Survival curve image data reconstruction method based on deep learning

Abstract

The invention discloses a survival curve image data reconstruction method based on deep learning, and the method comprises the steps: S1, obtaining to-be-reconstructed survival curve image data; s2, carrying out standardization processing, de-noising processing and blocking processing on the survival curve image data to be reconstructed; s3, constructing an improved deep learning network model; s4, optimizing hyper-parameters and network weights of the deep learning network model based on a spider bee optimization algorithm; and S5, inputting survival curve image data to be reconstructed into the trained deep learning network model to generate reconstructed survival curve image data. According to the method, the problems of feature loss and overfitting are effectively avoided, the reconstruction precision is remarkably improved, and the consistency of the generated survival curve image data at the pixel level and the structure level is ensured.

Description

Technical Field

[0001] The present invention relates to the technical field of image data reconstruction, and in particular to a survival curve image data reconstruction method based on deep learning. Background Art

[0002] With the rapid development of deep learning and artificial intelligence technologies, data reconstruction technology has been widely used in the fields of medicine and biostatistics. Survival curves, as an important tool in medical statistics and biological research, are used to describe the probability of events over time, such as patient survival rate or disease recurrence rate. However, due to the limitations of the capabilities of data acquisition equipment and the interference of environmental factors, the integrity and quality of survival curve image data are often difficult to guarantee, which brings many challenges to research and clinical applications in related fields.

[0003] At present, for the reconstruction and optimization of survival curve image data, traditional methods mainly rely on interpolation algorithms or statistical modeling techniques. Linear interpolation is used to fill in the data by performing simple linear fitting between missing points. Statistical modeling methods fit the data based on certain mathematical models. However, traditional methods have obvious limitations in practical applications. First, linear interpolation is too simple and can only be applied to scenarios where the data change trend is relatively linear. It cannot accurately restore the complex nonlinear characteristics that may exist in the survival curve. Second, although statistical modeling methods can handle a certain degree of nonlinear problems, they are highly dependent on model assumptions. When there are many missing data or the data quality is poor, it is often difficult to generate a survival curve that conforms to reality.

[0004] In recent years, image reconstruction technology based on deep learning has gradually emerged and demonstrated powerful data nonlinear modeling and feature extraction capabilities. Some studies have attempted to use convolutional neural networks or generative adversarial networks to reconstruct survival curve image data, but these methods still have the following problems in practical applications: On the one hand, the existing deep learning models rely heavily on hyperparameter settings and weight initialization. When the input data has high noise or severe missing data, the model is prone to fall into local optimality; on the other hand, existing methods often ignore the statistical characteristics and structural consistency requirements of survival curves in medical statistics, resulting in insufficient credibility of the reconstruction results. In addition, the current deep learning model training process has a high computational cost and is difficult to meet the requirements for efficiency in practical applications.

[0005] In summary, the existing technology faces the problems of low reconstruction accuracy, poor adaptability and high computational cost when processing survival curve image data, and cannot effectively meet the needs of high-quality reconstruction of survival curve data in medical statistics and clinical research. Therefore, there is an urgent need for an improved method based on deep learning, which can improve the reconstruction accuracy based on the characteristics of survival curve image data and combine optimization algorithms, while improving the efficiency and robustness of the model, so as to overcome the defects of the existing technology and meet the needs of practical applications. Summary of the invention

[0006] One object of the present invention is to propose a survival curve image data reconstruction method based on deep learning, which effectively avoids feature loss and overfitting problems, significantly improves the reconstruction accuracy, and ensures the consistency of the generated survival curve image data at the pixel level and the structural level.

[0007] A survival curve image data reconstruction method based on deep learning according to an embodiment of the present invention comprises the following steps:

[0008] S1, obtaining survival curve image data to be reconstructed;

[0009] S2, performing standardization, denoising and block processing on the survival curve image data to be reconstructed;

[0010] S3, build an improved deep learning network model;

[0011] S4, optimize the hyperparameters and network weights of the deep learning network model based on the spider bee optimization algorithm;

[0012] S5. Inputting the survival curve image data to be reconstructed into the trained deep learning network model to generate reconstructed survival curve image data.

[0013] Optionally, the S1 specifically includes:

[0014] S11. Obtain survival curve image data D from a database, wherein the survival curve image data is used to describe the probability of occurrence of an event over time, and the survival curve image data D includes complete survival curve image data, incomplete survival curve image data, or damaged survival curve image data, wherein:

[0015] Complete survival curve image data D c It indicates the survival curve image data that was not lost or damaged during the data collection process;

[0016] Incomplete survival curve image data D i Indicates survival curve image data with missing time points or event probabilities due to data loss during the acquisition process;

[0017] Damaged survival curve image data D d Represents survival curve image data where parts of the image are distorted or blurred due to noise interference, acquisition equipment limitations, or storage errors;

[0018] S12. Check the time range and event probability range in the survival curve image data:

[0019] Define the time range T = {t 1 ,t 2 ,...,t n}, where t 1 and t n They are the start time and end time of the survival curve image data respectively;

[0020] Define the event probability range P(t)∈[0,1], where P(t) is the probability of the event occurring at time point t;

[0021] S13. For incomplete survival curve image data D i and damaged survival curve image data D d Mark the outliers and generate an outlier set E = {e 1 ,e 2 ,...,e m}, where e m Indicates the position of the abnormal point on the time axis, and the time axis range is T;

[0022] S15, the annotated complete survival curve image data, incomplete survival curve image data, damaged survival curve image data and abnormal point set E are combined into a survival curve image data set:

[0023] D'=(D,E).

[0024] Optionally, the S2 specifically includes:

[0025] S21, performing standardization processing on the survival curve image data D' to be reconstructed, defining a time axis range, and normalizing the time range of all survival curve image data to a standard time interval;

[0026] S22, denoising the noise in the survival curve image data D', smoothing the local noise in the survival curve image by using a median filter method, decomposing the signal S into an approximate component A and a detail component D by using a wavelet transform for global noise, and applying a soft threshold function to process the detail component D;

[0027] S23, block processing is performed on the denoised survival curve image data D'', and the time axis range T' is evenly divided into k sub-intervals T k , each subinterval is defined as:

[0028] T k = {t′ j |t′ k-1 ≤ t′ j < t′ k}, k ∈ {1, 2,..., K};

[0029] where t′ k represents the boundary value of the sub - interval;

[0030] Take the event probability value P(T k ) of the corresponding sub - interval T k as the feature subset after chunking, and generate the survival curve data set after chunking:

[0031] D b = {(T k , P(T k ))|k ∈ {1, 2,..., K}}.

[0032] Optionally, the S3 specifically includes:

[0033] S31. Build an improved deep learning network model, where the improved deep learning network model includes an input layer, a parallel multi - scale feature extraction network, a cross - attention mechanism module, a residual reconstruction network, and an output layer;

[0034] S32. Input the processed survival curve image data set D b into the input layer. The input layer receives and normalizes the data dimension, and distributes the survival curve image data set D b to multiple branches of the parallel multi - scale feature extraction network;

[0035] S33. In the parallel multi - scale feature extraction network, use multiple groups of convolutional layers to extract features from survival curve image data of different scales. Let the feature map extracted by the m - th branch be:

[0036] F m = σ(W m * D b,m + b m );

[0037] where D b,m is the survival curve image data distributed by the input layer to the m - th branch, W m and b m respectively represent the weight and bias of the convolutional layer of the m - th branch, σ is the activation function, and m ∈ {1, 2,..., M} represents the parallel branch number;

[0038] Fuse the feature maps extracted from multiple branches through concatenation or summation to obtain the multi - scale fusion feature Fms ;

[0039] S34, in the cross attention mechanism module, the multi-scale fusion feature F ms Perform comprehensive weighting of channel and spatial dimensions to calculate the channel attention weight α c , the channel characteristic value is h c , the number of channels is C, then:

[0040]

[0041] Calculate the spatial attention map β(x,y), let f(x,y) be the eigenvalue corresponding to the position (x,y), and the spatial attention weight is:

[0042]

[0043] Among them, U represents all pixel positions of the feature map, and the channel attention weight and spatial attention weight are jointly applied to the multi-scale fusion feature F ms Get the weighted feature F ca ;

[0044] S35, in the residual reconstruction network, the weighted feature F ca The reconstruction process is performed by combining deconvolution operation with skip connection. The residual signal of input and output is retained at each layer during the reconstruction process. The output of the rth layer deconvolution is:

[0045]

[0046] Among them, R r-1 is the output feature map of the previous layer, S r-1 is the skip connection feature of the previous layer, which is used to preserve the local survival curve image data details. and Represent the deconvolution weight and bias of the rth layer respectively;

[0047] The output of the last layer is used as the reconstructed survival curve image data R;

[0048] S36, inputting the survival curve image data R output by the residual reconstruction network into the output layer, and the output layer performs dimension or format standardization processing on the reconstruction result to obtain the final reconstructed survival curve image data.

[0049] Optionally, the S4 specifically includes:

[0050] S41. Initialize the spider population, set the population size to N, each spider individual represents a hyperparameter combination or network weight combination of the deep learning network model, and the initial state of the i-th individual is:

[0051] X i ={p 1 ,p 2 ,...,p m};

[0052] Among them, X i represents the parameter combination of the ith individual, p j represents the jth hyperparameter or weight of the deep learning network, and m is the total number of parameters;

[0053] S42, define a fitness function, the fitness function is evaluated based on the reconstruction error of the generated reconstructed survival curve image data, and the fitness value F(X i ) is calculated as:

[0054] F(X i )=α·L MSE +β·L SSIM ;

[0055] Among them, F(X i ) is the fitness value of the i-th individual. The smaller the fitness value, the better the individual optimization effect. MSE is the pixel-level mean square error between the reconstructed survival curve image data and the real survival curve image data, L SSIM is the structural similarity measure between the reconstructed survival curve image data and the real survival curve image data, α and β are weight coefficients;

[0056] S43, based on the local neighborhood information, the position of the spider individual is optimized to calculate the interaction influence between the i-th individual and the neighboring individuals:

[0057]

[0058] in, and are the positions of the ith individual in round t and round t+1, respectively, N i is the neighborhood set of the ith individual, w k is the weight of the neighborhood individual k, indicating its influence on individual i, λ 1 is the step size factor, used to control the update amplitude;

[0059] S44, perform a global search for weight parameters, and randomly select some individuals to move closer to the global optimal position by simulating the information propagation behavior of the spider web:

[0060]

[0061] in, is the individual position with the lowest fitness value in the current iteration, γ is the global search factor, which is used to control the global exploration step size;

[0062] S45, iteratively update the fitness function value, the iteration termination condition is to reach the maximum number of iterations T or the fitness function converges to the set threshold, and finally output the global optimal parameter combination:

[0063]

[0064] Among them, X opt Represents the optimized combination of hyperparameters and network weights of the deep learning network model.

[0065] Optionally, the S42 specifically includes:

[0066] S421, L MSE Represents the mean square error, which is used to measure the average difference between the reconstructed survival curve image data and the real survival curve image data at the pixel level:

[0067]

[0068] Among them, y j is the value of the jth pixel in the real survival curve image data, is the value of the jth pixel in the reconstructed survival curve image data, and n is the total number of pixels;

[0069] S422, L SSIM Represents the structural similarity index, which is used to evaluate the consistency of the visual structure of the reconstructed survival curve image data:

[0070]

[0071] in, Represents the real data Y and the reconstructed data The structural similarity index of:

[0072]

[0073] Among them, μ Y and Y and The mean of and Y and The variance of For Y and The covariance, C 1 and C 2 is a constant used to avoid the denominator being zero;

[0074] S423, based on the reconstructed survival curve image data R generated by the optimized deep learning network model, calculate its fitness value F(X i ), which comprehensively reflects the pixel accuracy and visual structure consistency of the reconstructed data;

[0075] S424, by adjusting the weight coefficients α and β to control the fitness function for the mean square error L MSE and structural similarity L SSIM sensitivity to meet the optimization requirements of different application scenarios.

[0076] Optionally, the S5 specifically includes:

[0077] S51, the survival curve image data set D to be reconstructed b Input to the optimized deep learning network model;

[0078] S52, in the parallel multi-scale feature extraction network, extract the multi-scale features of the input survival curve image data based on the convolution operation, and generate a fused multi-scale feature map F ms And the multi-scale feature map F is ms Perform weighted processing to generate an optimized multi-scale feature map F ca ;

[0079] S53, the optimized multi-scale feature map F ca Input to the residual reconstruction network, the reconstruction network combines multi-layer deconvolution operations with jump connections to generate reconstructed survival curve image data R;

[0080] S54, the output layer processes the reconstruction result of the last layer to generate the final reconstructed survival curve image data R final , and the final reconstructed survival curve image data R final The output is the reconstruction result of the deep learning network;

[0081] S55, the generated reconstructed survival curve image data R final Compared with the fitness function evaluation standard, the reconstruction effect is verified, so that the reconstructed survival curve image data meets the consistency requirements of pixel-level accuracy and structural similarity.

[0082] The beneficial effects of the present invention are:

[0083] (1) The present invention adopts an improved deep learning network model, constructs a parallel multi-scale feature extraction network, a cross-attention mechanism module and a residual reconstruction network. The parallel multi-scale feature extraction network extracts features of different scales through convolution operations, thereby realizing a comprehensive capture of the complex structural features of the survival curve image. The cross-attention mechanism module improves the network's attention to the key areas of the survival curve by dynamically weighting in the channel and spatial dimensions. The residual reconstruction network combines the jump connection mechanism to retain the detailed features in the survival curve image, effectively avoiding feature loss and overfitting problems, significantly improving the reconstruction accuracy, and ensuring the consistency of the generated survival curve image data at the pixel level and structure level.

[0084] (2) The present invention globally optimizes the hyperparameters and weights of the deep learning network model through the spider bee optimization algorithm. The spider bee algorithm uses local neighborhood optimization and global search strategies to effectively solve the problem that the deep learning model is prone to fall into local optimality when facing high-noise data or data missing. The fitness function combines the mean square error and structural similarity indicators to comprehensively evaluate the pixel accuracy and visual consistency of the reconstruction results, ensuring the reliability of the optimization results. Experimental results show that the optimization algorithm of the present invention makes the network model more adaptable to different data qualities and shows stronger robustness in the reconstruction of survival curve image data.

[0085] (3) The present invention realizes the full process automation from the preprocessing of survival curve image data, network training to reconstruction output, avoiding a large amount of manual intervention in parameter adjustment and model adjustment in traditional methods. The efficient search mechanism of the spider bee optimization algorithm significantly reduces the number of iterations required for deep learning network training. The optimized model not only improves the computing efficiency, but also reduces the dependence on hardware resources. It shows higher efficiency in the reconstruction task of large-scale survival curve image data and can meet the timeliness requirements of data processing in practical applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0086] 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:

[0087] Figure 1 This is a flow chart of a survival curve image data reconstruction method based on deep learning proposed by the present invention. DETAILED DESCRIPTION

[0088] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, which only illustrate the basic structure of the present invention in a schematic manner, and therefore only show the components related to the present invention.

[0089] refer to Figure 1, a survival curve image data reconstruction method based on deep learning, the method comprises the following steps:

[0090] S1, obtaining survival curve image data to be reconstructed;

[0091] S2, performing standardization, denoising and block processing on the survival curve image data to be reconstructed;

[0092] S3, build an improved deep learning network model;

[0093] S4, optimize the hyperparameters and network weights of the deep learning network model based on the spider bee optimization algorithm;

[0094] S5. Inputting the survival curve image data to be reconstructed into the trained deep learning network model to generate reconstructed survival curve image data.

[0095] In this implementation, S1 specifically includes:

[0096] S11. Obtain survival curve image data D from a database. The survival curve image data is used to describe the probability of an event changing over time. The survival curve image data D includes complete survival curve image data, incomplete survival curve image data, or damaged survival curve image data, wherein:

[0097] Complete survival curve image data D c It indicates the survival curve image data that was not lost or damaged during the data collection process;

[0098] Incomplete survival curve image data D i Indicates survival curve image data with missing time points or event probabilities due to data loss during the acquisition process;

[0099] Damaged survival curve image data D d Represents survival curve image data where parts of the image are distorted or blurred due to noise interference, acquisition equipment limitations, or storage errors;

[0100] S12. Check the time range and event probability range in the survival curve image data:

[0101] Define the time range T = {t 1 ,t 2 ,...,t n}, where t 1 and t n They are the start time and end time of the survival curve image data respectively;

[0102] Define the event probability range P(t)∈[0,1], where P(t) is the probability of the event occurring at time point t;

[0103] S13. For incomplete survival curve image data D i and damaged survival curve image data D d Mark the outliers and generate an outlier set E = {e 1 ,e 2 ,...,e m}, where e m Indicates the position of the abnormal point on the time axis, and the time axis range is T;

[0104] S15, the annotated complete survival curve image data, incomplete survival curve image data, damaged survival curve image data and abnormal point set E are combined into a survival curve image data set:

[0105] D'=(D,E).

[0106] In this implementation, S2 specifically includes:

[0107] S21, performing standardization processing on the survival curve image data D' to be reconstructed, defining a time axis range, and normalizing the time range of all survival curve image data to a standard time interval;

[0108] S22, denoising the noise in the survival curve image data D', smoothing the local noise in the survival curve image by using a median filter method, decomposing the signal S into an approximate component A and a detail component D by using a wavelet transform for global noise, and applying a soft threshold function to process the detail component D;

[0109] S23, block processing is performed on the denoised survival curve image data D'', and the time axis range T' is evenly divided into k sub-intervals T k , each subinterval is defined as:

[0110] T k ={t′ j ∣t′ k-1 ≤t′ j <t′ k}, k∈{1,2,...,K};

[0111] Among them, t′ k Represents the boundary value of the subinterval;

[0112] The corresponding subinterval T k The event probability value P(T k ) as the feature subset after block division, and generate the survival curve data set after block division:

[0113] D b ={(T k ,P(T k))|k∈{1,2,...,K}}.

[0114] In this implementation, S3 specifically includes:

[0115] S31. Build an improved deep learning network model, which includes an input layer, a parallel multi-scale feature extraction network, a cross-attention mechanism module, a residual reconstruction network, and an output layer;

[0116] S32, the processed survival curve image data set D b Input to the input layer, the input layer receives and normalizes the data dimension, and converts the survival curve image data set D b Assigned to multiple branches of a parallel multi-scale feature extraction network;

[0117] S33. In the parallel multi-scale feature extraction network, multiple groups of convolutional layers are used to extract features from survival curve image data of different scales, and the feature map extracted by the mth branch is:

[0118] F m =σ(W m *D b,m +b m );

[0119] Among them, D b,m is the survival curve image data assigned to the mth branch by the input layer, W m and b m They represent the weight and bias of the m-th branch convolutional layer, σ is the activation function, and m∈{1,2,...,M} represents the parallel branch number;

[0120] The feature maps extracted by multiple branches are fused by cascading or adding to obtain the multi-scale fusion feature F. ms ;

[0121] S34. In the cross attention mechanism module, the multi-scale fusion feature F ms Perform comprehensive weighting of channel and spatial dimensions to calculate the channel attention weight α c , the channel characteristic value is h c , the number of channels is C, then:

[0122]

[0123] Calculate the spatial attention map β(x,y), let f(x,y) be the eigenvalue corresponding to the position (x,y), and the spatial attention weight is:

[0124]

[0125] Among them, U represents all pixel positions of the feature map, and the channel attention weight and spatial attention weight are jointly applied to the multi-scale fusion feature F ms Get the weighted feature F ca ;

[0126] S35. Weighted feature F in residual reconstruction network ca The reconstruction process is performed by combining deconvolution operation with skip connection. The residual signal of input and output is retained at each layer during the reconstruction process. The output of the rth layer deconvolution is:

[0127]

[0128] Among them, R r-1 is the output feature map of the previous layer, S r-1 is the skip connection feature of the previous layer, which is used to preserve the local survival curve image data details. and Represent the deconvolution weight and bias of the rth layer respectively;

[0129] The output of the last layer is used as the reconstructed survival curve image data R;

[0130] S36, inputting the survival curve image data R output by the residual reconstruction network into the output layer, and the output layer performs dimension or format standardization processing on the reconstruction result to obtain the final reconstructed survival curve image data.

[0131] In this implementation, S4 specifically includes:

[0132] S41. Initialize the spider population, set the population size to N, each spider individual represents a hyperparameter combination or network weight combination of the deep learning network model, and the initial state of the i-th individual is:

[0133] X i ={p 1 ,p 2 ,...,p m};

[0134] Among them, X i represents the parameter combination of the ith individual, p j represents the jth hyperparameter or weight of the deep learning network, and m is the total number of parameters;

[0135] S42, define a fitness function, the fitness function is evaluated based on the reconstruction error of the generated reconstructed survival curve image data, and the fitness value F(X i ) is calculated as:

[0136] F(X i )=α·L MSE +β·LSSIM ;

[0137] Among them, F(X i ) is the fitness value of the i-th individual. The smaller the fitness value, the better the individual optimization effect. MSE is the pixel-level mean square error between the reconstructed survival curve image data and the real survival curve image data, L SSIM is the structural similarity measure between the reconstructed survival curve image data and the real survival curve image data, α and β are weight coefficients;

[0138] S43, based on the local neighborhood information, the position of the spider individual is optimized to calculate the interaction influence between the i-th individual and the neighboring individuals:

[0139]

[0140] in, and are the positions of the ith individual in round t and round t+1, respectively, N i is the neighborhood set of the ith individual, w k is the weight of the neighborhood individual k, indicating its influence on individual i, λ 1 is the step size factor, used to control the update amplitude;

[0141] S44, perform a global search for weight parameters, and randomly select some individuals to move closer to the global optimal position by simulating the information propagation behavior of the spider web:

[0142]

[0143] in, is the individual position with the lowest fitness value in the current iteration, γ is the global search factor, which is used to control the global exploration step size;

[0144] S45, iteratively update the fitness function value, the iteration termination condition is to reach the maximum number of iterations T or the fitness function converges to the set threshold, and finally output the global optimal parameter combination:

[0145]

[0146] Among them, X opt Represents the optimized combination of hyperparameters and network weights of the deep learning network model.

[0147] In this implementation, S42 specifically includes:

[0148] S421, L MSE Represents the mean square error, which is used to measure the average difference between the reconstructed survival curve image data and the real survival curve image data at the pixel level:

[0149]

[0150] Among them, y j is the value of the jth pixel in the real survival curve image data, is the value of the jth pixel in the reconstructed survival curve image data, and n is the total number of pixels;

[0151] S422, L SSIM Represents the structural similarity index, which is used to evaluate the consistency of the visual structure of the reconstructed survival curve image data:

[0152]

[0153] in, Represents the real data Y and the reconstructed data The structural similarity index of:

[0154]

[0155] Among them, μ Y and Y and The mean of and Y and The variance of For Y and The covariance, C 1 and C 2 is a constant used to avoid the denominator being zero;

[0156] S423, based on the reconstructed survival curve image data R generated by the optimized deep learning network model, calculate its fitness value F(X i ), which comprehensively reflects the pixel accuracy and visual structure consistency of the reconstructed data;

[0157] S424, by adjusting the weight coefficients α and β to control the fitness function for the mean square error L MSE and structural similarity L SSIM sensitivity to meet the optimization requirements of different application scenarios.

[0158] In this implementation, S5 specifically includes:

[0159] S51, the survival curve image data set D to be reconstructed b Input to the optimized deep learning network model;

[0160] S52, in the parallel multi-scale feature extraction network, extract the multi-scale features of the input survival curve image data based on the convolution operation, and generate a fused multi-scale feature map F msAnd the multi-scale feature map F is ms Perform weighted processing to generate an optimized multi-scale feature map F ca ;

[0161] S53, the optimized multi-scale feature map F ca Input to the residual reconstruction network, the reconstruction network combines multi-layer deconvolution operations with jump connections to generate reconstructed survival curve image data R;

[0162] S54, the output layer processes the reconstruction result of the last layer to generate the final reconstructed survival curve image data R final , and the final reconstructed survival curve image data R final The output is the reconstruction result of the deep learning network;

[0163] S55, the generated reconstructed survival curve image data R final Compared with the fitness function evaluation standard, the reconstruction effect is verified, so that the reconstructed survival curve image data meets the consistency requirements of pixel-level accuracy and structural similarity.

[0164] Embodiment 1:

[0165] Embodiment During the period from June 2024 to December 2024, a provincial cancer center analyzed the patient survival information in its database and found a batch of missing or damaged survival curve image data. The data involved about 500 patients, of which 60% were complete data, and the remaining 40% of the data had varying degrees of missing and noise. The main reasons for the missing data included incomplete follow-up records and data loss during equipment conversion. The noise problem mainly came from the aging of the acquisition equipment and the interference during the storage format conversion process. Since these data directly affect the statistical analysis of subsequent treatment plans, the hospital decided to introduce the method of the present invention for data reconstruction.

[0166] In a test in July 2024, the research team conducted a detailed analysis of a follow-up record of patient A (No.: P001). The survival curve of patient A recorded a time span of 3 years, with a total of 36 time points, of which 10 time points had missing event probability data. At the same time, the recorded curve contained about 5% high-frequency noise points. Using the traditional linear interpolation method, the reconstructed curve showed obvious deviations near the time point "18th month", and the data fluctuation range exceeded a reasonable range, resulting in the analysis results being unable to be used for medical decision-making.

[0167] In order to solve this problem, the research team used the method of the present invention to process the survival curve image data of patient A. First, the team standardized and denoised the original data and found that there was significant high-frequency noise in the curve parts near "the 6th month" and "the 24th month". After median filtering, the proportion of noise points dropped to 0.3%. Then, the team divided the time axis into blocks of 3 months and divided the survival curve into 12 time intervals. The probability value of the event occurrence in each interval was extracted as the block feature of the model input.

[0168] In a model training on August 15, 2024, the team input patient A's data together with the survival curve image data of other patients into the deep learning network model for training. After multiple iterations of the Spider Bee optimization algorithm, the model's hyperparameters were gradually adjusted to the optimal state. Finally, the reconstruction result generated by the model had an event probability of 0.57 near the "18th month", which was consistent with the data change trend at adjacent time points, and was verified by the structural similarity index, with an SSIM value of 0.92.

[0169] To further verify the effectiveness of the method, the team processed the survival curve data of patient B (No.: P002). The data of patient B covered 60 months and included 50 time points. The curves near the "10th month" and "40th month" were severely damaged. After repairing them using traditional statistical fitting methods, the event probability was shown as 0.4 at the "10th month", which was quite different from the real data of 0.25, resulting in a high survival analysis result. After using the method of the present invention, the reconstructed event probability of the model at the "10th month" was 0.26, which was almost consistent with the real data.

[0170] In a comprehensive test in November 2024, the team divided the survival curve image data of 500 patients into a training set (70%) and a test set (30%), and compared them with traditional methods. The test results are shown in Table 1:

[0171] Table 1 Comparison results of survival curve image data reconstruction methods

[0172] Test indicators Traditional linear interpolation Traditional statistical fitting Method of the present invention Average MSE 0.031 0.024 0.011 Average SSIM 0.813 0.845 0.926 Reconstruction time (seconds / image) 0.03 0.2 0.12 Error at key time point (%) 7.2 4.5 1.3

[0173] Through detailed analysis, the data of patient C (number: P003) in the test set performed particularly well. The data of patient C covered 30 months, of which 5 time points were missing and the event probability showed obvious fluctuations in the "20th month". The method of the present invention successfully repaired the fluctuations in the "20th month", and the event probability of the generated reconstructed curve at this time point was 0.48, which was close to the real data 0.47, while the results of the traditional method were 0.60 and 0.52 respectively.

[0174] On December 31, 2024, the research team officially completed the reconstruction of all data, and the generated high-quality survival curve image data was used for statistical analysis of tumor treatment plans. According to the statistical results, the hospital found that the new treatment method significantly improved the three-year survival rate of a certain patient group (from the original 62% to 74%), providing important data support for subsequent large-scale promotion.

[0175] In summary, this embodiment verifies the feasibility and effectiveness of the method of the present invention through detailed scenarios and data, and fully demonstrates its technical advantages and clinical application value in the reconstruction of survival curve image data.

[0176] The present invention adopts an improved deep learning network model, constructs a parallel multi-scale feature extraction network, a cross-attention mechanism module and a residual reconstruction network. The parallel multi-scale feature extraction network extracts features of different scales through convolution operations, thereby realizing a comprehensive capture of the complex structural features of the survival curve image. The cross-attention mechanism module improves the network's attention to the key areas of the survival curve by dynamically weighting in the channel and spatial dimensions. The residual reconstruction network combines the jump connection mechanism to retain the detailed features in the survival curve image, effectively avoiding feature loss and overfitting problems, significantly improving the reconstruction accuracy, and ensuring the consistency of the generated survival curve image data at the pixel level and the structural level.

[0177] The present invention globally optimizes the hyperparameters and weights of the deep learning network model through the spider bee optimization algorithm. The spider bee algorithm utilizes local neighborhood optimization and global search strategies to effectively solve the problem that the deep learning model is prone to fall into local optimality when faced with high-noise data or missing data. The fitness function combines the mean square error and structural similarity indicators to comprehensively evaluate the pixel accuracy and visual consistency of the reconstruction results, ensuring the reliability of the optimization results. Experimental results show that the optimization algorithm of the present invention makes the network model more adaptable to different data qualities and exhibits stronger robustness in the reconstruction of survival curve image data.

[0178] The present invention realizes the full-process automation processing from preprocessing of survival curve image data, network training to reconstruction output, avoiding a large amount of manual intervention in parameter adjustment and model adjustment in traditional methods. The efficient search mechanism of the spider bee optimization algorithm significantly reduces the number of iterations required for deep learning network training. The optimized model not only improves the computing efficiency, but also reduces the dependence on hardware resources. It shows higher efficiency in the reconstruction task of large-scale survival curve image data and can meet the timeliness requirements of data processing in practical applications.

[0179] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A survival curve image data reconstruction method based on deep learning, characterized in that: The steps include: S1, obtaining survival curve image data to be reconstructed; S2, performing standardization, denoising and block processing on the survival curve image data to be reconstructed; S3, build an improved deep learning network model; S4, optimize the hyperparameters and network weights of the deep learning network model based on the spider bee optimization algorithm; S5. Inputting the survival curve image data to be reconstructed into the trained deep learning network model to generate reconstructed survival curve image data.

2. The method for reconstructing survival curve image data based on deep learning according to claim 1, characterized in that: The S1 specifically includes: S11. Obtain survival curve image data D from a database, wherein the survival curve image data is used to describe the probability of occurrence of an event over time, and the survival curve image data D includes complete survival curve image data, incomplete survival curve image data, or damaged survival curve image data, wherein: Complete survival curve image data D c It indicates the survival curve image data that was not lost or damaged during the data collection process; Incomplete survival curve image data D i Indicates the survival curve image data with some time points or event probabilities missing due to data loss during the acquisition process; Damaged survival curve image data D d Represents survival curve image data where parts of the image are distorted or blurred due to noise interference, acquisition equipment limitations, or storage errors; S12. Check the time range and event probability range in the survival curve image data: Define the time range T = {t1, t2, ..., t n }, where t1 and t n They are the start time and end time of the survival curve image data respectively; Define the event probability range P(t)∈[0,1], where P(t) is the probability of the event occurring at time point t; S13. For incomplete survival curve image data D i and damaged survival curve image data D d Mark the outliers and generate an outlier set E = {e1, e2, ..., e m }, where e m Indicates the position of the abnormal point on the time axis, and the time axis range is T; S15, the annotated complete survival curve image data, incomplete survival curve image data, damaged survival curve image data and abnormal point set E are combined into a survival curve image data set: D'=(D,E).

3. The method for reconstructing survival curve image data based on deep learning according to claim 1, characterized in that: The S2 specifically includes: S21, performing standardization processing on the survival curve image data D' to be reconstructed, defining a time axis range, and normalizing the time range of all survival curve image data to a standard time interval; S22, denoising the noise in the survival curve image data D', smoothing the local noise in the survival curve image by using a median filter method, decomposing the signal S into an approximate component A and a detail component D by using a wavelet transform for global noise, and applying a soft threshold function to process the detail component D; S23, block processing is performed on the denoised survival curve image data D'', and the time axis range T' is evenly divided into k sub-intervals T k , each subinterval is defined as: T k ={t′ j ∣t′ k-1 ≤t′ j <t′ k },k∈{1,2,...,K}; Among them, t′ k Represents the boundary value of the subinterval; The corresponding subinterval T k The event probability value P(T k ) as the feature subset after block division, and generate the survival curve data set after block division: D b ={(T k ,P(T k ))∣k∈{1,2,...,K}}。 4. The method for reconstructing survival curve image data based on deep learning according to claim 1, characterized in that: The S3 specifically includes: S31. Build an improved deep learning network model, which includes an input layer, a parallel multi-scale feature extraction network, a cross-attention mechanism module, a residual reconstruction network, and an output layer; S32, the processed survival curve image data set D b Input to the input layer, the input layer receives and normalizes the data dimension, and converts the survival curve image data set D b Assigned to multiple branches of a parallel multi-scale feature extraction network; S33. In the parallel multi-scale feature extraction network, multiple groups of convolutional layers are used to extract features from survival curve image data of different scales, and the feature map extracted by the mth branch is: F m =σ(W m *D b,m +b m ); Among them, D b,m is the survival curve image data assigned to the mth branch by the input layer, W m and b m They represent the weight and bias of the m-th branch convolution layer, σ is the activation function, and m∈{1,2,...,M} represents the parallel branch number; The feature maps extracted by multiple branches are fused by cascading or adding to obtain the multi-scale fusion feature F. ms ; S34, in the cross attention mechanism module, the multi-scale fusion feature F ms Perform comprehensive weighting of channel and spatial dimensions to calculate the channel attention weight α c , the channel characteristic value is h c , the number of channels is C, then: Calculate the spatial attention map β(x,y), let f(x,y) be the eigenvalue corresponding to the position (x,y), and the spatial attention weight is: Among them, U represents all pixel positions of the feature map, and the channel attention weight and spatial attention weight are jointly applied to the multi-scale fusion feature F ms Get the weighted feature F ca ; S35, in the residual reconstruction network, the weighted feature F ca The reconstruction process is performed by combining deconvolution operation with skip connection. The residual signal of input and output is retained at each layer during the reconstruction process. The output of the rth layer deconvolution is: Among them, R r-1 is the output feature map of the previous layer, S r-1 is the skip connection feature of the previous layer, which is used to preserve the local survival curve image data details. and Represent the deconvolution weight and bias of the rth layer respectively; The output of the last layer is used as the reconstructed survival curve image data R; S36, inputting the survival curve image data R output by the residual reconstruction network into the output layer, and the output layer performs dimension or format standardization processing on the reconstruction result to obtain the final reconstructed survival curve image data.

5. The method for reconstructing survival curve image data based on deep learning according to claim 1, characterized in that: The S4 specifically includes: S41. Initialize the spider population, set the population size to N, each spider individual represents a hyperparameter combination or network weight combination of the deep learning network model, and the initial state of the i-th individual is: X i ={p1,p2,...,p m }; Among them, X i represents the parameter combination of the ith individual, p j represents the jth hyperparameter or weight of the deep learning network, and m is the total number of parameters; S42, define a fitness function, the fitness function is evaluated based on the reconstruction error of the generated reconstructed survival curve image data, and the fitness value F(X i ) is calculated as: F(X i )=α·L MSE +β·L SSIM ; Among them, F(X i ) is the fitness value of the i-th individual. The smaller the fitness value, the better the individual optimization effect. MSE is the pixel-level mean square error between the reconstructed survival curve image data and the real survival curve image data, L SSIM is the structural similarity measure between the reconstructed survival curve image data and the real survival curve image data, α and β are weight coefficients; S43, based on the local neighborhood information, the position of the spider individual is optimized to calculate the interaction influence between the i-th individual and the neighboring individuals: in, and are the positions of the ith individual in round t and round t+1, respectively, N i is the neighborhood set of the ith individual, w k is the weight of neighborhood individual k, indicating its influence on individual i, and λ1 is the step size factor, which is used to control the update amplitude; S44, perform a global search for weight parameters, and randomly select some individuals to move closer to the global optimal position by simulating the information propagation behavior of the spider web: in, is the individual position with the lowest fitness value in the current iteration, γ is the global search factor, which is used to control the global exploration step size; S45, iteratively update the fitness function value, the iteration termination condition is to reach the maximum number of iterations T or the fitness function converges to the set threshold, and finally output the global optimal parameter combination: Among them, X opt Represents the optimized combination of hyperparameters and network weights of the deep learning network model.

6. The method for reconstructing survival curve image data based on deep learning according to claim 5, characterized in that: The S42 specifically includes: S421, L MSE Represents the mean square error, which is used to measure the average difference between the reconstructed survival curve image data and the real survival curve image data at the pixel level: Among them, y j is the value of the jth pixel in the real survival curve image data, is the value of the jth pixel in the reconstructed survival curve image data, and n is the total number of pixels; S422, L SSIM Represents the structural similarity index, which is used to evaluate the consistency of the visual structure of the reconstructed survival curve image data: in, Represents the real data Y and the reconstructed data The structural similarity index of: Among them, μ Y and Y and The mean of and Y and The variance of For Y and The covariance of , C1 and C2 are constants used to avoid the denominator being zero; S423, based on the reconstructed survival curve image data R generated by the optimized deep learning network model, calculate its fitness value F(X i ), which comprehensively reflects the pixel accuracy and visual structure consistency of the reconstructed data; S424, by adjusting the weight coefficients α and β to control the fitness function for the mean square error L MSE and structural similarity L SSIM sensitivity to meet the optimization requirements of different application scenarios.

7. The method for reconstructing survival curve image data based on deep learning according to claim 1, characterized in that: The S5 specifically includes: S51, the survival curve image data set D to be reconstructed b Input to the optimized deep learning network model; S52, in the parallel multi-scale feature extraction network, extract the multi-scale features of the input survival curve image data based on the convolution operation, and generate a fused multi-scale feature map F ms And the multi-scale feature map F is ms Perform weighted processing to generate an optimized multi-scale feature map F ca ; S53, the optimized multi-scale feature map F ca Input to the residual reconstruction network, the reconstruction network combines multi-layer deconvolution operations with jump connections to generate reconstructed survival curve image data R; S54, the output layer processes the reconstruction result of the last layer to generate the final reconstructed survival curve image data R final , and the final reconstructed survival curve image data R final The output is the reconstruction result of the deep learning network; S55, the generated reconstructed survival curve image data R final Compared with the fitness function evaluation standard, the reconstruction effect is verified, so that the reconstructed survival curve image data meets the consistency requirements of pixel-level accuracy and structural similarity.

Images