Glioma H3K27M change prediction model construction method and system

Abstract

The invention discloses a glioma H3K27M change prediction model construction method and a glioma H3K27M change prediction model construction system, and belongs to the technical field of diffuse midline glioma gene prediction models, and the glioma H3K27M change prediction model construction method comprises the following steps: constructing a prediction model in combination with CNN and Transformer; mRI data are acquired, and the MRI data are preprocessed; taking the preprocessed MRI data as input data, and pre-training an encoder part of the prediction model by using a mask automatic encoder to obtain a first model; training label data for executing the classification task by using a cooperative training method to obtain a second model; and accessing the first model to the second model by using a transfer learning method, and training to obtain a glioma H3K27M change prediction model. According to the method, the prediction precision of the model on the glioma H3K27M change is remarkably improved by combining the CNN and the Transform architecture.

Description

Technical Field

[0001] The present invention relates to the technical field of diffuse midline glioma gene prediction models, and more specifically to a method and system for constructing a glioma H3K27M change prediction model. Background Art

[0002] Diffuse midline glioma (DMG) is a highly malignant brain tumor, in which H3K27M mutation has a significant impact on the patient's prognosis, and targeted therapy for this mutation is expected to bring survival benefits to patients. Currently, there are many limitations to the methods for detecting H3K27M mutations. Although cerebrospinal fluid examination can determine whether H3K27M is mutated, the accuracy rate is only about 10%. Although histological biopsy is the main means of detection, it requires craniotomy and is an invasive and high-risk examination.

[0003] In deep learning research related to glioma classification, the classification task is often divided into two steps: first segment the tumor area, and then classify the segmented area. However, these two steps are independent of each other and do not consider the correlation between tasks. The segmentation task locates the tumor by distinguishing whether the pixel belongs to normal brain tissue or tumor, while the classification task aims to classify the H3K27M mutant genotype from the representation of the located tumor. The two are closely related, and the homogeneous representation information in the H3K27M mutant genotype is also conducive to the tumor segmentation task.

[0004] At present, there are studies based on head MRI to predict the mutation of H3K27M molecular marker in DMG. The traditional machine learning model constructed by Kandemirli et al. uses 5 imaging features extracted from traditional MRI sequences to classify H3K27M, with AUC values ​​of 0.95 and 0.90 for the training set and test set, but the prediction accuracy is not reported. Pan et al. combined 36 MRI image features and 3 clinical information features of 151 patients in a single center to generate a prediction model with an accuracy of 84.44%. In 2020, a single-center machine learning model released by a certain hospital included 122 patients, trained 10 models to select the best model, and the prediction accuracy of the validation set was 85.5%. However, these studies have shortcomings: the images input to the model need to be manually segmented in advance, which is time-consuming and labor-intensive, and subjective operations lead to individual model differences; the model is not end-to-end and difficult to be actually applied in clinical practice; the generalization has not been verified in an independent external test set; the traditional machine learning method only selects some image features, which is easy to lose important features.

[0005] Therefore, how to provide a prediction model construction method and system for accurately predicting H3K27M mutations is an urgent problem that technicians in this field need to solve. Summary of the invention

[0006] In view of this, the present invention provides a method and system for constructing a prediction model for H3K27M changes in glioma, which constructs an automated prediction model for H3K27M mutations in DMG based on deep learning and whole-brain MRI, and can accurately predict H3K27M mutations.

[0007] In order to achieve the above object, the present invention adopts the following technical solution: In one aspect, the present invention provides a method for constructing a glioma H3K27M change prediction model, comprising: Combine CNN and Transformer to build a prediction model; Acquiring MRI data, and preprocessing the MRI data; Using the preprocessed MRI data as input data, pre-training the encoder part of the prediction model by using masked automatic encoding to obtain a first model; Use the collaborative training method to train the labeled data for performing the classification task to obtain the second model; The first model was connected to the second model using the transfer learning method, and training was performed to obtain a glioma H3K27M change prediction model.

[0008] Preferably, after obtaining the glioma H3K27M change prediction model, the method further comprises: The glioma H3K27M change prediction model is evaluated to obtain an evaluation result, and according to the evaluation result, the model with the best model performance is determined as the final glioma H3K27M change prediction model.

[0009] Preferably, a prediction model is constructed by combining CNN and Transformer, including: Encoder, used for feature extraction; A segmentation head, used for upsampling the features extracted by the encoder and segmenting the MRI data; The classification head identifies H3K27M mutations in the segmented MRI data; The attention layer is used to realize information exchange between the segmentation head and the classification head through the attention mechanism.

[0010] Preferably, the preprocessed MRI data is used as input data, and a masked autoencoder is used to pre-train the encoder part of the prediction model to obtain a first model, comprising: The encoder part of the prediction model is pre-trained using MRI data of a healthy head. In the pre-training, the Adam optimizer is used. In the pre-training stage, the mean square error between the masked reconstructed image block and the original image block is used as the loss function; After completing the model pre-training, various parameters in the encoder are fixed to obtain the first model.

[0011] Preferably, a collaborative training method is used to train the label data for performing the classification task to obtain a second model: The features X3, X4, and X5 extracted from the three parts of the segmentation head are given different weights through the attention mechanism. Then the three extracted features X3, X4, and X5 are concatenated into a new feature X, which is input into the classification head to finally obtain the classification result.

[0012] In another aspect, the present invention provides a system for constructing a glioma H3K27M change prediction model, comprising: Modeling module, used to build prediction models by combining CNN and Transformer; A data acquisition module, used for acquiring MRI data and preprocessing the MRI data; A first training module is used to pre-train the encoder part of the prediction model using the pre-processed MRI data as input data using a masked autoencoder to obtain a first model; A second training module is used to train the label data for performing the classification task using a collaborative training method to obtain a second model; The third training module is used to use the transfer learning method to connect the first model to the second model, and train it to obtain a glioma H3K27M change prediction model.

[0013] It can be seen from the above technical solutions that, compared with the prior art, the present invention discloses a method and system for constructing a glioma H3K27M change prediction model. By combining the CNN and Transformer architectures, the convolutional neural network's ability to extract local features and the Transformer's ability to model global features are fully utilized, significantly improving the model's prediction accuracy for glioma H3K27M changes. The use of a masked autoencoder to pre-train the encoder part can effectively extract key features from MRI data, while reducing dependence on a large amount of labeled data and improving the generalization ability of the model. The weighted splicing of features through collaborative training methods and attention mechanisms further optimizes the classification performance of the model and improves the model's ability to respond to different features. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0015] Figure 1The overall schematic diagram of the model is shown in Figure 2.

[0016] Figure 2 A detailed overall structure diagram of the model.

[0017] Figure 3 The detailed structure diagram of Transformer, resblock1 and resblock2 modules.

[0018] Figure 4 Construct a flow chart for the dataset in this invention. DETAILED DESCRIPTION

[0019] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. 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.

[0020] The present invention discloses a method for constructing a glioma H3K27M change prediction model. Figure 1 As shown, including: Combine CNN and Transformer to build a prediction model. Figure 1 The upper middle part shows the MAE pre-training process. Figure 1 The lower part of indicates that the downstream classification task starts after pre-training.

[0021] MAE is a new generative self-supervised learning model. It has an asymmetric encoder-decoder structure. Its mechanism is to randomly mask some pixel blocks and then let the model use the remaining pixel blocks to reconstruct the randomly masked pixel blocks. This forces the model to learn the features in the image and increases the model's adaptability to tasks in the same scene.

[0022] At present, convolutional neural network (CNN) is widely used in the field of computer vision. Its convolution operation is only applicable to local neighboring areas, which is effective for extracting local features, but it is difficult to extract long-distance interactive features, which affects the accuracy of classification to a certain extent. Therefore, the present invention introduces Transformer into the model, which achieves significant performance improvement in global feature extraction through the self-attention mechanism compared with CNN. Therefore, the present invention combines CNN and Transformer to fully utilize the ability of CNN to extract local features and the advantages of Transformer in global feature extraction.

[0023] In the embodiment of the present invention, a branch task - tumor automatic segmentation task - is added on the basis of the main task (classification task), and a collaborative training method is adopted. Since segmentation and classification are two related tasks, the features in the classification head and the segmentation head are communicated with each other to help each other learn during training, and the attention mechanism is used to focus on the features that can improve the classification accuracy of the model. The segmentation head branch is added to take advantage of the collaborative training to improve the accuracy of H3K27M mutation classification.

[0024] Acquire MRI data and pre-process the MRI data; Using the preprocessed MRI data as input data, the encoder part of the prediction model is pre-trained using a masked autoencoder to obtain a first model; Use the collaborative training method to train the labeled data for performing the classification task to obtain the second model; The first model was connected to the second model using the transfer learning method, and training was performed to obtain a glioma H3K27M change prediction model.

[0025] Furthermore, after obtaining the glioma H3K27M change prediction model, it also includes: The glioma H3K27M change prediction model was evaluated to obtain evaluation results, and the model with the best model performance was determined as the final glioma H3K27M change prediction model based on the evaluation results.

[0026] Specifically, Figure 2-3 As shown, a prediction model is constructed by combining CNN and Transformer, including: Encoder, used for feature extraction; The segmentation head is used to upsample the features extracted by the encoder and segment the MRI data; The classification head identifies H3K27M mutations in the segmented MRI data; The attention layer is used to realize information exchange between the segmentation head and the classification head through the attention mechanism.

[0027] Preferably, the preprocessed MRI data is used as input data, and a masked autoencoder is used to pre-train the encoder part of the prediction model to obtain a first model, comprising: The encoder part of the prediction model is pre-trained using MRI data of healthy heads. In the pre-training, the Adam optimizer is used. In the pre-training stage, the mean square error between the masked reconstructed image block and the original image block is used as the loss function; After completing the model pre-training, various parameters in the encoder are fixed to obtain the first model.

[0028] Preferably, a collaborative training method is used to train the label data for performing the classification task to obtain a second model: The features X3, X4, and X5 extracted from the three parts of the segmentation head are given different weights through the attention mechanism. Then the three extracted features X3, X4, and X5 are concatenated into a new feature X, which is input into the classification head to finally obtain the classification result.

[0029] Further, such as Figure 4 As shown, MRI data was obtained and preprocessed, including: In this study, all patients were from West China Hospital of Sichuan University (training set) and Chengdu Shangjin Nanfu Hospital (test set) between February 2016 and April 2022. In the embodiment of the present invention, patients newly diagnosed with DMG and completed H3K27M status detection were screened from patients who underwent tumor resection or biopsy. The exclusion criteria are as follows: (1) history of surgery, radiotherapy or chemotherapy before surgery; (2) preoperative patients lack T1C or T2 weighted sequences; (3) preoperative MRI has artifact interference. The study of the present invention collected T1C and T2 weighted sequences in the patient's preoperative MRI because the imaging features of these two sequences are the most effective in distinguishing the H3K27M mutation status (5,13,14).

[0030] The present invention collected a total of 75,365 MRI images of approximately 600 healthy subjects from the public dataset IXI dataset (https: / / brain-development.org / ) for model pre-training.

[0031] MRI data preparation: In the training set, MRI was performed on a 3.0-T MRI scanner (Philips Achieva; GE MR 750 W; Siemens Healthcare) or a 1.5-T MRI scanner (Toshiba Medical Systems; Alltech Medical Systems). The main MRI scanning protocols included: axial T1C sequence (repetition time [TR]: 1550 ms, echo time [TE]: 1.98 ms, slice thickness [ST]: 5-6 mm); axial T2 sequence (TR: 4500 ms, TE: 105 ms, ST: 5-6 mm). In the test set, MRI was mainly performed on a 3.0-T (uMR780 3.0T, UIH) or 1.5-T (Achieva 1.5 T, Philips Medical Systems) MRI scanner. The main MRI scanning schemes include the following sequences: axial T1C (TR\TE: 151.8\2.40 ms, ST: 5-6 mm) and axial T2 sequence (TR\TE: 3943\100 ms, ST: 5.5-6 mm). The detailed MRI equipment information and scanning schemes in the training set and test set are shown in Tables 1 and 2, respectively.

[0032] Table 1 Detailed display of commonly used MRI image acquisition parameters in the training set

[0033] Note: TR: repetition time, repetition time; TE: echo time, echo time Table 2 Detailed display of commonly used MRI image acquisition parameters in the test set

[0034] Note: TR: repetition time, repetition time; TE: echo time, echo time The regions of interest (ROI) of each image include the tumor enhancement area in the T1C sequence and the tumor edema area in the T2-weighted sequence. A neurosurgeon with 5 years of experience used LabelMe 3.16.2 version software to manually segment and annotate each slice on the axial MRI. After the annotation was completed, a neurosurgeon with 10 years of clinical experience checked the accuracy of the segmentation. All clinical information of the case was hidden from the annotator during the annotation. It took about 5 months to annotate the ROI in this embodiment.

[0035] In order to verify whether MAE and collaborative training can improve the classification accuracy of the model, an ablation experiment was designed in the embodiment of the present invention, and the following six models were trained: Model 1: It consists of the basic framework of a neural network and only performs classification tasks, not side tasks.

[0036] Model 2: Add MAE pre-training on the basis of Model 1, only perform classification tasks, and do not perform side tasks.

[0037] Model 3: Based on Model 1, a side task is added, and both classification and side tasks are performed at the same time. However, there is no information exchange between classification and side tasks during the training process.

[0038] Model 4: Based on Model 3, an attention mechanism is added between the branch task and the classification task to realize information exchange between the branch task and the classification task.

[0039] Model 5: MAE pre-training is added based on Model 3.

[0040] Model 6: Add MAE pre-training based on Model 4.

[0041] The probability threshold for accuracy calculation is set to 0.5, that is, when the final probability calculated for the image is >0.5, the image is classified as H3K27M mutant type, and when the final probability calculated for the image is ≤0.5, the image is classified as wild type. The patient's H3K27M type is determined based on the average mutation probability of all MRI images of the patient. Accuracy, sensitivity, and specificity are used to describe the performance of the classification model. The formula is defined as follows: ; ; ; In the formula, TP represents the number of true positive samples in the diagnostic results, FP represents the number of false positive samples, TN represents the number of true negative samples, and FN represents the number of false negative samples. The receiver operating characteristic (ROC) curves and precision-recall curves (PRCs) were plotted for each fold of the model using the Matplotlib library, and the area under the curve (AUC) was calculated. When AUC = 1, it is considered a perfect classifier, and there is at least one threshold for the model to obtain perfect predictions. When 0.5 < AUC < 1, it is considered that the classification performance of the model is better than random guessing, and if the threshold is properly set, the model can have predictive value. When AUC = 0.5, it is considered that the model performance is the same as random guessing and has no predictive value. SPSS version 25.0 software was used for statistical analysis, and when P < 0.05, it was considered that the difference was statistically significant. The Dice similarity coefficient (DSC) was used to evaluate the performance of the segmentation task, and this coefficient reflects the amount of spatial overlap between the segmentation image blocks automatically generated by the model and the true segmentation image blocks outlined by neurosurgeons. The DSC formula is defined as follows: ; where A and B represent the labels segmented by the model and the true labels respectively, ( ) represents the intersection of the two sets of labels, represents the total number of the two sets of labels. The value range of DSC is between 0 and 1. The closer it is to 1, the greater the coincidence degree of A and B. If it is equal to 1, it means that A and B completely coincide.

[0042] In the present invention, the classification accuracy, sensitivity, and specificity of the model trained using the T1C sequence are shown in Table 3. The classification sensitivity and specificity of the model trained using the T2-weighted sequence are shown in Table 4, and the accuracy and DSC are shown in Table 5. By comparison, it is found that the model trained with T2 is superior to the model trained with T1C in all aspects. To save space, only the model based on the T2 sequence will be introduced below.

[0043] Table 3 Performance display of all models for predicting H3K27M mutations based on the T1C sequence

[0044] Table 4 Sensitivity and specificity results of all models for predicting H3K27M mutations based on the T2 sequence

[0045] Table 5 Accuracy and DSC results of all models for predicting H3K27M mutations based on the T2 sequence

[0046] By comparison, it was found that model 6 had the highest prediction accuracy for H3K27M genotype, reaching 0.905, with a 95% confidence interval (95% CI) of 0.853-0.961. In the external test set, its classification accuracy was also the highest, at 0.851 (95% CI, 0.802-0.906). Although the baseline model had good prediction accuracy in the training set (0.855, 95% CI, 0.805-0.897), its prediction accuracy in the test set was very low (0.697, 95% CI, 0.641-0.753). In the subsequent model construction, MAE pre-training and dual-task collaborative training were added, and the generalization performance of the model was significantly improved. Compared with model 1, model 2 using MAE increased the classification accuracy by 2.5% in the training set and 3.4% in the external test set. After adding MAE pre-training, the classification accuracy of Model 4 increased by 3% in the training set and 3.4% in the external test set. These results show that MAE pre-training can improve the accuracy of classification predictions and the generalization performance of the model. Compared with Models 3 and 5, after adding dual-task co-training, the classification accuracy of the model in the external test set also increased by 1.1%-2.3%, indicating that the co-training of segmentation and classification tasks can improve the accuracy and generalization of deep learning models for H3K27M genotype classification. In the external test set, the DSC of Model 6 was 0.785 (95% CI, 0.759-0.811), which was 4.2% higher than that of Model 3.

[0047] In Table 6, the 5-fold cross validation results of Model 6 on the training set and test set are detailed. Compared with the training set, the specificity of the test set has almost no decrease (0.932 vs 0.933), while the accuracy of the external test set has decreased (0.905 vs 0.851). However, the accuracy of 0.851 in the external test set is also very high. From Table 6, it can be concluded that the model not only has good prediction accuracy, but also has good generalization performance as verified in the external test set.

[0048] Table 6 Cross-validation results of model 6 in predicting H3K27M mutations in the training set and test set

[0049] In another aspect, the present invention provides a system for constructing a glioma H3K27M change prediction model, comprising: Modeling module, used to build prediction models by combining CNN and Transformer; A data acquisition module, used for acquiring MRI data and preprocessing the MRI data; A first training module is used to pre-train the encoder part of the prediction model using the pre-processed MRI data as input data using a masked autoencoder to obtain a first model; A second training module is used to train the label data for performing the classification task using a collaborative training method to obtain a second model; The third training module is used to connect the first model to the second model using the transfer learning method, and train to obtain a glioma H3K27M change prediction model.

[0050] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0051] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for constructing a glioma H3K27M change prediction model, characterized in that: include: Combine CNN and Transformer to build a prediction model; Acquiring MRI data, and preprocessing the MRI data; Using the preprocessed MRI data as input data, pre-training the encoder part of the prediction model by using masked automatic encoding to obtain a first model; Use the collaborative training method to train the labeled data for performing the classification task to obtain the second model; The first model was connected to the second model using the transfer learning method, and training was performed to obtain a glioma H3K27M change prediction model.

2. The method for constructing a glioma H3K27M change prediction model according to claim 1, characterized in that: After obtaining the glioma H3K27M change prediction model, it also includes: The glioma H3K27M change prediction model is evaluated to obtain an evaluation result, and according to the evaluation result, the model with the best model performance is determined as the final glioma H3K27M change prediction model.

3. The method for constructing a glioma H3K27M change prediction model according to claim 1, characterized in that: Combining CNN and Transformer to build a prediction model, including: Encoder, used for feature extraction; A segmentation head, used for upsampling the features extracted by the encoder and segmenting the MRI data; The classification head identifies H3K27M mutations in the segmented MRI data; The attention layer is used to realize information exchange between the segmentation head and the classification head through the attention mechanism.

4. The method for constructing a glioma H3K27M change prediction model according to claim 1, characterized in that: The preprocessed MRI data is used as input data, and a masked autoencoder is used to pretrain the encoder part of the prediction model to obtain a first model, including: The encoder part of the prediction model is pre-trained using MRI data of a healthy head. In the pre-training, the Adam optimizer is used. In the pre-training stage, the mean square error between the masked reconstructed image block and the original image block is used as the loss function; After completing the model pre-training, various parameters in the encoder are fixed to obtain the first model.

5. The method for constructing a glioma H3K27M change prediction model according to claim 1, characterized in that: Use the collaborative training method to train the labeled data for the classification task and obtain the second model: The features X3, X4, and X5 extracted from the three parts of the segmentation head are given different weights through the attention mechanism. Then the three extracted features X3, X4, and X5 are concatenated into a new feature X, which is input into the classification head to finally obtain the classification result.

6. A glioma H3K27M change prediction model construction system, characterized in that: include: Modeling module, used to build prediction models by combining CNN and Transformer; A data acquisition module, used for acquiring MRI data and preprocessing the MRI data; A first training module is used to pre-train the encoder part of the prediction model using the pre-processed MRI data as input data using a masked autoencoder to obtain a first model; A second training module is used to train the label data for performing the classification task using a collaborative training method to obtain a second model; The third training module is used to use the transfer learning method to connect the first model to the second model, and train it to obtain a glioma H3K27M change prediction model.

Images