 54 patients after bone tumor surgery and successfully identified 18 individuals with high-risk trends in advance, with an average advance time of 6.3 days, which is 2.1 days earlier than the traditional method. The actual intervention success rate has increased by about 19.7%.

[0159] This Example 1 fully demonstrates the effectiveness and clinical significance of the method of the present invention in the rehabilitation management of real bone tumor surgery. It not only realizes the deep fusion modeling of multi-source rehabilitation data, but also effectively avoids the shortcomings of low training efficiency and poor discrimination accuracy of traditional methods, providing data support and intelligent decision-making capabilities for complex individualized postoperative management.

[0160] The present invention divides the pre-processed multi-source postoperative bone tumor rehabilitation dataset into high-frequency and low-frequency domains, introduces a lesion attention mechanism to guide the Octave convolution channel, so that the high-frequency path focuses on local structural changes and the low-frequency path captures global trend information. Through channel decoupling, complementary residual update and cross-channel fusion, the joint modeling of fine-grained changes and trend evolution in postoperative rehabilitation data is achieved. Compared with traditional single-channel convolution, the frequency feature fusion strength can be dynamically adjusted, and higher discrimination accuracy is shown when processing the joint representation of lesion area changes and postoperative systemic state fluctuations.

[0161] The present invention embeds model prediction error, trend deviation and training efficiency into the fitness function design, and combines trend deviation feedback to guide the search direction, introduces staged behavior regulation and local disturbance jump mechanism, effectively solving the problems of slow convergence and easy falling into local optimality of traditional optimization algorithms in high-dimensional space, and realizes a trend-guided dynamic evolution path in discriminant model training.

[0162] In the model prediction output stage, the present invention constructs a three-in-one prediction result structure consisting of a rehabilitation trend curve, rehabilitation progress level, and prognostic risk level. It also interprets the rehabilitation path in segments through trend slope and volatility indicators, and performs graded risk identification based on changes in predicted values in future time periods. This can intuitively reflect changes in the patient's rehabilitation rhythm and risk trends, providing doctors with a data-driven basis for staged management.

[0163] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A bone tumor rehabilitation process evaluation method based on deep learning, characterized in that: The steps include: S1. Collect multi-source rehabilitation data after bone tumor surgery, construct a multi-source rehabilitation dataset after bone tumor surgery, and perform preprocessing to obtain a preprocessed multi-source rehabilitation dataset after bone tumor surgery; S2. Perform frequency domain division on the preprocessed multi-source rehabilitation dataset after bone tumor surgery based on the adaptive frequency allocation factor to generate a high-frequency rehabilitation feature set and a low-frequency rehabilitation feature set; S3. Input the high-frequency rehabilitation feature set into the high-frequency improved Octave convolution channel, and input the low-frequency rehabilitation feature set into the low-frequency improved Octave convolution channel to form a fused rehabilitation feature representation; S4. Input the fused rehabilitation feature representation into the initial bone tumor postoperative rehabilitation trend discrimination model, initialize the hybrid seagull search optimization module, use the bone tumor postoperative rehabilitation trend discrimination model parameters as search individuals, combine the global exploration strategy with the local perturbation strategy, establish a hybrid seagull search population, construct a multi-objective fitness function, perform iterative optimization on the hybrid seagull search population, and obtain the optimized model parameters; S5. Use the optimized model parameters to replace the corresponding parameters in the bone tumor postoperative rehabilitation trend discrimination model, perform supervised training on the preprocessed bone tumor postoperative multi-source rehabilitation dataset, and output the trained bone tumor postoperative rehabilitation trend discrimination model; S6. Apply the trained bone tumor postoperative rehabilitation trend discrimination model to the fused rehabilitation feature representation of the patient to be discriminated, and generate a rehabilitation trend prediction result consisting of rehabilitation progress level, rehabilitation trend curve and prognostic risk level.

2. A bone tumor rehabilitation process evaluation method based on deep learning according to claim 1, characterized in that: Said S1 comprises the following steps: S11. Collect raw multi-source rehabilitation data after bone tumor surgery and set a collection time range to construct a raw multi-source rehabilitation data set after bone tumor surgery. Each piece of raw multi-source rehabilitation data after bone tumor surgery includes a timestamp, a medical image vector, a physiological indicator vector, and a rehabilitation scale vector. The medical image vector represents the patient's corresponding medical image feature information at the time of collection, the physiological indicator vector represents the patient's continuous physiological state measurement value at that time, and the rehabilitation scale vector represents the corresponding assessment questionnaire score information. S12. Perform standardization, denoising, time alignment, and missing completion on the multi-source rehabilitation dataset after bone tumor surgery to obtain a preprocessed multi-source rehabilitation dataset after bone tumor surgery.

3. A bone tumor rehabilitation process evaluation method based on deep learning according to claim 2, characterized in that: The S2 comprises the following steps: S21. Extract each piece of postoperative bone tumor multi-source rehabilitation data from the preprocessed postoperative bone tumor multi-source rehabilitation dataset. Based on the adaptive frequency allocation factor, divide the medical image vector, physiological indicator vector, and rehabilitation scale vector in each piece of postoperative bone tumor multi-source rehabilitation data into frequency dimensions. Perform a frequency domain division operation on each set of vectors on the feature channel according to the adaptive frequency allocation factor to obtain the high-frequency and low-frequency components of the corresponding vector. The adaptive frequency allocation factor controls the proportion of the high-frequency dimension of each vector after division, and the remaining dimensions are divided into the low-frequency component. S22. The high-frequency parts obtained after frequency domain division of each postoperative multi-source rehabilitation data of bone tumor surgery are aggregated to form a high-frequency feature set of multi-source rehabilitation after bone tumor surgery; the low-frequency parts obtained after frequency domain division of each postoperative multi-source rehabilitation data of bone tumor surgery are aggregated to form a low-frequency feature set of multi-source rehabilitation after bone tumor surgery.

4. A bone tumor rehabilitation process evaluation method based on deep learning according to claim 3, characterized in that: The S3 includes the following steps: S31. Obtaining high-frequency feature set D for multi-source rehabilitation after bone tumor surgery high and low-frequency feature set D of multi-source rehabilitation after bone tumor surgery low , and construct the initial input feature tensor pair S32. Calculate a lesion region significance weight map, which is used to represent the importance distribution of the postoperative lesion region in the bone tumor postoperative medical image. The lesion region significance weight map is obtained by inputting the lesion mask tensor in the bone tumor postoperative medical image as input to the lesion attention feature extraction network to extract the lesion semantic representation, and then normalizing it through the normalization function Softmax; S33. Perform an adaptive frequency reallocation operation to weight the high-frequency input tensor and the low-frequency input tensor of bone tumor postoperative rehabilitation with the lesion region saliency weight map element by element, and introduce a high-frequency adjustment factor and a low-frequency adjustment factor generated by a learnable adjustment factor network. The high-frequency adjustment factor is used to control the dynamic enhancement ratio of low-frequency information to high-frequency paths, and the low-frequency adjustment factor is used to control the dynamic enhancement ratio of high-frequency information to low-frequency paths. The element-by-element weighted results are used as the high-frequency input tensor and the low-frequency input tensor after lesion guidance, respectively. S34. Construct a channel decoupled convolution structure in the first improved Octave convolution layer, and input the high-frequency input tensor after lesion guidance and the low-frequency input tensor after lesion guidance into the channel decoupled convolution structure for spatial-aware convolution operation. The channel decoupled convolution structure performs spatial convolution operation independently in each channel to optimize the response ability of each channel to local structural changes. The results of the spatial-aware convolution operation generate high-frequency convolution output tensors respectively. Convolution with low frequency output tensor S35. Perform cross-path residual update on the high-frequency convolution output tensor and the low-frequency convolution output tensor, use the low-frequency convolution output tensor to extract channel perception information through the global average pooling function, and then input it into the Sigmoid activation function to generate a high-frequency complementary adjustment factor At the same time, the high-frequency convolution output tensor is used to generate the low-frequency complementary adjustment factor The high-frequency complementary adjustment factor is used to adjust the influence of the high-frequency input tensor from the previous layer on the current high-frequency output. The low-frequency complementary adjustment factor is used to adjust the influence of the low-frequency input tensor from the previous layer on the current low-frequency output. The cross-path residual update operation adds the adjusted high-frequency input tensor of the previous layer to the current high-frequency output tensor, and adds the adjusted low-frequency input tensor of the previous layer to the current low-frequency output tensor to form a high-frequency residual update tensor. Update the tensor with the low-frequency residual S36. Concatenate the high-frequency residual update tensor and the low-frequency residual update tensor in the channel dimension, and iterate steps S32 to S36 until the Lth layer, and finally obtain the fused rehabilitation feature tensor The fused rehabilitation feature tensor contains the multi-scale dynamic response features of multi-source rehabilitation data after bone tumor surgery and the lesion area guidance features.

5. A bone tumor rehabilitation process evaluation method based on deep learning according to claim 4, characterized in that: The S4 comprises the following steps: S41. Fusion recovery feature tensor F fusion The data is input into the bone tumor postoperative rehabilitation trend discrimination model, which is used to output the patient's rehabilitation trend prediction value. The bone tumor postoperative rehabilitation trend prediction value is generated by the fusion rehabilitation feature tensor through the multi-layer fully connected network, activation function and output mapping function in the bone tumor postoperative rehabilitation trend discrimination model. All trainable parameters of the multi-layer fully connected network, activation function and output mapping function constitute the parameter set θ; S42. Construct a hybrid seagull search optimization module for trend perception in postoperative bone tumor rehabilitation. Use each parameter vector in the parameter set θ to construct a search individual and associate a trend feedback vector with each search individual. The trend feedback vector consists of the current predicted trend value, the true trend label, and the difference between the two. The trend feedback vector serves as a feedback signal for global direction adjustment during the optimization process. S43. Construct a trend-weighted fitness function based on the trend feedback vector of each search individual. The trend-weighted fitness function consists of three parts: a first-order measure of the trend error value, a normalized value of the loss function value, and the time consumed by model training. The trend-weighted fitness function is obtained by weighted combination of these three parts. The trend-weighted fitness function is used to measure the degree of adaptation of the search individual to the task of identifying the postoperative rehabilitation trend of bone tumors. S44. Adjust the search behavior of the hybrid seagull search optimization module based on the stage ratio of the current optimization process. The stage ratio is the current number of iterations divided by the maximum number of iterations. When the stage ratio is less than 0.3, the global search capability is enhanced; when the stage ratio is between 0.3 and 0.7, the global search capability and the local fine-tuning capability are balanced; when the stage ratio is greater than 0.7, the local convergence capability is enhanced; S45. Execute a trend-bias-aware hybrid seagull search update strategy for each search individual. The hybrid seagull search update strategy consists of two parts: a periodic perturbation update based on the optimal individual direction, used to perform global search; and a fine-grained adjustment of individual parameters based on the perturbation intensity of the current stage ratio to achieve local convergence. The final individual update result is a weighted synthesis of the two parts. S46. When it is detected that the validation error of an individual continues to decrease over multiple iterations but the training error remains unchanged, the individual is judged to have fallen into a local overfitting trap. At this time, the local perturbation jump mechanism is triggered. The local perturbation jump mechanism adds a slight Gaussian perturbation to the current parameters to achieve a jump out of the local optimal area. S47. Repeat the trend-weighted fitness function construction, stage-by-stage weight adjustment, trend deviation perception update, and local disturbance jump mechanism until the termination condition is met. The termination condition includes that the current number of iterations reaches the maximum number of iterations, or the change in the optimal individual between two consecutive rounds is less than a preset error threshold. The optimization process stops when any of the termination conditions is met. S48. Select the individual with the largest fitness value from all search individuals as the final optimization result. The parameter set corresponding to the final optimization result is the optimized parameter set θ. * , the optimized parameter set replaces the original parameter set in the bone tumor postoperative rehabilitation trend discrimination model to obtain the final bone tumor postoperative rehabilitation trend discrimination model.

6. A bone tumor rehabilitation process evaluation method based on deep learning according to claim 5, characterized in that: The S6 comprises the following steps: S61. Input the preprocessed multi-source rehabilitation dataset of the patient undergoing bone tumor surgery to the trained improved Octave convolutional network to generate a fused rehabilitation feature representation. This fused rehabilitation feature representation is then input into the trained bone tumor surgery rehabilitation trend discrimination model to obtain a series of rehabilitation trend prediction values continuously output by the model. This series of rehabilitation trend prediction values represents the changes in the patient's rehabilitation status at consecutive time points and is used to construct a rehabilitation trend curve. S62. Calculate the rehabilitation trend slope and volatility index based on the rehabilitation trend prediction value sequence. The rehabilitation trend slope is used to indicate the speed and direction of rehabilitation progress, and the volatility index is used to measure the stability of the rehabilitation process. Generate a rehabilitation progress grade based on the final rehabilitation trend prediction value, trend slope, and volatility index. S63. Based on the rehabilitation trend prediction value sequence and the patient's historical rehabilitation assessment reference threshold, calculate the probability of rehabilitation failure in the future prediction time period and output the prognostic risk level.

7. A bone tumor rehabilitation process evaluation method based on deep learning according to claim 6, characterized in that: The rehabilitation progress levels are divided according to the following rules: Rehabilitation progress level I is the rapid recovery stage: the slope of the rehabilitation trend is greater than 0.5 and the volatility index is less than 0.2; Rehabilitation progress level II is the stable rehabilitation stage: the slope of the rehabilitation trend is 0.20.5 and the volatility index is 0.20.4; Rehabilitation progress level III is the slow recovery stage: the slope of the rehabilitation trend is ≤0.2 or the volatility index is >0.

4.

8. The method for evaluating bone tumor rehabilitation process based on deep learning according to claim 6, characterized in that: The prognostic risk levels are divided according to the following rules: Prognostic risk level A is a low-risk stage: all predicted values in the next 7 days are higher than the stable threshold; Prognostic risk level B is the medium-risk stage: the predicted value will be lower than the stable threshold 1-2 times in the next 7 days; Prognostic risk level C is a high-risk stage: the predicted value in the next 7 days will be lower than the stable threshold for 3 or more times in a row.

9. A deep learning-based bone tumor rehabilitation process evaluation system, used to execute a deep learning-based bone tumor rehabilitation process evaluation method according to any one of claims 1 to 8, characterized in that: Includes the following modules: Data acquisition module, used to collect original data sets of multi-source rehabilitation after bone tumor surgery; The data preprocessing module is used to perform standardization, denoising, time series alignment, and missing value filling on the original multi-source rehabilitation data set after bone tumor surgery to generate a preprocessed multi-source rehabilitation data set after bone tumor surgery; The feature extraction module includes a frequency domain partitioning submodule and an improved Octave convolution feature extraction submodule. The frequency domain partitioning submodule is used to divide the preprocessed rehabilitation dataset into high-frequency features and low-frequency features according to the adaptive frequency allocation factor. The improved Octave convolution feature extraction submodule is used to perform lesion-guided decoupled convolution operations, cross-channel residual fusion, and dynamic complementary updates on the high-frequency features and low-frequency features, respectively, to ultimately generate a fused rehabilitation feature representation. The trend discrimination model construction module is used to input the fusion rehabilitation feature representation into the initial bone tumor postoperative rehabilitation trend discrimination model to complete the structure construction; The model parameter optimization module includes a hybrid seagull search optimization submodule. This submodule iteratively optimizes the model parameters based on the trend feedback vector and the multi-objective fitness function, combining a phased search adjustment mechanism with a local perturbation mechanism to obtain the final optimal parameter set and complete the training of the bone tumor postoperative rehabilitation trend discrimination model. The trend result output module is used to receive the output results of the trained bone tumor postoperative rehabilitation trend discrimination model, and combine the trend prediction value sequence to generate a rehabilitation trend curve, rehabilitation progress level and prognostic risk level for bone tumor rehabilitation process evaluation.

Images