Kidney cancer risk grade analysis method based on deep learning

Abstract

The invention provides a kidney cancer risk grade analysis method based on deep learning. The method comprises the following steps: acquiring multi-modal medical data of a patient and preprocessing the multi-modal medical data to construct a training sample set; training a risk level prediction model by using the sample set until a preset condition is met; inputting data of a to-be-detected patient into the trained model to obtain a predicted risk level, if a low risk threshold is met, suggesting conventional follow-up visit, otherwise, optimizing treatment scheme parameters by adopting a multi-objective optimization algorithm, evaluating the risk level of each scheme through the model, and calculating a fitness function value in combination with a side effect score and a treatment cost; and finally outputting an optimal scheme as a clinical intervention basis. According to the invention, accurate prediction of the risk level of the kidney cancer can be realized, and the accuracy of clinical decision and the benefit-cost ratio of a treatment scheme are improved.

Description

Technical Field

[0001] This invention relates to the field of medical artificial intelligence technology, and more specifically, to a method for analyzing the risk level of kidney cancer based on deep learning. Background Technology

[0002] In the medical field, early diagnosis and risk assessment of renal cell carcinoma are crucial for patient treatment and prognosis. With the continuous development of medical technology, various techniques such as imaging examinations, pathological analysis, and gene testing are widely used in the diagnosis and risk assessment of renal cell carcinoma. However, most existing renal cell carcinoma risk assessment methods rely on single-modality data, such as analysis based solely on imaging features or clinical indicators. While these methods can provide valuable information to some extent, the lack of comprehensive analysis of multimodal data often makes it difficult to fully and accurately assess the individualized risk level of renal cell carcinoma patients.

[0003] In radiology, traditional CT image analysis relies heavily on physician experience and visual judgment, making it susceptible to subjective influences and unable to fully uncover subtle features hidden within the image data. While pathological analysis provides detailed histological information, its process is complex and its ability to detect minute lesions is limited. Genetic testing, though capable of revealing the molecular characteristics of renal cell carcinoma, often requires specialized knowledge for interpretation, and individual gene information cannot directly reflect a patient's clinical risk level. Furthermore, the selection and optimization of existing treatment methods also face challenges. When developing treatment plans, physicians typically need to comprehensively consider various factors such as the patient's condition, physical status, and the side effects and costs of treatment. However, a tool for systematically optimizing treatment plans is currently lacking, especially when considering individualized patient risk levels.

[0004] In implementing the embodiments of the present invention, the prior art has at least the following problems or defects: First, most existing methods for assessing renal cell carcinoma risk are limited to the analysis of single-modality data, failing to fully utilize the rich information contained in multimodal data, thus limiting the accuracy and comprehensiveness of the assessment results. Second, traditional imaging and pathological analysis methods suffer from strong subjectivity and complex analysis, making it difficult to meet the needs of precision medicine. Furthermore, there is currently a lack of a treatment plan optimization method that can effectively combine factors such as patient risk level, treatment side effects, and costs, failing to provide scientific and personalized decision support for clinical treatment. Therefore, a new method is urgently needed that can comprehensively analyze renal cell carcinoma risk levels using multimodal data and optimize treatment plans based on this analysis. Summary of the Invention

[0005] This invention provides a deep learning-based method for analyzing the risk level of renal cell carcinoma, comprising: Acquire multimodal medical data of patients, and perform preprocessing operations on the multimodal medical data to form a training sample set; The risk level prediction model is trained based on the training sample set to obtain a target risk level prediction model that meets preset conditions. Obtain the target input data corresponding to the current patient to be tested, and input it into the target risk level prediction model to obtain the predicted risk level of the patient; If the predicted risk level meets the low-risk threshold, routine follow-up recommendations are output; if it does not meet the low-risk threshold, a multi-objective optimization algorithm is used to optimize the treatment plan parameters, and the optimized plan is used as the basis for clinical intervention. During the optimization process, the treatment plan parameters represented by each individual are input into the target risk level prediction model to obtain the risk level of each individual. Based on the risk level, treatment side effect score and treatment cost of each individual, an fitness assessment function is generated.

[0006] Furthermore, the multimodal medical data includes the patient's CT image sequences, pathological slide images, genomic mutation characteristics, clinical indicator data, and risk level labels.

[0007] Furthermore, the preprocessing operation includes: The CT image sequence is processed based on wavelet transform algorithm and maximum information coefficient to extract texture feature map and reconstruct it into first fusion feature; A morphological segmentation algorithm was used to separate the tumor region boundary from the pathological slide image and to calculate the morphological parameters of the cell nucleus. Pathogenicity scoring and pathway enrichment analysis were performed on genomic mutation characteristics to generate a vector of mutation functional impact. The first fusion feature, nuclear morphology parameters, mutation function impact vector, clinical indicator data, and risk level labels are used as sample data to form a training sample set.

[0008] Furthermore, the training process includes: The risk level prediction model is trained using the risk level labels in the training sample set as the ground truth and the remaining data as input. When the change in the model's loss function value over multiple consecutive iterations is less than a preset iteration threshold or the total number of iterations reaches a preset upper limit, training is terminated and the target risk level prediction model is obtained.

[0009] Furthermore, the loss function is: Where RMSE is the root mean square error and MAE is the mean absolute error. For pathological grading smoothing loss, These are preset weighting coefficients.

[0010] Furthermore, the pathological grading smoothing loss is calculated as follows: in For the first The risk level probability distribution output by the next iteration As a preset smoothing factor, The iteration number is the number of iterations and the iteration number condition is satisfied.

[0011] Furthermore, the CT image sequence was acquired from multi-phase data of the patient's enhanced abdominal scan.

[0012] Furthermore, the wavelet transform algorithm and the maximum information coefficient processing steps include: Three-dimensional wavelet decomposition is performed on the CT image sequence to generate high-frequency and low-frequency components; the maximum information coefficient between each component and the tumor region is calculated, and components with coefficients greater than a preset correlation threshold are selected to form the key component set. The key component sets are sorted in descending order according to the maximum information coefficient value, and the components in different sorting intervals are assigned a preset decreasing weight; the key component sets are weighted and summed and reconstructed into the first fusion feature.

[0013] Furthermore, the multi-objective optimization algorithm includes: A similar case database is formed by selecting samples from the training sample set whose clinical indicators are more similar to those of the patients to be tested than a preset similarity threshold. The input data to be tested and the treatment plans of similar case databases are used as the initial population individuals; If the initial population size is insufficient, supplementary individuals are generated in the parameter space using a spatial sampling algorithm. The risk level and fitness assessment function value for each individual are calculated using the target risk level prediction model. An elite retention strategy is employed to select individuals whose fitness meets a preset advantage standard for entry into the next generation; Perform crossover and mutation operations to generate offspring populations; The optimization is terminated when the number of iterations reaches the preset optimization depth or the fitness change value is less than the preset convergence threshold, and the individual with the highest fitness is output as the optimization solution.

[0014] Furthermore, the fitness evaluation function is: in This is the normalized value for the risk level. The normalized value for the side effect score. This is the normalized value for treatment costs. The preset weighting coefficients for treatment factors.

[0015] The embodiments of the present invention have at least the following beneficial effects: 1. By fusing multimodal data such as CT image texture features, pathological section cell nuclear morphology parameters, genomic mutation functional influence vectors, and clinical indicators, and using wavelet transform and maximum information coefficient for feature screening and reconstruction, the problem that traditional single-modal analysis methods cannot fully reflect tumor heterogeneity is solved, thus improving the accuracy and robustness of renal cell carcinoma risk level prediction.

[0016] 2. Innovatively, a composite loss function including pathological grading smoothing loss is introduced into the model training. By constraining the difference in the probability distribution of risk level output between adjacent iterations, the problem of fluctuation in prediction results commonly found in traditional deep learning models in tumor grading tasks is effectively alleviated, making the risk level prediction results more consistent with the continuity characteristics of clinical pathology.

[0017] 3. For high-risk patients, a multi-objective optimization plan generation mechanism based on a similar case database is proposed. By jointly optimizing three core clinical indicators—risk level, treatment side effects, and treatment cost—it solves the problem of over-reliance on experience and lack of quantitative assessment in current clinical decision-making, and provides doctors with personalized treatment plan recommendations that take into account both efficacy and safety. Attached Figure Description

[0018] The above and other objects, features, and advantages of exemplary embodiments of the present invention will become readily apparent from the following detailed description taken in conjunction with the accompanying drawings. Several embodiments of the invention are illustrated in the drawings by way of example and not limitation, wherein: Figure 1 This is a flowchart illustrating a deep learning-based method for analyzing the risk level of renal cell carcinoma, as provided in an embodiment of the present invention. Detailed Implementation

[0019] The principles and spirit of the invention will now be described with reference to several exemplary embodiments. It should be understood that these embodiments are provided merely to enable those skilled in the art to better understand and implement the invention, and are not intended to limit the scope of the invention in any way. Rather, these embodiments are provided to make the invention more thorough and complete, and to fully convey the scope of the invention to those skilled in the art.

[0020] Those skilled in the art will recognize that embodiments of the present invention can be implemented as a system, apparatus, device, method, or computer program product. Therefore, the present invention can be specifically implemented in the following forms: entirely hardware, entirely software (including firmware, resident software, microcode, etc.), or a combination of hardware and software.

[0021] It should be noted that the number of any elements in the accompanying drawings is for illustrative purposes only and not as a limitation, and any naming is for distinction only and has no limiting meaning.

[0022] The following is for reference. Figure 1 , Figure 1 This is a flowchart illustrating a deep learning-based method for analyzing renal cell carcinoma risk levels according to an embodiment of the present invention. Figure 1 As shown, a deep learning-based method for analyzing the risk level of kidney cancer includes: S1. Acquire the patient's multimodal medical data, and perform preprocessing operations on the multimodal medical data to form a training sample set; S2. Train the risk level prediction model based on the training sample set to obtain a target risk level prediction model that meets preset conditions. S3. Obtain the input data corresponding to the current patient to be tested, and input it into the target risk level prediction model to obtain the predicted risk level of the patient; S4. If the predicted risk level meets the low-risk threshold, then routine follow-up recommendations are output; if it does not meet the low-risk threshold, then a multi-objective optimization algorithm is used to optimize the treatment plan parameters, and the optimized plan is used as the basis for clinical intervention. S5. During the optimization process, the treatment plan parameters represented by each individual are input into the target risk level prediction model to obtain the risk level of each individual. Based on the risk level, treatment side effect score and treatment cost of each individual, an fitness assessment function is generated.

[0023] The core of the deep learning-based renal cell carcinoma risk level analysis method described in this invention lies in the comprehensive integration of multimodal medical data, including patient CT image sequences, pathological slide images, genomic mutation characteristics, and clinical indicator data, to accurately predict the risk level of renal cell carcinoma. Multimodal medical data refers to patient information obtained from different medical testing methods, covering multiple aspects such as imaging, pathology, genetics, and clinical indicators. By preprocessing this data and forming a training sample set, a model capable of accurately predicting the risk level of renal cell carcinoma can be trained. Preprocessing refers to standardizing and feature extraction of the raw data to make it suitable for model training. The training sample set is the preprocessed dataset used to train the model to learn the relationship between data and risk level. The target risk level prediction model is a deep learning model trained to accurately predict the risk level of renal cell carcinoma, meeting preset conditions such as achieving a certain accuracy rate. The input data to be tested refers to the multimodal medical data of the current patient to be tested, which, after being input into the model, yields the predicted risk level. The low-risk threshold is a pre-set risk level standard used to determine whether a patient belongs to a low-risk group. Multi-objective optimization algorithms are used to optimize treatment plans by comprehensively considering factors such as risk level, treatment side effects, and cost to generate the optimal treatment plan. The fitness evaluation function is a function used to evaluate the merits of a treatment plan; it is calculated based on risk level, treatment side effect score, and treatment cost.

[0024] Specifically, in multimodal medical data, CT image sequences refer to multi-phase data of enhanced abdominal scans of patients, used to reflect the morphology and location information of tumors. Pathological slide images are used to observe the morphological characteristics of tumor cells, such as cell nucleus size and shape. Genomic mutation features refer to mutation information present in the patient's genome, which may be related to the occurrence and development of renal cell carcinoma. Clinical indicator data include the patient's age, gender, and renal function indicators, which can reflect the patient's overall health status. Preprocessing operations include performing wavelet transform and maximum information coefficient processing on CT image sequences to extract texture feature maps and reconstruct them into the first fusion feature; performing morphological segmentation algorithms on pathological slide images to separate tumor region boundaries and calculate cell nucleus morphological parameters; and performing pathogenicity scoring filtering and pathway enrichment analysis on genomic mutation features to generate mutation functional influence vectors. The training sample set consists of the first fusion feature, cell nucleus morphological parameters, mutation functional influence vectors, clinical indicator data, and risk level labels. The training process of the target risk level prediction model uses the risk level labels in the training sample set as ground truth, and the remaining data as input, and trains the model by optimizing uyjjj. The loss function consists of the root mean square error (RMSE), the mean absolute error (MAE), and the pathological grading smoothing loss. The components are: RMSE (Record Value Error) and MAE (Magnitude of Error). RMSE measures the error between the predicted and actual values, while MAE measures the average magnitude of the error. This is used to smooth the probability distribution of risk levels and prevent the model output from jumping too much. Preset weighting coefficients α, β, and γ are used to balance the contributions of different error terms in the loss function. Pathological grading smoothing loss. This is achieved by calculating the difference between the risk level probability distributions of adjacent iterations, where a preset smoothing factor μ is used to adjust the degree of smoothing.

[0025] Preferably, in the preprocessing step of the CT image sequence, a wavelet transform algorithm is used to perform three-dimensional wavelet decomposition on the CT image sequence to generate high-frequency and low-frequency components to extract texture features from the image. The maximum information coefficient is used to calculate the correlation between each component and the tumor region, and components with a correlation greater than a preset correlation threshold are selected to form a key component set. The key component set is arranged in descending order of maximum information coefficient value, and components in different sorting intervals are assigned preset decreasing weights, and reconstructed into the first fusion feature through weighted summation. In the multi-objective optimization algorithm, samples with a similarity greater than a preset similarity threshold to the clinical indicators of the patient to be tested are first selected from the training sample set to form a similar case library. The input data to be tested and the treatment plan in the similar case library are used as the initial population individuals. If the number of individuals in the initial population is less than the preset size, supplementary individuals are generated in the parameter space through a spatial sampling algorithm. The risk level and fitness evaluation function value of each individual are calculated using a target risk level prediction model. The fitness evaluation function F(x) consists of a risk level normalized value R, a side effect score normalized value S, and a treatment cost normalized value C. Preset treatment factor weighting coefficients w1, w2, and w3 are used to balance the importance of different factors in fitness evaluation. An elite retention strategy is used to select individuals with fitness reaching a preset advantage standard to enter the next generation, and crossover and mutation operations are performed to generate the offspring population. Optimization terminates when the number of iterations reaches a preset optimization depth or the fitness change value is less than a preset convergence threshold, and the individual with the highest fitness is output as the optimal solution.

[0026] In some embodiments, the multimodal medical data includes the patient's CT image sequences, pathological slide images, genomic mutation characteristics, clinical indicator data, and risk level labels.

[0027] It should be noted that the multimodal medical data mentioned in this invention refers to patient information obtained from various medical testing methods. This information covers different levels of medical data, including CT image sequences, pathological slide images, genomic mutation characteristics, clinical indicator data, and labeled risk level tags. CT image sequences refer to multi-phase data of the patient's abdominal enhanced scan, used to reflect information such as the morphology, size, and location of the tumor; pathological slide images are obtained by observing the cellular structure of tumor tissue under a microscope, providing morphological characteristics of the cell nucleus; genomic mutation characteristics refer to mutation information related to renal cell carcinoma in the patient's genome, which may affect the occurrence and development of the tumor; clinical indicator data includes the patient's age, gender, renal function indicators, etc., which can reflect the patient's overall health status; risk level tags are pre-labeled categories based on the severity of the patient's condition, used to train the model to identify patients with different risk levels.

[0028] Specifically, CT imaging sequences are obtained through multi-phase contrast-enhanced scanning, providing information on contrast agent enhancement of the tumor at different time points, thus helping to identify the tumor's blood supply characteristics. Pathological slide images, after histological staining, clearly display morphological parameters of tumor cells, such as nuclear size, shape, and mitotic figures; these parameters are crucial for determining the malignancy of the tumor. Genomic mutation characteristics are obtained through high-throughput sequencing technology, including single nucleotide variants (SNVs) and insertion / deletion mutations (InDel), which may be closely related to the pathogenesis of renal cell carcinoma. Clinical indicators encompass the patient's physiological indicators and laboratory test results, such as serum creatinine levels and blood pressure, reflecting the patient's renal function and overall health status. Risk level labels, based on the experience and diagnostic criteria of clinical experts, classify the patient's condition into low-risk, intermediate-risk, and high-risk categories to guide subsequent treatment decisions.

[0029] Preferably, the acquisition of CT image sequences requires the use of specialized medical imaging equipment, such as multi-slice spiral CT, and multi-phase scanning, including arterial, venous, and delayed phases, after intravenous injection of contrast agent to obtain imaging information of the tumor at different time points. Obtaining pathological slide images requires biopsy of the patient's tumor tissue, followed by a series of processes including fixation, dehydration, embedding, sectioning, and staining, before high-resolution images are captured using a microscope. Detection of genomic mutation characteristics requires extracting DNA from the patient's tumor tissue and performing whole-exome sequencing or targeted sequencing using high-throughput sequencing technology to analyze mutations in the genome, and screening for mutation sites related to renal cell carcinoma through bioinformatics analysis. The collection of clinical indicator data requires combining the patient's medical history, physical examination results, and laboratory test reports to ensure the accuracy and completeness of the data. Risk level labeling requires experienced clinicians to make a comprehensive judgment based on the patient's imaging findings, pathological characteristics, gene testing results, and clinical symptoms to ensure the accuracy of the labels. The integration of this multimodal medical data provides comprehensive data support for the accurate prediction of renal cell carcinoma risk levels.

[0030] In some embodiments, the preprocessing operation includes: The CT image sequence is processed based on wavelet transform algorithm and maximum information coefficient to extract texture feature map and reconstruct it into first fusion feature; A morphological segmentation algorithm was used to separate the tumor region boundary from the pathological slide image and to calculate the morphological parameters of the cell nucleus. Pathogenicity scoring and pathway enrichment analysis were performed on genomic mutation characteristics to generate a vector of mutation functional impact. The first fusion feature, nuclear morphology parameters, mutation function impact vector, clinical indicator data, and risk level labels are used as sample data to form a training sample set.

[0031] It should be noted that the preprocessing operations mentioned in this invention are a series of processing steps performed on multimodal medical data, aiming to convert the raw data into a format suitable for model training. Specifically, the preprocessing operations include performing wavelet transform and maximum information coefficient processing on CT image sequences to extract texture feature maps and reconstruct them into first fusion features; performing morphological segmentation algorithms on pathological slide images to separate tumor region boundaries and calculate cell nuclear morphological parameters; and performing pathogenicity scoring filtering and pathway enrichment analysis on genomic mutation features to generate mutation functional influence vectors. These preprocessing steps can effectively extract key information from the data, reduce noise interference, and improve the training effect and prediction accuracy of the model.

[0032] Specifically, the first fusion feature refers to the texture feature map obtained by processing CT image sequences using wavelet transform and maximum information coefficient (MIC). Wavelet transform is a mathematical tool that decomposes a signal into different frequency components, used to extract high-frequency and low-frequency features from images. The MIC is a statistic that measures the correlation between variables, used to screen feature components highly correlated with the tumor region. Nuclear morphology parameters refer to the size, shape, texture, and other features of the cell nucleus extracted from pathological slide images; these parameters reflect the atypia of tumor cells. The mutation functional impact vector is a vector obtained by filtering genomic mutation features through pathogenicity scoring and pathway enrichment analysis, used to represent the potential impact of mutations on tumor occurrence and development. Pathogenicity scoring filtering scores mutations based on a known pathogenicity database to screen for potentially pathogenic mutations; pathway enrichment analysis assesses the impact of mutations on cell function by analyzing their distribution in biological pathways. These preprocessing steps convert data from different modalities into a unified feature representation, facilitating subsequent model training and risk level prediction.

[0033] Preferably, in the preprocessing of CT image sequences, the image sequences are first decomposed into three-dimensional wavelet components to obtain high-frequency and low-frequency components. High-frequency components typically contain edge and texture information, while low-frequency components contain overall structural information. Then, the maximum information coefficient between each component and the tumor region is calculated, and components with coefficients greater than a preset correlation threshold are selected as the key component set. These key components are sorted in descending order according to their maximum information coefficient values ​​and assigned different weights based on the sorting interval. Finally, a weighted summation is used to reconstruct the first fusion feature. In the preprocessing of pathological slide images, morphological segmentation algorithms are used to separate the boundaries of the tumor region from the image, and morphological parameters of the cell nucleus, such as nuclear area, perimeter, and shape factor, are calculated. For the preprocessing of genomic mutation features, potentially pathogenic mutation sites are filtered out using pathogenicity scoring, and pathway enrichment analysis is performed to generate a mutation function influence vector. These preprocessing steps effectively extract key information from multimodal data, providing reliable data support for the accurate prediction of renal cell carcinoma risk levels.

[0034] In some embodiments, the training process includes: The risk level prediction model is trained using the risk level labels in the training sample set as the ground truth and the remaining data as input. When the change in the model's loss function value over multiple consecutive iterations is less than a preset iteration threshold or the total number of iterations reaches a preset upper limit, training is terminated and the target risk level prediction model is obtained.

[0035] It should be noted that the training process mentioned in this invention refers to training the risk level prediction model using preprocessed multimodal medical data, i.e., the training sample set, to accurately predict the individualized risk level of kidney cancer patients. The risk level labels in the training sample set serve as the ground truth for model training, i.e., the known correct answers, while other data, such as imaging features and genetic features, are used as input features to train the model to learn the relationship between input features and risk levels. The training process determines whether to terminate training based on the model's loss function value. Training ends when the change in the loss function value over multiple iterations is less than a preset iteration threshold or the total number of iterations reaches a preset upper limit. The resulting target risk level prediction model can then be used for actual kidney cancer risk level prediction.

[0036] Specifically, the training sample set consists of preprocessed multimodal medical data, including first fusion features, nuclear morphology parameters, mutational functional impact vectors, clinical indicator data, and risk level labels. After standardization and feature extraction, this data better reflects the patient's risk level characteristics. Risk level labels are pre-labeled according to clinical diagnostic criteria to guide model learning. The model's loss function measures the difference between the model's predicted values ​​and the actual values; the smaller the value, the more accurate the prediction. The change in the loss function value refers to the magnitude of change in the loss function value over multiple iterations. The preset iteration threshold is a small value; when the change in the loss function value is less than this threshold, the model has converged, and training can terminate. The total number of iterations is the maximum number of iterations allowed during training to prevent excessively long training periods or getting trapped in local optima.

[0037] Preferably, the risk level prediction model can employ architectures such as Convolutional Neural Networks (CNNs) or Recurrent Neural Networks (RNNs) in deep learning. Taking CNNs as an example, the model construction steps include defining the network structure, initializing network parameters, setting the optimizer and loss function, etc. Input parameters include preprocessed multimodal medical data features, such as first fusion features and cell nuclear morphology parameters. The network structure can include multiple convolutional layers, pooling layers, and fully connected layers for feature extraction and classification. The optimizer can employ algorithms such as Stochastic Gradient Descent (SGD) or Adam to update network parameters to minimize the loss function. The loss function can be Root Mean Square Error (RMSE), Mean Absolute Error (MAE), or Cross-Entropy Loss, the specific choice depending on the model's output type and task requirements. During training, the model learns based on the input features and risk level labels, adjusting network parameters through backpropagation to gradually reduce the loss function value. Training ends when the change in the loss function value is less than a preset iteration threshold, such as 0.001, or when the total number of iterations reaches a preset upper limit, such as 1000. The resulting target risk level prediction model can then be used to predict the risk level of new patient data.

[0038] In some embodiments, the loss function is: Where RMSE is the root mean square error and MAE is the mean absolute error. For pathological grading smoothing loss, These are preset weighting coefficients.

[0039] It should be noted that the loss function mentioned in this invention is used to measure the difference between the predicted values ​​and the actual values ​​of the risk level prediction model, and is an important basis for optimizing model performance. The loss function consists of the root mean square error (RMSE), the mean absolute error (MAE), and the pathological grading smoothing loss. The model is composed of three parts, each evaluating its predictive performance from a different perspective. RMSE measures the mean of the squared differences between predicted and true values, MAE measures the mean of the absolute differences between predicted and true values, and... This is used to smooth the probability distribution of risk levels and prevent the model output from jumping too much. By setting the loss function appropriately, a balance can be achieved between prediction accuracy and output stability.

[0040] Specifically, the root mean square error (RMSE) is the square root of the mean of the sum of squared prediction errors. It can amplify large errors, thus requiring high prediction accuracy from the model. The mean absolute error (MAE), on the other hand, is the mean of the absolute values ​​of prediction errors. It treats all errors equally, preventing a few large errors from excessively affecting the overall assessment. (Pathological grading smoothing loss) This is achieved by calculating the difference in risk level probability distributions between adjacent iterations. The aim is to smooth the model's output and avoid output fluctuations caused by overfitting. Preset weighting coefficients α, β, and γ are used to balance RMSE, MAE, and... These weighting coefficients, which contribute to the loss function, can be adjusted according to the specific circumstances to achieve optimal training results. For example, if higher prediction accuracy is required, the value of α can be increased; if a smoother output is needed, the value of γ can be increased.

[0041] Preferably, the loss function is calculated as follows: First, the RMSE is calculated, which is the square root of the mean of the squares of the differences between the predicted and actual values. Specifically, for each sample, the square of the difference between its predicted and actual values ​​is calculated, then the mean of the squared differences of all samples is taken, and finally the square root is calculated. The MAE is calculated by averaging the absolute differences between the predicted and actual values ​​for each sample. It achieves this by calculating the difference between the risk level probability distributions of adjacent iterations. Specifically, for each iteration's output risk level probability distribution, the difference between it and the distribution of the previous iteration's output is calculated, and these differences are summed with weights. The weight coefficients α, β, and γ can be adjusted based on the validation set performance during model training, for example, by determining the optimal weight combination through cross-validation. The final loss function values ​​are RMSE, MAE, and The weighted sum of the loss function is used to train the model, and the goal is to minimize this loss function value.

[0042] In some embodiments, the pathological grading smoothing loss is calculated as follows: in For the first The risk level probability distribution output by the next iteration As a preset smoothing factor, The iteration number is the number of iterations and the iteration number condition is satisfied.

[0043] It should be noted that the pathological grading smoothing loss mentioned in this invention is a component of the loss function used to optimize a risk level prediction model. Its purpose is to improve the model's stability and generalization ability by constraining the smoothness of the risk level probability distribution output by the model, thus preventing excessive fluctuations during iteration. Specifically, the pathological grading smoothing loss is achieved by calculating the difference between the risk level probability distributions output in adjacent iterations; this difference reflects the continuity and stability of the model's output. By introducing the smoothing loss, prediction accuracy and output stability can be balanced during model training, making the model's output smoother at different iteration stages and avoiding inaccurate predictions caused by local optima.

[0044] Specifically, the calculation of the smoothing loss for pathological grading involves the following key parameters and concepts. First, the risk level probability distribution refers to the predicted probability of a patient's risk level output by the model after each iteration. For example, the model outputs a probability distribution indicating whether a patient belongs to low, medium, or high risk. The smoothing factor is a preset parameter used to adjust the degree of smoothing, typically ranging from 0 to 1. The magnitude of the smoothing factor determines the weight of the difference between the current iteration and the previous iteration in the smoothing loss. When the smoothing factor is large, the difference between the current iteration and the previous iteration contributes more to the smoothing loss, resulting in a smoother model output; when the smoothing factor is small, the model focuses more on the difference between the previous two iterations, resulting in relatively weaker smoothing. The number of iterations refers to the current step in the model training process. To calculate the smoothing loss, at least two iterations' output results are required; therefore, the number of iterations must meet a certain condition, namely, at least 2.

[0045] Preferably, the calculation process of the pathological grading smoothing loss can be further refined. First, for the risk level probability distribution output in each iteration, the absolute difference between it and the output of the previous iteration is calculated. This difference reflects the degree of change between the current iteration and the previous iteration. Then, the absolute difference between the output of the previous iteration and the output of the two iterations before that is calculated, which reflects the degree of change between the two iterations before that. These two differences are multiplied by a smoothing factor and 1 minus the smoothing factor, respectively, and then added together to obtain the pathological grading smoothing loss. The choice of the smoothing factor can be adjusted according to the model's training status and the performance on the validation set. For example, in the early stages of training, a larger smoothing factor can be set to enhance the smoothness of the model output and avoid overfitting; in the later stages of training, the smoothing factor can be appropriately reduced to improve the model's prediction accuracy. In this way, the pathological grading smoothing loss can effectively balance the stability of the model output and the prediction accuracy, allowing the model to gradually converge to the optimal solution during training.

[0046] In some embodiments, the CT image sequence is acquired from multi-phase data of a patient's abdominal enhanced scan.

[0047] It should be noted that the CT image sequences mentioned in this invention are acquired from multi-phase data of enhanced abdominal scans of patients. This multi-phase data refers to the acquisition of tumor imaging characteristics at different stages by injecting contrast agent and taking images at different time points during a CT scan. This multi-phase data provides richer information, such as tumor blood perfusion and contrast agent uptake and elimination rates, which are crucial for assessing the risk level of renal cell carcinoma. By using multi-phase data, the pathological characteristics of the tumor can be more comprehensively reflected, thereby improving the accuracy of the model in predicting the risk level of renal cell carcinoma.

[0048] Specifically, multi-phase data in CT imaging sequences typically include three phases: the arterial phase, the venous phase, and the delayed phase. The arterial phase refers to images taken shortly after contrast agent injection, primarily used to observe the arterial blood supply to the tumor. The venous phase refers to images taken slightly longer after contrast agent injection, used to observe venous return from the tumor. The delayed phase refers to images taken a considerable amount of time after contrast agent injection, used to observe the long-term uptake and excretion of the contrast agent by the tumor. These different phases of imaging data provide information on the dynamic changes of the tumor, helping doctors and models to more accurately determine the nature and risk level of the tumor. Acquiring CT imaging sequences requires specialized medical imaging equipment, such as multi-slice spiral CT, to ensure high resolution and high quality. Furthermore, parameters such as the type of contrast agent, injection rate, and dosage need to be optimized based on the patient's specific situation to ensure that the acquired imaging data accurately reflects the characteristics of the tumor.

[0049] Preferably, the acquisition process of the CT image sequence can be further refined. First, the patient needs to undergo an abdominal contrast-enhanced scan, which is usually performed in the radiology department of a hospital. Before the scan, the patient needs to fast and undergo necessary preparations, such as removing metal objects. Then, a contrast agent, usually iodine, is injected intravenously. The injection rate and dosage of the contrast agent need to be adjusted according to the patient's weight and renal function to ensure that the contrast agent can effectively enter the bloodstream and reach the kidneys. After the contrast agent injection, multi-phase scanning is performed at set time intervals, such as an arterial phase scan 30 seconds after contrast agent injection, a venous phase scan 90 seconds after injection, and a delayed phase scan 180 seconds after injection. The selection of these time points is based on the flow and distribution patterns of the contrast agent in the human body. The acquired multi-phase CT image sequence will contain rich tumor feature information, which will be used in subsequent preprocessing and model training steps to improve the accuracy and reliability of renal cell carcinoma risk level prediction.

[0050] In some embodiments, the wavelet transform algorithm and the maximum information coefficient processing step include: Three-dimensional wavelet decomposition is performed on the CT image sequence to generate high-frequency and low-frequency components; the maximum information coefficient between each component and the tumor region is calculated, and components with coefficients greater than a preset correlation threshold are selected to form the key component set. The key component sets are sorted in descending order according to the maximum information coefficient value, and the components in different sorting intervals are assigned a preset decreasing weight; the key component sets are weighted and summed and reconstructed into the first fusion feature.

[0051] It should be noted that the wavelet transform algorithm and the maximum information coefficient processing step mentioned in this invention are key steps in feature extraction from CT image sequences. The wavelet transform algorithm is a mathematical tool that can decompose a signal into different frequency components. Through three-dimensional wavelet decomposition, high-frequency and low-frequency features in CT images can be extracted. These features correspond to texture details and overall structural information in the images, respectively. The maximum information coefficient is a statistic that measures the correlation between variables and is used to screen feature components that are highly correlated with the tumor region. Through these processing steps, key information in the CT image sequence can be extracted and reconstructed into the first fusion feature, providing high-quality input data for subsequent risk level prediction models.

[0052] Specifically, the application of wavelet transform algorithms in CT image sequence processing includes performing three-dimensional wavelet decomposition on the image sequence to generate high-frequency and low-frequency components. High-frequency components typically contain edge and texture information, while low-frequency components contain overall structural information. The maximum information coefficient is used to calculate the correlation between each component and the tumor region, and components with coefficients greater than a preset correlation threshold are selected to form a key component set. The preset correlation threshold is a set value used to determine which components have a sufficiently high correlation with the tumor region, thus retaining these components for extracting key features. The key component set is sorted in descending order of maximum information coefficient value, and components in different sorting intervals are assigned preset decreasing weights. The weights are set to highlight the influence of important feature components. The key component sets are weighted and summed to reconstruct the first fusion feature. This process integrates the information from multiple components into a comprehensive feature for subsequent model training.

[0053] Preferably, the three-dimensional wavelet decomposition process of the wavelet transform algorithm can be further refined. First, wavelet decomposition is performed on each voxel and three-dimensional pixel in the CT image sequence, decomposing the image into components of different frequencies. For example, Daubechies wavelets or other wavelet bases suitable for medical images can be used for decomposition. The high-frequency and low-frequency components obtained after decomposition reflect the detailed information and overall structural information in the image, respectively. Next, the maximum information coefficient between each component and the tumor region is calculated. The calculation of the maximum information coefficient can be achieved by evaluating the statistical dependence between the component and the tumor region, such as using mutual information or correlation analysis. Components with a maximum information coefficient greater than a preset correlation threshold, such as 0.5, are selected, and these components are considered to have a strong correlation with the tumor region. Then, these key components are sorted in descending order according to their maximum information coefficient values, and different weights are assigned according to the sorting interval. For example, the top 10% of components are given the highest weight, the components ranked 10%-30% are given the second highest weight, and so on. Finally, the weighted components are summed to reconstruct the first fusion feature. This fusion feature can more accurately reflect the imaging characteristics of tumors, providing more valuable input data for risk level prediction models.

[0054] In some embodiments, the multi-objective optimization algorithm includes: A similar case database is formed by selecting samples from the training sample set whose clinical indicators are more similar to those of the patients to be tested than a preset similarity threshold. The input data to be tested and the treatment plans of similar case databases are used as the initial population individuals; If the initial population size is insufficient, supplementary individuals are generated in the parameter space using a spatial sampling algorithm. The risk level and fitness assessment function value for each individual are calculated using the target risk level prediction model. An elite retention strategy is employed to select individuals whose fitness meets a preset advantage standard for entry into the next generation; Perform crossover and mutation operations to generate offspring populations; The optimization is terminated when the number of iterations reaches the preset optimization depth or the fitness change value is less than the preset convergence threshold, and the individual with the highest fitness is output as the optimization solution.

[0055] It should be noted that the multi-objective optimization algorithm mentioned in this invention is an algorithm for optimizing renal cell carcinoma treatment plans. Its core idea is to comprehensively consider multiple factors such as the predicted risk level of renal cell carcinoma patients, treatment side effect scores, and treatment costs, and find the optimal treatment plan through optimization algorithms. Specifically, the algorithm first selects samples from the training sample set whose clinical indicators are more similar to those of the patients to be tested than a preset similarity threshold, forming a similar case library. These similar cases provide a reference basis for optimizing the treatment plan for the patients to be tested. Then, the input data of the patients to be tested and the treatment plans in the similar case library are used as the initial population individuals. Through a series of optimization operations, such as crossover, mutation, and selection, iterative optimization is performed step by step, finally outputting the treatment plan with the highest fitness. The fitness evaluation function is used to evaluate the merits of each treatment plan, providing an evaluation standard for the optimization algorithm by comprehensively considering factors such as risk level, side effects, and costs.

[0056] Specifically, the similar case library in the multi-objective optimization algorithm is composed of samples from the training sample set whose clinical indices are more than a preset similarity threshold with those of the patients to be tested. Similarity can be assessed by calculating the Euclidean distance, cosine similarity, or other similarity measures between clinical indices. The preset similarity threshold is a set value used to determine whether the similarity between samples is high enough to decide whether to include them in the similar case library. The initial population includes the input data of the patients to be tested and the treatment plans from the similar case library; these individuals constitute the starting point of the optimization algorithm. If the number of individuals in the initial population is insufficient for the preset size, supplementary individuals can be generated in the parameter space using a spatial sampling algorithm to ensure population diversity and the effectiveness of the optimization process. The target risk level prediction model is used to calculate the risk level and fitness evaluation function value of each individual; these values ​​reflect the overall effect of the treatment plan. The elite preservation strategy is a selection mechanism used to select individuals with fitness reaching a preset advantage standard to enter the next generation, thereby preserving superior genes. Crossover and mutation operations are common operations in genetic algorithms, used to generate offspring populations, increase population diversity, and avoid the algorithm getting trapped in local optima. The optimization process terminates when the number of iterations reaches the preset optimization depth or the change in fitness is less than the preset convergence threshold, and the individual with the highest fitness is output as the optimization solution. The preset optimization depth and preset convergence threshold are parameters that control the stopping conditions of the optimization process. The optimization depth is usually set based on experience or experimental results, while the convergence threshold is used to determine whether the fitness has stabilized.

[0057] Preferably, the implementation of the multi-objective optimization algorithm can be further refined. First, the construction of the similar case database can be accomplished by calculating the similarity between the clinical indicators of the patient to be tested and each sample in the training sample set. For example, for each clinical indicator, the difference between the patient to be tested and the sample is calculated, and then the differences of all indicators are weighted and summed to obtain a comprehensive similarity score. If the score is greater than a preset similarity threshold, such as 0.8, the sample is included in the similar case database. The construction of the initial population includes the input data of the patient to be tested and the treatment plan in the similar case database. If the number of individuals in the initial population is less than the preset size, such as 100 individuals, supplementary individuals can be generated in the parameter space using spatial sampling algorithms such as Latin hypercube sampling to ensure the diversity of the population. The input parameters of the target risk level prediction model include the multimodal medical data features of the patient to be tested, such as CT image features and genomic features. The model output is the risk level prediction value. The fitness evaluation function can be calculated by weighting and summing the normalized value of the risk level, the normalized value of the side effect score, and the normalized value of the treatment cost according to preset weight coefficients to obtain the fitness value of each individual. The elite retention strategy selects the top 10% of individuals with the highest fitness to directly enter the next generation. Crossover operations can be performed through single-point crossover or uniform crossover, while mutation operations can be achieved by randomly changing certain parameters of individuals. When the number of iterations reaches a preset optimization depth, such as 1000, or the fitness change value is less than a preset convergence threshold, such as 0.001, the optimization process terminates, and the individual with the highest fitness is output as the optimization solution.

[0058] In some embodiments, the fitness evaluation function is: in This is the normalized value for the risk level. The normalized value for the side effect score. This is the normalized value for treatment costs. The preset weighting coefficients for treatment factors.

[0059] It should be noted that the fitness assessment function mentioned in this invention is a key tool for evaluating the merits of treatment regimens. It comprehensively considers multiple factors, including the risk level of renal cell carcinoma patients, treatment side effects, and treatment costs, using a unified evaluation standard to measure the overall effectiveness of different treatment regimens. The purpose of the fitness assessment function is to find a balance between risk level, side effects, and cost, thereby providing clinicians with a scientific and rational basis for decision-making. By optimizing the fitness assessment function, it can be ensured that treatment regimens reduce patient risk while minimizing treatment side effects and economic burden.

[0060] Specifically, the fitness assessment function is in the form of: Fitness Value = Risk Grade Weight × (1 - Risk Grade Normalized Value) - Side Effect Weight × Side Effect Score Normalized Value - Cost Weight × Treatment Cost Normalized Value. Here, the risk grade normalized value converts the patient's risk level into a value between 0 and 1, with a lower value indicating lower risk; the side effect score normalized value converts the potential side effects of the treatment plan into a value between 0 and 1, with a lower value indicating fewer side effects; and the treatment cost normalized value converts the economic cost of the treatment plan into a value between 0 and 1, with a lower value indicating lower cost. The weighting coefficients—risk grade weight, side effect weight, and cost weight—are preset parameters used to balance the importance of different factors in fitness assessment. These weights can be adjusted according to clinical needs and the patient's specific situation. For example, if more emphasis is placed on reducing risk, the risk grade weight can be increased; if more emphasis is placed on reducing side effects, the side effect weight can be increased.

[0061] Preferably, the construction and calculation process of the fitness assessment function can be further refined. First, the risk level normalization value can be calculated by mapping the predicted risk level to a range of 0 to 1, for example, low risk is mapped to 0.1, medium risk to 0.5, and high risk to 0.9. The side effect score normalization value can be obtained by standardizing existing side effect score data, for example, mapping the minimum side effect score to 0 and the maximum to 1. The treatment cost normalization value can be obtained by standardizing treatment costs, for example, mapping the minimum cost to 0 and the maximum cost to 1. The weighting coefficients can be set according to clinical experience and the specific needs of patients, for example, the risk level weight can be set to 0.5, the side effect weight can be set to 0.3, and the cost weight can be set to 0.2. These weighting coefficients need to be adjusted according to the actual application scenario to ensure that the fitness assessment function can accurately reflect the comprehensive benefits of the treatment plan. When calculating the fitness value, the above normalization values ​​are substituted into the fitness assessment function formula, and the fitness value of each treatment plan is obtained by weighted summation. A higher fitness score indicates that the treatment plan is more effective after taking into account risks, side effects, and costs.

[0062] The above embodiments of the present invention have the following beneficial effects: 1. By fusing multimodal data such as CT image texture features, pathological section cell nuclear morphology parameters, genomic mutation functional influence vectors, and clinical indicators, and using wavelet transform and maximum information coefficient for feature screening and reconstruction, the problem that traditional single-modal analysis methods cannot fully reflect tumor heterogeneity is solved, thus improving the accuracy and robustness of renal cell carcinoma risk level prediction. 2. Innovatively, a composite loss function including pathological grading smoothing loss is introduced into the model training. By constraining the difference in the probability distribution of risk level output between adjacent iterations, the problem of fluctuation in prediction results commonly found in traditional deep learning models in tumor grading tasks is effectively alleviated, making the risk level prediction results more consistent with the continuity characteristics of clinical pathology. 3. For high-risk patients, a multi-objective optimization plan generation mechanism based on a similar case database is proposed. By jointly optimizing three core clinical indicators—risk level, treatment side effects, and treatment cost—it solves the problem of over-reliance on experience and lack of quantitative assessment in current clinical decision-making, and provides doctors with personalized treatment plan recommendations that take into account both efficacy and safety.

[0063] Furthermore, the storage medium in the embodiments of this application stores program instructions capable of implementing all the above methods. These program instructions can be stored in the storage medium in the form of a software product, including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks, or terminal devices such as computers, servers, mobile phones, and tablets.

[0064] The above description is merely an explanation of some preferred embodiments of the present invention and the technical principles employed. Those skilled in the art should understand that the scope of the invention as described in the embodiments of the present invention is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the above-described inventive concept. For example, technical solutions formed by substituting the above-described features with (but not limited to) technical features with similar functions disclosed in the embodiments of the present invention.

Claims

1. A method for analyzing the risk level of renal cell carcinoma based on deep learning, characterized in that, include: Acquire multimodal medical data of patients, and perform preprocessing operations on the multimodal medical data to form a training sample set; The risk level prediction model is trained based on the training sample set to obtain a target risk level prediction model that meets preset conditions. Obtain the target input data corresponding to the current patient to be tested, and input it into the target risk level prediction model to obtain the predicted risk level of the patient; If the predicted risk level meets the low-risk threshold, then a routine follow-up recommendation is output. If the low-risk threshold is not met, a multi-objective optimization algorithm is used to optimize the treatment plan parameters, and the optimized plan is used as the basis for clinical intervention. During the optimization process, the treatment plan parameters represented by each individual are input into the target risk level prediction model to obtain the risk level of each individual. Based on the risk level, treatment side effect score and treatment cost of each individual, an fitness assessment function is generated.

2. The method according to claim 1, characterized in that, The multimodal medical data includes the patient's CT image sequences, pathological slide images, genomic mutation characteristics, clinical indicator data, and risk level labels.

3. The method according to claim 2, characterized in that, The preprocessing operations include: The CT image sequence is processed based on wavelet transform algorithm and maximum information coefficient to extract texture feature map and reconstruct it into first fusion feature; A morphological segmentation algorithm was used to separate the tumor region boundary from the pathological slide image and to calculate the morphological parameters of the cell nucleus. Pathogenicity scoring and pathway enrichment analysis were performed on genomic mutation characteristics to generate a vector of mutation functional impact. The first fusion feature, nuclear morphology parameters, mutation function impact vector, clinical indicator data, and risk level labels are used as sample data to form a training sample set.

4. The method according to claim 3, characterized in that, The training process includes: The risk level prediction model is trained using the risk level labels in the training sample set as the ground truth and the remaining data as input. When the change in the model's loss function value over multiple consecutive iterations is less than a preset iteration threshold or the total number of iterations reaches a preset upper limit, training is terminated and the target risk level prediction model is obtained.

5. The method according to claim 4, characterized in that, The loss function is: Where RMSE is the root mean square error and MAE is the mean absolute error. For pathological grading smoothing loss, These are preset weighting coefficients.

6. The method according to claim 5, characterized in that, The pathological grading smoothing loss is calculated as follows: in, For the first The risk level probability distribution output by the next iteration As a preset smoothing factor, The iteration number is given and the iteration number condition is satisfied. Let i be the preset smoothing factor, and let i be the iteration number, satisfying the iteration number condition.

7. The method according to any one of claims 1-6, characterized in that, The CT image sequence was acquired from multi-phase data of the patient's abdominal enhanced scan.

8. The method according to claim 3, characterized in that, The wavelet transform algorithm and the maximum information coefficient processing steps include: Three-dimensional wavelet decomposition is performed on the CT image sequence to generate high-frequency and low-frequency components; the maximum information coefficient between each component and the tumor region is calculated, and components with coefficients greater than a preset correlation threshold are selected to form the key component set. The key component sets are sorted in descending order according to the maximum information coefficient value, and the components in different sorting intervals are assigned a preset decreasing weight; the key component sets are weighted and summed and reconstructed into the first fusion feature.

9. The method according to claim 1, characterized in that, The multi-objective optimization algorithm includes: A similar case database is formed by selecting samples from the training sample set whose clinical indicators are more similar to those of the patients to be tested than a preset similarity threshold. The input data to be tested and the treatment plans of similar case databases are used as the initial population individuals; If the initial population size is insufficient, supplementary individuals are generated in the parameter space using a spatial sampling algorithm. The risk level and fitness assessment function value for each individual are calculated using the target risk level prediction model. An elite retention strategy is employed to select individuals whose fitness meets a preset advantage standard for entry into the next generation; Perform crossover and mutation operations to generate offspring populations; The optimization is terminated when the number of iterations reaches the preset optimization depth or the fitness change value is less than the preset convergence threshold, and the individual with the highest fitness is output as the optimization solution.

10. The method according to claim 9, characterized in that, The fitness evaluation function is: in This is the normalized value for the risk level. The normalized value for the side effect score. This is the normalized value for treatment costs. The preset weighting coefficients for treatment factors.

Images