Method and device for predicting hemotoxicity of patient after radiotherapy based on machine learning

By constructing a machine learning-based LKB-NTCP model, combined with multi-center data and deep learning technology, the accuracy and universality of predicting hematologic toxicity in rectal cancer radiotherapy were solved, enabling more precise radiotherapy planning and improving the treatment effect and quality of life for rectal cancer patients.

CN119007978APending Publication Date: 2024-11-22FUDAN UNIV SHANGHAI CANCER CENT
View PDF 0 Cites 2 Cited by

Patent Information

Application Number
CN202411006150.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-25
Publication Date
2024-11-22

AI Technical Summary

Technical Problem

Existing radiotherapy planning for rectal cancer mainly relies on physician experience and lacks precise design through artificial intelligence and big data platforms. The Lyman model does not fit well under non-uniform irradiation conditions, and its generalization ability and universality are insufficient. In particular, it cannot effectively predict the risk of hematologic toxicity in short-course radiotherapy for locally advanced rectal cancer.

Method used

By collecting patient information and CT images, the pelvic bone cavity structure was segmented using a 3D U-Net deep learning model. Combined with the LKB-NTCP model based on machine learning, key predictive factors were screened, and a machine learning-based method for predicting blood toxicity was constructed. The LKB model was built using multi-center data, dosimetric parameters were screened, iterative optimization was carried out, and an artificial intelligence platform was established to assist in radiotherapy planning.

Benefits of technology

It improves the accuracy and universality of predicting post-radiotherapy hematologic toxicity, breaks through the traditional planning method that relies on physician experience, provides more precise treatment plan design, enhances the model's generalization ability and robustness, and can better predict the probability of HT events.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119007978A_ABST
    Figure CN119007978A_ABST
Patent Text Reader

Abstract

The invention relates to a machine learning-based post-radiotherapy hematotoxicity prediction method and device for a patient, and the method comprises the steps: collecting clinical information, hematology parameter information and CT images of the patient subjected to radiotherapy and chemotherapy, forming a data set, and carrying out the preprocessing; performing automatic segmentation of each structure of the pelvic cavity and automatic positioning of the pelvis on the CT image so as to calculate a cumulative dose-volume histogram and a differential dose-volume histogram; screening key predictive factors of the dosimetry parameters based on a cumulative dose-volume histogram; according to the differential dose-volume histogram, parameter fitting is carried out on the LKB-NTCP model through a machine learning method; and carrying out iterative optimization on the LKB-NTCP model by taking the obtained key predictive factor, the differential dose-volume histogram and the hematotoxicity result as model input to obtain an optimal LKB-NTCP model for predicting the hematotoxicity of the patient after radiotherapy. Compared with the prior art, the method has the advantages of being high in model generalization, good in universality, high in model robustness and the like.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of clinical auxiliary diagnosis technology based on artificial intelligence, and in particular to a method and device for predicting blood toxicity after patient radiotherapy based on machine learning. Background Art

[0002] Rectal cancer is one of the most common digestive tract malignancies in my country, and neoadjuvant chemoradiotherapy (CRT) is increasingly used in the treatment of rectal cancer. Acute hematotoxicity is the most common complication of preoperative chemoradiotherapy, which may lead to prolonged treatment time and inadequate planned treatment, which is not conducive to local control of the tumor. According to clinical studies, the risk of acute hematotoxicity depends not only on the radiation dose, but also on the chemotherapy regimen. Therefore, for some rectal cancer patients who cannot tolerate surgery and have a strong desire to preserve their anus, how to assist in predicting and evaluating the efficacy of neoadjuvant chemoradiotherapy is of great guiding significance. If the risk of hematotoxicity caused by short-term radiotherapy in patients with rectal cancer can be predicted by this model before the patient receives neoadjuvant treatment, it will be better able to plan precise treatment plans for rectal cancer.

[0003] When rectal cancer patients receive intensity-modulated radiation therapy (IMRT), the normal tissues and organs endangered by radiation are the small intestine, colon, bladder, and femoral head. When planning the treatment of rectal cancer, it is necessary to locate the preoperative and postoperative radiotherapy for rectal cancer, delineate the target area and organs at risk (OARs), set the radiotherapy dose, and summarize the dose distribution and limitations of the organs at risk and the side effects. If the patient's treatment plan is preoperative and postoperative radiotherapy, it is necessary to plan the dose of 95% PTV or pelvic irradiation and shrink the field to the mesorectal area before and after surgery; if the patient's treatment plan is preoperative short-term radiotherapy, it is necessary to plan the minimum acceptable dose of 95% PTV. This ensures that the therapeutic effect of radiotherapy on the target area is maximized while minimizing the impact on the organs at risk. The pelvic bone marrow (PBM) is divided into three subunits: ilium, lower pelvis (LP), and lumbosacral spine (LSS). Therefore, the dose distribution of the PBM and its substructure target volumes and OARs needs to be considered during dose planning.

[0004] Before doctors perform radiotherapy on patients, they mainly design treatment plans based on the patient's personal medical history, cancer staging and other clinical diagnostic conditions. When planning the treatment of locally advanced rectal cancer, doctors mainly consider the treatment effect and normal tissue complications based on their experience. There is still a lack of research related to the precise design of radiotherapy planning based on artificial intelligence big data platforms to accurately give the target area sufficient doses while minimizing damage to surrounding normal tissues, thereby improving patients' tumor control rates, prolonging survival time or improving quality of life.

[0005] In response to this, Lyman proposed the Lyman model in 1985, which characterizes the probability of normal tissue complications by calculating the uniform radiation dose of some organs. et al. proposed that the Lyman model is one of the most commonly used radiobiological models for calculating NTCP, and used the error function curve to approximate the "S"-shaped dose-risk relationship caused by radiation. The NTCP calculation formula based on the Lyman model is as follows:

[0006]

[0007] Among them, TD 50 is the dose corresponding to NTCP with a correlation volume of 50% after irradiation; m is the slope of the NTCP curve; n is the volume-dependent parameter of the complication probability, and the volume effect is maximum when n is close to 1 and minimum when n is close to 0.

[0008]

[0009] However, the Lyman model only considers the case of uniform irradiation, while the irradiation of organs at risk in radiotherapy for locally advanced rectal cancer is non-uniform, so the Lyman model cannot fit the model well.

[0010] In summary, the existing solutions have the following defects:

[0011] The chemoradiotherapy planning for locally advanced rectal cancer relies solely on the physician's experience to determine the possibility of HT events after radiotherapy. This is dependent on the physician's experience and has individual differences. There is a lack of an artificial intelligence model platform for simulation and prediction to assist physician planning.

[0012] The generalization ability of the model is weak: The Lyman model is fitted on the basis of uniform dose to the irradiated organs. However, in the treatment planning of rectal cancer patients undergoing IMRT, the irradiation dose to the organs at risk near the target area is highly uneven. Therefore, the Lyman model cannot be robustly applied to the prediction of HT toxicity in chemoradiotherapy for locally advanced rectal cancer.

[0013] The universality of the model needs to be improved: The HT prediction model for rectal cancer chemoradiotherapy is limited to long-term radiotherapy. The small amount of data for model training and the data source is a single center, resulting in weak model generalization ability. The model may not be extended to other centers. In addition, there is only a normal tissue complication model for HT toxicity in long-term radiotherapy for locally advanced rectal cancer, but a short-term normal tissue complication model with a PBM structure is missing. Summary of the invention

[0014] The purpose of the present invention is to provide a method and device for predicting blood toxicity after radiotherapy for patients based on machine learning in order to overcome the defects of the above-mentioned prior art, namely, weak generalization ability and need to improve universality.

[0015] The purpose of the present invention can be achieved by the following technical solutions:

[0016] A method for predicting blood toxicity after radiotherapy for patients based on machine learning, comprising the following steps:

[0017] Collect information about patients who have undergone radiotherapy and chemotherapy, including clinical information, hematological parameter information, and CT images of the patients, to form a data set, and preprocess the collected data set;

[0018] Automatically segment the pelvic bone cavity structures and automatically locate the pelvis on the CT images in the data set; obtain the dose-volume histogram of the pelvic bone cavity and each substructure of each patient based on the positioning results and the patient information in the data set, and calculate the corresponding metrological evaluation indicators, which include cumulative dose-volume histogram and differential dose-volume histogram;

[0019] Based on the cumulative dose-volume histogram, key predictors of dosimetric parameters are screened; and according to the differential dose-volume histogram, a pre-established LKB-NTCP model is fitted with parameters by a machine learning method;

[0020] The key predictors, differential dose-volume histograms, and blood toxicity results obtained based on the data set were used as model inputs to iteratively optimize the LKB-NTCP model to obtain the optimal LKB-NTCP model.

[0021] The patient data to be predicted is obtained and input into the optimal LKB-NTCP model to obtain the prediction result of the patient's blood toxicity after radiotherapy.

[0022] Furthermore, the clinical information includes name, gender, age, BMI, T stage, N stage and HT grade, and the hematological parameter information includes the lowest value, initial value and grade of white blood cells, neutrophils, hemoglobin and platelets.

[0023] Further, the CT image includes PBM and dose distribution of three substructures, including ilium, lower pelvis and lumbar spine.

[0024] Furthermore, the method uses a 3D U-Net deep learning model to automatically segment each PBM structure and automatically locate the pelvis on the CT image. During the training process of the 3D U-Net deep learning model, CT images are acquired and the results of manual annotation by doctors are used as labels to construct a CT image set, and the 3D U-Net deep learning model is trained. The Dice similarity coefficient, volume similarity and Hausdorff distance are used as indicators to evaluate the segmentation efficiency of the model.

[0025] Furthermore, the preprocessing process of the data set includes: using principal component analysis to reduce the dimension of CT images and clinical information, and determining whether there is a batch effect. If so, the batch effect is corrected by using Z-score standardization and the ComBat algorithm based on the empirical Bayesian method.

[0026] Furthermore, the screening process of the key predictors includes: using the cumulative dose-volume histogram, single-factor logistic regression, multi-factor logistic regression and stepwise logistic regression based on machine learning to screen the key predictors of dosimetric parameters.

[0027] Furthermore, the process of parameter fitting for the LKB-NTCP model is as follows:

[0028] The differential dose-volume histogram is converted into an equivalent uniform dose and a biological equivalent dose. The patient information is used to determine whether the patient belongs to the long-term radiotherapy group or the short-term radiotherapy group. If the patient is in the long-term radiotherapy group, an equivalent uniform dose and 2Gy fractionated biological equivalent dose model are established based on the equivalent uniform dose, and the LKB-NTCP model is fitted with parameters. If the patient is in the short-term radiotherapy group, a biological equivalent dose model is established based on the biological equivalent dose, and the LKB-NTCP model is fitted with parameters.

[0029] Furthermore, in the process of parameter fitting of the LKB-NTCP model, cross-validation is used for hyperparameter selection. The Akaike information criterion is used to weigh the complexity of the estimated model and the goodness of fitting the data. The LKB-NTCP model with the smallest Akaike information criterion value is selected, and a loss function based on probability distribution measurement is introduced for machine learning fitting.

[0030] Furthermore, the method uses maximum likelihood estimation to iteratively optimize the LKB-NTCP model; and uses accuracy, precision, recall, receiver operating characteristic curve and area under the curve to comprehensively evaluate the prediction effect of blood toxicity after radiotherapy in patients.

[0031] Furthermore, the method is used to predict blood toxicity in patients with rectal cancer after radiotherapy.

[0032] The present invention also provides a device for predicting blood toxicity after patient radiotherapy based on machine learning, comprising a memory and a processor, wherein the memory stores a computer program, and the processor calls the computer program to execute the steps of the method described above.

[0033] Compared with the prior art, the present invention has the following advantages:

[0034] At present, the dose planning before radiotherapy in clinical practice mainly relies on doctors to set the dose of organs at risk, which is related to subjective factors such as doctor experience and years of work. In addition, the existing prediction model only focuses on long-term radiotherapy for rectal cancer, and the universality and robustness of the model need to be improved.

[0035] Previous models only focused on long-term radiotherapy for rectal cancer, and the generalization and robustness of the model need to be improved. The present invention uses a large number of PBM and DVH data of short-term radiotherapy of locally advanced rectal cancer patients from multiple centers, screens key predictive factors through diverse data sources and complete model structures, and constructs an LKB model based on the MLE algorithm of machine learning, and then builds an artificial intelligence prediction platform to assist in evaluating the probability of HT events after short-term chemoradiotherapy for locally advanced rectal cancer. Doctors can input the regimen design into the model when performing dose planning, and can predict the probability of HT events after radiotherapy, thereby assisting doctors in adjusting and optimizing the regimen; it has the following advantages:

[0036] (1) Strong model generalization: The short-term radiotherapy for locally advanced rectal cancer that this invention focuses on is a highlight. Compared with other similar models, it has created a wider range of applicable frameworks. It breaks through the traditional method of radiotherapy dose planning based solely on doctor experience, and builds an AI platform from the perspective of machine learning algorithms and artificial intelligence big data to construct a universal HT event prediction model.

[0037] (2) Strong model robustness: The MLE algorithm of machine learning is used to jointly construct the LKB model with data from multiple centers. The ComBat algorithm is used to remove batch effects caused by data sources, target area delineation, CT scanning equipment, etc. Logistic regression is used to fully utilize cDVH to screen key predictors of dosimetric parameters, and dDVH is used to fit the NTCP model to train an artificial intelligence platform that can assist radiotherapists in planning treatment plans. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 Schematic diagram of a process for predicting blood toxicity after radiotherapy for patients based on machine learning provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0039] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Generally, the components of the embodiments of the present invention described and shown in the drawings here can be arranged and designed in various different configurations.

[0040] Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the invention claimed for protection, but merely represents selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0041] It should be noted that similar reference numerals and letters denote similar items in the following drawings, and therefore, once an item is defined in one drawing, further definition and explanation thereof is not required in subsequent drawings.

[0042] Example 1

[0043] like Figure 1 As shown, this embodiment provides a method for predicting blood toxicity after radiotherapy of patients based on machine learning, comprising the following steps:

[0044] S1: Collect information about patients who have undergone radiotherapy and chemotherapy, including clinical information, hematological parameter information, and CT images of patients, form a data set, and preprocess the collected data set;

[0045] S2: Automatically segment the pelvic bone cavity structures and automatically locate the pelvis on the CT images in the data set; obtain the dose-volume histogram of the pelvic bone cavity and each substructure of each patient based on the positioning results and the patient information in the data set, and calculate the corresponding metrological evaluation indicators, which include the cumulative dose-volume histogram and the differential dose-volume histogram;

[0046] S3: Screening key predictors of dosimetric parameters based on the cumulative dose-volume histogram; performing parameter fitting on the pre-established LKB-NTCP model by machine learning method according to the differential dose-volume histogram;

[0047] S4: Using the key predictors, differential dose-volume histograms, and hematologic toxicity results obtained based on the dataset as model inputs, the LKB-NTCP model was iteratively optimized to obtain the optimal LKB-NTCP model;

[0048] S5: Obtain the patient data to be predicted, and input it into the optimal LKB-NTCP model to obtain the prediction result of the patient's blood toxicity after radiotherapy.

[0049] The following is a detailed description of each step:

[0050] 1. Data collection and preprocessing in step S1

[0051] 1.1. Collection of clinical baseline information

[0052] The clinical information includes name, gender, age, BMI, T stage, N stage and HT grade, and the hematological parameter information includes the lowest value, initial value and grade of white blood cells, neutrophils, hemoglobin and platelets.

[0053] In this embodiment, the method is used to predict blood toxicity in patients with rectal cancer after radiotherapy.

[0054] Specifically, the information of patients with locally advanced rectal cancer who received neoadjuvant chemoradiotherapy, followed by total mesorectal excision and adjuvant therapy, and IMRT at the Affiliated Cancer Hospital of Fudan University, the Sixth Affiliated Hospital of Sun Yat-sen University, the West China Hospital of Sichuan University, and the First Affiliated Hospital of Zhengzhou University from 2010 to 2023 was retrospectively collected. Cases with distant metastasis, those who had received anti-tumor drugs, and those with incomplete clinical information or missing DICOM (digital imaging and communications in medicine) images were excluded. The clinical information to be collected was name, gender, age, BMI, T stage, N stage, HT grade, and the lowest value, initial value, and grade of hematological parameters (white blood cells, neutrophils, hemoglobin, and platelets). Among them, HT grade was the endpoint indicator. The enrolled cases were grouped according to the treatment regimen.

[0055] 1.2 Image Collection

[0056] The collected CT images included PBM and dose distribution of three substructures including the ilium, lower pelvis, and lumbosacral spine.

[0057] In this embodiment, CT images of each group of patients were specifically collected, wherein the image information included the dose distribution of the PBM and three substructures (ilium, lower pelvis, and lumbosacral spine), and the volumes receiving 5, 10, 15, 20, 25, 30, 35, 40, and 45 Gy (corresponding to V5, V10, V15, V20, V25, V30, V35, V40, and V45);

[0058] 1.3 Data Preprocessing

[0059] Principal component analysis (PCA) was used to reduce the dimension of CT images and clinical information, and to observe whether there were batch effects caused by factors such as data source, CT scanning equipment, and ROI delineation criteria. Z-score standardization and ComBat algorithm based on empirical Bayesian method were used to correct potential batch effects, and PCA was used again to visualize the removal effect of batch effects.

[0060] 2. CT Image Processing in Step S2

[0061] 2.1 Structural segmentation of CT images

[0062] A 3D U-Net deep learning model is used to automatically segment the PBM structures and automatically locate the pelvis on CT images. During the training process of the 3D U-Net deep learning model, CT images are acquired and the results of manual annotation by doctors are used as labels to construct a CT image set for training the 3D U-Net deep learning model. The Dice similarity coefficient, volume similarity and Hausdorff distance are used as indicators to evaluate the segmentation efficiency of the model.

[0063] The specific processing process in this embodiment is as follows:

[0064] The patients were fixed in the prone position and underwent CT scans at the L3-L4 junction, with a scanning layer thickness of 5 mm, a reconstruction layer thickness of 1-1.25 mm, and a layer spacing of 1-1.25 mm. The original DICOM images of the CT scan of the pelvic bone cavity (PBM) were converted to NIFTI, and a deputy senior physician with 10 years of imaging diagnosis experience used ITK-SNAP software to annotate the PBM and its substructures, ilium, lower pelvis, and lumbosacral spine, layer by layer within the scanning range. Deep learning training was performed on the 3000 image data included in the data. The 3D U-Net deep learning model was used to automatically segment the PBM structures and automatically locate the pelvis on the CT images. The manual annotation results of the doctors were used as labels, and the Dice similarity coefficient (DSC), volume similarity (VS), and Hausdorff distance (HD) of the test set were used as indicators to evaluate the segmentation efficiency of the model.

[0065] 2.2. Volume data collection and collation of PBM structures

[0066] In this embodiment, for each IMRT plan of the enrolled patient, the dose volume histogram (DVH) data of the PBM and each substructure is extracted from the planning system, the dose volume data is extracted using R language, and the dosimetric evaluation index is calculated, which is used as a characteristic parameter to describe the dose situation of the target area and organs at risk (normal tissue). DVH is the most powerful tool for evaluating plan design schemes in radiation physics. The suitability of high-dose areas and target areas can be directly evaluated based on the DVH graph to select a better treatment plan. Among them, DVH is further subdivided into cumulative dose-volume histogram (cDVH) and differential dose-volume histogram (dDVH). The former is used to calculate the dosimetric parameters of PBM and its substructures; the latter is used to fit the NTCP model.

[0067] 3. Model building process in step S3

[0068] 3.1 Dosimetric parameter analysis

[0069] (1) Using machine learning-based univariate logistic regression, multivariate logistic regression, and stepwise logistic regression, cDVH was used to screen key predictors of dosimetric parameters;

[0070] (2) Machine learning modeling of dosimetric parameters and construction of LKB-NTCP model.

[0071] 3.2 NTCP model fitting

[0072] (1) The specific process of parameter fitting for the LKB-NTCP model is as follows:

[0073] The differential dose-volume histogram is converted into an equivalent uniform dose and a biological equivalent dose. The patient information is used to determine whether the patient belongs to the long-term radiotherapy group or the short-term radiotherapy group. If the patient is in the long-term radiotherapy group, an equivalent uniform dose and 2Gy fractionated biological equivalent dose model are established based on the equivalent uniform dose, and the LKB-NTCP model is fitted with parameters. If the patient is in the short-term radiotherapy group, a biological equivalent dose model is established based on the biological equivalent dose, and the LKB-NTCP model is fitted with parameters.

[0074] Specifically, in this embodiment, dDVH is converted into equivalent uniform dose (EUD) and biological equivalent dose (BED). Since the equivalent uniform dose (EUD) has a unique uniform dose distribution with the same number of fractions in the same total time, it produces the same radiobiological effect, and dose uniformity is an important feature of EUD. Therefore, it is necessary to establish EUD and 2Gy fractionated biological equivalent dose (EQD2) models in the long-term radiotherapy group, and to establish a biological equivalent dose (BED) model in the short-term radiotherapy group to correct dDVH. The calculation formula of the EQD2 model is as follows:

[0075]

[0076] For PBM parallel organs, this model sets α / β to 9.

[0077] (2) Machine learning fitting of LKB-NTCP model

[0078] In the process of parameter fitting of the LKB-NTCP model, cross-validation is used for hyperparameter selection. The Akaike information criterion is used to weigh the complexity of the estimated model and the quality of the fitted data. The LKB-NTCP model with the smallest Akaike information criterion value is selected, and a loss function based on probability distribution measurement is introduced for machine learning fitting.

[0079] In this embodiment, the model is machine-learned and fitted based on the LKB model, and TD50, m, and n are respectively used as the endpoints of grade 2 HT toxicity and grade 3 HT toxicity. A 10-fold cross validation is used for hyperparameter selection, and the Akaike information criterion (AIC) is used to weigh the complexity of the estimated model and the goodness of the fitted data. The model with the smallest AIC is selected to improve the model fit (i.e., maximum likelihood), and a loss function based on probability distribution measurement is introduced to minimize the model parameters while reducing overfitting. The formula of the cross entropy loss function is:

[0080]

[0081] (3) Iterative Optimization of LKB-NTCP Model

[0082] The LKB-NTCP model was iteratively optimized using maximum likelihood estimation, and the accuracy, precision, recall, receiver operating characteristic curve and area under the curve were used to comprehensively evaluate the prediction effect of hematotoxicity after radiotherapy.

[0083] Specifically, in this embodiment, in order to find a model with the highest probability of generating the currently observed data, the present invention uses the log-likelihood likelihood estimation in machine learning to construct a likelihood function, and uses the important parameters of the fitted model, dDVH and HT endpoint event labels as modeling inputs. The NTCP model is iteratively optimized by maximum likelihood estimation (Maximum Likelihood Estimate, MLE). The log-likelihood function calculation formula of MLE is as follows:

[0084]

[0085] The accuracy, precision, recall, F1 score and area under the receiver operating characteristic curve (ROC) were used to comprehensively evaluate the model's effect in predicting HT events after chemoradiotherapy for rectal cancer.

[0086] A total of 3,000 patients with locally advanced rectal cancer from the Affiliated Cancer Hospital of Fudan University and 4,000 patients from the Sixth Affiliated Hospital of Sun Yat-sen University were used as training sets, 3,000 patients from the West China Hospital of Sichuan University were used as external validation sets, and the First Affiliated Hospital of Zhengzhou University was used for prospective validation.

[0087] This embodiment also provides a device for predicting blood toxicity after radiotherapy for patients based on machine learning, comprising a memory and a processor, wherein the memory stores a computer program, and the processor calls the computer program to execute the steps of the method for predicting blood toxicity after radiotherapy for patients based on machine learning as described above.

[0088] The preferred specific embodiments of the present invention are described in detail above. It should be understood that a person skilled in the art can make many modifications and changes based on the concept of the present invention without creative work. Therefore, any technical solution that can be obtained by a person skilled in the art through logical analysis, reasoning or limited experiments based on the concept of the present invention on the basis of the prior art should be within the scope of protection determined by the claims.

Claims

1. A method for predicting blood toxicity after radiotherapy in patients based on machine learning, characterized in that: The following steps are involved: Collect information about patients who have undergone radiotherapy and chemotherapy, including clinical information, hematological parameter information, and CT images of the patients, to form a data set, and preprocess the collected data set; Automatically segment the pelvic bone cavity structures and automatically locate the pelvis on the CT images in the data set; obtain the dose-volume histogram of the pelvic bone cavity and each substructure of each patient based on the positioning results and the patient information in the data set, and calculate the corresponding metrological evaluation indicators, which include cumulative dose-volume histogram and differential dose-volume histogram; Screening of key predictors of dosimetric parameters based on cumulative dose-volume histograms; According to the differential dose-volume histogram, a pre-established LKB-NTCP model is fitted with parameters by a machine learning method; The key predictors, differential dose-volume histograms, and blood toxicity results obtained based on the data set were used as model inputs to iteratively optimize the LKB-NTCP model to obtain the optimal LKB-NTCP model. The patient data to be predicted is obtained and input into the optimal LKB-NTCP model to obtain the prediction result of the patient's blood toxicity after radiotherapy.

2. The method for predicting blood toxicity after radiotherapy based on machine learning according to claim 1, characterized in that: The clinical information includes name, gender, age, BMI, T stage, N stage and HT grade, and the hematological parameter information includes the lowest value, initial value and grade of white blood cells, neutrophils, hemoglobin and platelets.

3. The method for predicting blood toxicity after radiotherapy based on machine learning according to claim 1, characterized in that: The CT images include PBM and dose distribution of three substructures, including ilium, lower pelvis, and lumbosacral spine; The method uses a 3D U-Net deep learning model to automatically segment various PBM structures and automatically locate the pelvis on CT images. During the training process of the 3D U-Net deep learning model, CT images are acquired and the results of manual annotation by doctors are used as labels to construct a CT image set, and the 3D U-Net deep learning model is trained. The Dice similarity coefficient, volume similarity and Hausdorff distance are used as indicators to evaluate the segmentation efficiency of the model.

4. The method for predicting blood toxicity after radiotherapy of patients based on machine learning according to claim 1, characterized in that: The preprocessing process of the data set includes: using principal component analysis to reduce the dimension of CT images and clinical information, and determining whether there is a batch effect. If so, Z-score standardization and ComBat algorithm based on empirical Bayesian method are used to correct the batch effect.

5. The method for predicting blood toxicity after radiotherapy of patients based on machine learning according to claim 1, characterized in that: The screening process of the key predictors includes: using the cumulative dose-volume histogram, single-factor logistic regression, multi-factor logistic regression and stepwise logistic regression based on machine learning to screen the key predictors of dosimetric parameters.

6. The method for predicting blood toxicity after radiotherapy of patients based on machine learning according to claim 1, characterized in that: The specific process of parameter fitting for the LKB-NTCP model is as follows: The differential dose-volume histogram is converted into an equivalent uniform dose and a biological equivalent dose. The patient information is used to determine whether the patient belongs to the long-term radiotherapy group or the short-term radiotherapy group. If the patient is in the long-term radiotherapy group, an equivalent uniform dose and 2Gy fractionated biological equivalent dose model are established based on the equivalent uniform dose, and the LKB-NTCP model is fitted with parameters. If the patient is in the short-term radiotherapy group, a biological equivalent dose model is established based on the biological equivalent dose, and the LKB-NTCP model is fitted with parameters.

7. The method for predicting blood toxicity after radiotherapy of patients based on machine learning according to claim 1, characterized in that: In the process of parameter fitting of the LKB-NTCP model, cross-validation is used for hyperparameter selection. The Akaike information criterion is used to weigh the complexity of the estimated model and the quality of the fitted data. The LKB-NTCP model with the smallest Akaike information criterion value is selected, and a loss function based on probability distribution measurement is introduced for machine learning fitting.

8. The method for predicting blood toxicity after radiotherapy of patients based on machine learning according to claim 1, characterized in that: The method uses maximum likelihood estimation to iteratively optimize the LKB-NTCP model; and uses accuracy, precision, recall, receiver operating characteristic curve and area under the curve to comprehensively evaluate the prediction effect of blood toxicity after radiotherapy in patients.

9. The method for predicting blood toxicity after radiotherapy of patients based on machine learning according to claim 1, characterized in that: The method is used to predict blood toxicity in patients with rectal cancer after radiotherapy.

10. A device for predicting blood toxicity after radiotherapy for patients based on machine learning, characterized in that: The method comprises a memory and a processor, wherein the memory stores a computer program, and the processor calls the computer program to execute the steps of any one of the methods according to claims 1 to 9.

Citation Information

Cited By

  • Batch effect analysis method and system based on medical image

    CN119541779A

  • A batch effect analysis method and system based on medical images

    CN119541779B