A method, storage medium and device for predicting CSF1R in brain glioma

Abstract

The present application relates to a method, storage medium and device for predicting CSF1R in brain glioma, and belongs to the field of medical information analysis technology. The method includes preprocessing magnetic resonance image data; extracting the imaging genomics feature CR of the magnetic resonance image data; performing multi-angle and multi-slice 2D image analysis and processing on the magnetic resonance image data, inputting the processed data into a convolutional neural network for training, extracting the deep learning feature DL of the last convolutional layer; fusing the imaging genomics feature CR and the deep learning feature DL to obtain feature S; inputting feature S into the pre-trained machine learning ensemble learning strategy to predict the CSF1R label, and finally obtaining the CSF1R prediction result. The present application realizes an end-to-end automated process, combines the advantages of multimodal data and imaging genomics and deep learning, significantly improves the accuracy and efficiency of prediction, and provides strong support for the diagnosis and treatment of brain glioma.

Description

Technical Field

[0001] The present application relates to the field of medical information analysis technology, and in particular to a method, storage medium, and device for predicting CSF1R in brain glioma. Background Art

[0002] Glioma is the most common and aggressive primary tumor in the central nervous system. Tumor-associated macrophages (TAMs) in the tumor microenvironment are key players in glioma immune evasion and the formation of an immunosuppressive environment. Colony-stimulating factor 1 receptor (CSF1R) signaling plays an important role in the differentiation and survival of TAMs, and CSF1R blockade is considered a potential therapeutic strategy.

[0003] Magnetic resonance imaging (MRI) plays an important role in the diagnosis and treatment of gliomas. Radiomics technology can assess the risk of tumors based on signal intensity changes, shape characteristics, and texture characteristics within the tumor volume of interest, and deep learning technology is gradually being applied to the field of medical imaging.

[0004] Existing technologies for processing MRI glioma data typically require a series of preprocessing steps and segmentation of the glioma region as training data. However, segmenting the glioma region on MRI is a very time-consuming process, which greatly limits the further development of the technology and is one of the main obstacles affecting the end-to-end application of such models.

[0005] Furthermore, noninvasive preoperative prediction of CSF1R abundance is expected to provide physicians with insights into tumor prognosis and assist in making more informed treatment decisions. This prediction is particularly valuable for immunotherapy decisions, given the association between CSF1R and tumor-associated macrophages (TAMs). Summary of the Invention

[0006] In order to address one or more of the problems in the prior art, the present application provides a method for predicting CSF1R in gliomas in the first aspect, comprising:

[0007] Preprocessing magnetic resonance image data;

[0008] extracting radiomics features CR of the preprocessed magnetic resonance image data;

[0009] Performing multi-angle and multi-slice 2D image analysis and processing on the pre-processed magnetic resonance image data, inputting the processed data into a convolutional neural network for training, and extracting the deep learning features DL of the last convolutional layer;

[0010] Fusing the radiomics feature CR and the deep learning feature DL to obtain feature S;

[0011] The feature S is input into the pre-trained machine learning ensemble learning strategy to predict the CSF1R label, and finally obtain the prediction result of CSF1R.

[0012] Preferably, the preprocessing of the magnetic resonance image data includes:

[0013] converting the magnetic resonance image data into a format;

[0014] The format-converted magnetic resonance image data are resampled and registered in standard space;

[0015] Perform N4 bias field correction on magnetic resonance image data;

[0016] The magnetic resonance image data were normalized for signal intensity;

[0017] The MRI data was skull-stripped to extract brain images;

[0018] Glioma region segmentation is performed on magnetic resonance image data.

[0019] Preferably, the magnetic resonance image data includes a T1WI enhancement sequence and a FLAIR sequence.

[0020] Preferably, the pre-processed magnetic resonance image data is subjected to multi-angle and multi-slice 2D image analysis and processing, the processed data is input into a convolutional neural network for training, and the deep learning features DL of the last convolutional layer are extracted, including:

[0021] Based on the preprocessed magnetic resonance image data, extracting a bounding box of the tumor on the largest slice, and additionally extracting slices where the tumor area is located at multiple percentiles, to form a multi-layer image set to capture local tumor features and contextual information;

[0022] The registered FLAIR and T1CE sequences and their VOIs were rotated in different directions and then bounding boxes were extracted, so that each sequence had several typical slices.

[0023] The extracted slice images are resized, normalized, and converted into tensors.

[0024] Preferably, the convolutional neural network model includes five convolutional layers, a pooling layer, a fully connected layer and an output classification layer.

[0025] Preferably, during the training process, the cross entropy loss function is used to calculate the loss, and the weights and biases in the model are updated through the error back propagation algorithm. After the training is completed, the features of the last layer are extracted to obtain the deep learning features DL.

[0026] Preferably, the radiomics feature CR and the deep learning feature DL are fused to obtain the feature S including:

[0027] The radiomics feature CR and the deep learning feature DL are fused through one or more methods including feature concatenation, weighted averaging, feature selection and dimensionality reduction, and feature standardization and normalization to form feature S;

[0028] Preferably, the feature S is input into a pre-trained machine learning ensemble learning strategy to predict the CSF1R label, and the final prediction result of CSF1R includes:

[0029] Perform ensemble learning using a variety of machine learning algorithms including but not limited to logistic regression, support vector machines, naive Bayes, K-nearest neighbors, random forests, AdaBoost, gradient boosted decision trees, LightGBM, and multi-perceptrons;

[0030] When more than half of the algorithms predict that CSF1R is high or low, the result is determined to be the final prediction result, thereby improving the accuracy and reliability of the prediction.

[0031] In a second aspect, the present application provides a storage medium storing a computer program. When the computer program is executed, the above-mentioned CSF1R prediction method in brain glioma can be implemented.

[0032] A third aspect of the present application provides a device for predicting CSF1R in brain glioma, the device comprising:

[0033] processor; and

[0034] The memory is used to store the storage medium mentioned above. When the program in the storage medium is executed by the processor, the program implements the above-mentioned CSF1R prediction method in brain glioma.

[0035] One or more of the above embodiments of the present application have at least the following beneficial effects:

[0036] This application automates the entire process, from magnetic resonance imaging data input to CSF1R prediction output. By seamlessly integrating a series of steps—format conversion, preprocessing, feature extraction, model training, and prediction—without manual intervention, processing efficiency and result consistency are significantly improved. This end-to-end process design not only reduces human error but also enables the model to rapidly respond to new data inputs, providing timely support for clinical decision-making.

[0037] This application combines two MRI modalities: T1WI-enhanced sequences and FLAIR sequences, leveraging the strengths of each modality. T1WI-enhanced sequences clearly demonstrate blood-brain barrier disruption and the parenchymal portion of a tumor, while FLAIR sequences offer significant advantages in revealing areas of brain edema surrounding the tumor. By integrating data from these two modalities, the model can more comprehensively capture tumor characteristics and improve prediction accuracy.

[0038] This application combines traditional radiomics feature extraction with deep learning technology. Radiomics features can reflect information such as tumor signal intensity, shape, and texture, while deep learning features can capture complex, nonlinear patterns in images. By extracting and fusing these two features, the model can more comprehensively characterize the biological characteristics of tumors and significantly improve the prediction ability of CSF1R.

[0039] The prediction model utilizes a voting ensemble learning strategy involving nine machine learning algorithms. This strategy leverages the synergy of multiple algorithms, combining their strengths to effectively improve model stability and predictive accuracy. When more than half of the algorithms reach a consensus on their predictions, the final result becomes more reliable, reducing the bias and uncertainty that could be introduced by a single algorithm.

[0040] The prediction results of this application have important clinical significance. By accurately predicting CSF1R status in gliomas, physicians can better assess tumor prognosis and develop more precise treatment plans for patients. This prediction, in particular, provides a critical reference for immunotherapy decisions related to tumor-associated macrophages (TAMs), potentially improving treatment outcomes and possessing promising clinical application prospects and socioeconomic value. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] The accompanying drawings are used to provide a further understanding of the present application and constitute a part of the specification. Together with the embodiments of the present application, they are used to explain the present application and do not constitute a limitation of the present application. In the accompanying drawings:

[0042] Figure 1 1 is a flow chart of a method for predicting CSF1R in gliomas provided in some embodiments of the present application;

[0043] Figure 2 These are the feature names and correlation heatmaps of traditional radiomics features CR and deep learning features DL provided in some embodiments of the present application;

[0044] Figure 3-a 1 is a graph comparing ROC curves of five-fold cross validation of different models in the training set in the method for predicting CSF1R in brain glioma provided by some embodiments of the present application;

[0045] Figure 3-b 3. This is a performance comparison result diagram of different feature combinations in the training set of the CSF1R prediction method for brain glioma provided by some embodiments of the present application;

[0046] Figure 4-a 3. This is a graph comparing ROC curves of five-fold cross validation of different models in an independent test set in the CSF1R prediction method for brain glioma provided by some embodiments of the present application;

[0047] Figure 4-b 3. This is a performance comparison result diagram of different feature combinations in an independent test set in the CSF1R prediction method for brain glioma provided by some embodiments of the present application;

[0048] Figure 5-a is a KM survival curve diagram of the prediction efficacy of CSF1R prediction results obtained by the CSF1R prediction method in brain glioma provided by some embodiments of the present application on an external data set (cohort 4);

[0049] Figure 5-b is a KM survival curve diagram of the prediction efficacy of CSF1R prediction results obtained by the CSF1R prediction method in brain glioma provided by some embodiments of the present application on an external data set (cohort 4);

[0050] Figure 5-c is a KM survival curve diagram of the prediction efficacy of CSF1R prediction results obtained by the CSF1R prediction method in brain glioma provided by some embodiments of the present application on an external data set (cohort 4);

[0051] Figure 6-a is a statistical correlation diagram of the CSF1R prediction results obtained by the CSF1R prediction method in brain glioma provided by some embodiments of the present application and the macrophage marker CD163 on an external data set;

[0052] Figure 6-b This is a graph of imaging information results of CSF1R prediction results obtained by the CSF1R prediction method in brain glioma provided by some embodiments of the present application on an external data set;

[0053] Figure 6-c is an image information result diagram of the CSF1R prediction result obtained by the CSF1R prediction method in brain glioma provided by other embodiments of the present application on an external data set;

[0054] Figure 6-d is an image information result diagram of the CSF1R prediction result obtained by the CSF1R prediction method in brain glioma provided by some embodiments of the present application on an external data set;

[0055] Figure 6-eIt is a result diagram of imaging information of CSF1R prediction results obtained by the CSF1R prediction method in brain glioma provided in some embodiments of the present application on an external data set;

[0056] Figure 7-a Schematic diagram of the correlation between the CSF1R prediction results obtained by the CSF1R prediction method in brain gliomas provided by some embodiments of the present application and the single-cell sequencing results of lesion F1 macrophages in patients with multifocal gliomas on an external dataset;

[0057] Figure 7-b Schematic diagram of the correlation between the CSF1R prediction results obtained by the CSF1R prediction method in brain gliomas provided in some embodiments of the present application and the single-cell sequencing results of lesion F2 macrophages in patients with multifocal gliomas on an external dataset;

[0058] Figure 8-a Schematic diagram of the correlation between the CSF1R prediction results obtained by the CSF1R prediction method in brain gliomas provided in other embodiments of the present application and the single-cell sequencing results of lesion F1 macrophages in patients with multifocal gliomas on an external dataset;

[0059] Figure 8-b This is a schematic diagram of the correlation between the CSF1R prediction results obtained by the CSF1R prediction method in brain glioma provided in other embodiments of the present application and the single-cell sequencing results of lesion F2 macrophages in patients with multifocal gliomas on an external dataset. DETAILED DESCRIPTION

[0060] Embodiments of the present application will be described in detail below, examples of which are illustrated in the accompanying drawings. Generally, the components of the embodiments of the present application described and shown in the drawings herein may be arranged and designed in a variety of different configurations. Therefore, the following detailed description of the embodiments of the present application provided in the drawings is not intended to limit the scope of the claimed application, but rather merely represents selected embodiments of the present application.

[0061] Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making any creative work shall fall within the scope of protection of this application.

[0062] In the description of this application, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicating orientations or positional relationships, are based on the orientations or positional relationships shown in the accompanying drawings and are intended solely to facilitate the description of this application and simplify the description. They do not indicate or imply that the devices or components referred to must have a specific orientation, be constructed, or operate in a specific orientation. Therefore, they should not be construed as limitations on this application. Furthermore, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0063] In the description of this application, it should be noted that, unless otherwise expressly specified or limited, the terms "mounted," "connected," and "connected" should be understood in a broad sense. For example, they can refer to fixed connections, detachable connections, or integral connections; mechanical connections or electrical connections; direct connections or indirect connections through an intermediate medium; and internal connections between two components. Those skilled in the art will understand the specific meanings of the above terms in this application based on the specific circumstances.

[0064] The following will be combined Figure 1 FIG8 provides a clear and complete description of the technical solution of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all of the embodiments.

[0065] See also Figure 1 The present application provides a method for predicting CSF1R in brain glioma, comprising:

[0066] Preprocessing of magnetic resonance image (MRI) data;

[0067] Extract radiomics features CR from pre-processed magnetic resonance image data;

[0068] The pre-processed magnetic resonance image data is subjected to multi-angle and multi-slice 2D image analysis and processing, and the processed data is input into a convolutional neural network (CNN) for training to extract the deep learning features DL of the last convolutional layer;

[0069] The radiomics feature CR and the deep learning feature DL are integrated to obtain the feature S;

[0070] The feature S is input into the pre-trained machine learning ensemble learning strategy to predict the CSF1R label, and finally the prediction result of CSF1R is obtained.

[0071] In some embodiments, the magnetic resonance imaging data used covers multiple modalities to comprehensively capture the imaging characteristics of brain gliomas, including but not limited to:

[0072] T1-weighted imaging (T1WI) enhanced sequence:

[0073] This sequence uses contrast agents (such as gadopentetate dimeglumine) to enhance the scan on the basis of T1WI, which can significantly improve the contrast between gliomas and surrounding normal brain tissue. The enhanced T1WI images can clearly show the damage to the blood-brain barrier of the tumor, the solid part of the tumor, and possible necrotic areas. In the diagnosis of gliomas, the T1WI enhanced sequence plays a key role in assessing the size, shape, location, and presence of enhancement of the tumor, and is an important imaging basis for preoperative planning and postoperative follow-up.

[0074] Fluid-attenuated inversion recovery (FLAIR) sequence:

[0075] The FLAIR sequence is a T2-weighted imaging technique unique in its ability to effectively suppress the high signal intensity of cerebrospinal fluid (CSF), making lesions around the ventricles and in the sulci and fissures easier to detect. In the assessment of gliomas, the FLAIR sequence has a significant advantage in demonstrating areas of brain edema surrounding the tumor. Because brain edema appears as a high signal intensity on FLAIR images, contrasting sharply with normal brain tissue, the extent and degree of edema can be more accurately delineated, which is important for determining the aggressiveness of the tumor, assessing the patient's prognosis, and formulating treatment plans (such as determining the irradiation range for radiotherapy).

[0076] These multimodal MRI data together form the foundation for a comprehensive assessment of gliomas, providing a rich source of information for subsequent imaging feature extraction and deep learning analysis. By fusing image data from different modalities, we can more comprehensively capture the morphological, functional, and microstructural characteristics of tumors, thereby improving our understanding of the biological behavior of gliomas and the accuracy of our predictions.

[0077] In some embodiments, preprocessing the magnetic resonance image data specifically includes:

[0078] 1) Convert the original DICOM MRI image data to the NIFTI format. This process can be performed using professional medical image processing software such as 3D Slicer or Medical Image Processing, Analysis, and Visualization (MIPAV), or by utilizing programming libraries such as the nibabel library in Python. These tools can efficiently read image data and metadata from DICOM files and save them in the NIFTI format, ensuring uniformity and compatibility of the data format and paving the way for subsequent processing and analysis.

[0079] 2) The converted MRI data were spatially repositioned and resampled to the MNI standard space using the advanced normalization tool ANTs. Specifically, the image data were resized to 193 × 229 × 229, with voxel sizes precisely set to 1 mm × 1 mm × 1 mm. This process, through a nonlinear registration algorithm, enabled image data from different patients to be compared and analyzed in the same spatial framework, eliminating inter-individual differences in brain structure and improving the comparability and accuracy of subsequent feature extraction and analysis.

[0080] 3) N4 bias field correction is performed on the resampled MRI images. This step utilizes the N4 bias field correction algorithm, which is based on a discrete Gaussian mixture model and effectively corrects for intensity nonuniformity in the image. By estimating and removing slowly varying intensity deviations in the image, N4 bias field correction significantly improves image quality and reduces image intensity deviations caused by differences in equipment or imaging conditions. This ensures that the image's grayscale values ​​more accurately reflect tissue signal intensity, providing a more reliable data foundation for subsequent image analysis.

[0081] 4) Signal intensity normalization was performed on the corrected images using the White Stripe R package. This method calculates the statistical distribution of signal intensity within the white matter region and adjusts the signal intensity of all images to the same range, making the image signal intensities comparable across patients. This step effectively eliminates the influence of individual differences on subsequent feature extraction, ensuring that features extracted from images of different patients can be compared and analyzed within the same intensity reference frame, thereby improving the stability and reliability of the features.

[0082] 5) Skull stripping is performed on the standardized images using the MNI152 brain mask. Using the brain mask from the MNI152 standard brain template, the skull, scalp, and other non-brain tissue are precisely removed from the image, retaining only the brain tissue image data. This process, through template matching and image segmentation techniques, accurately isolates brain tissue, reducing interference from non-brain tissue on subsequent analysis, improving the accuracy of glioma region analysis, and enabling subsequent feature extraction and analysis to focus more on the tumor tissue itself.

[0083] 6) The CKD TransBTS framework is used to automatically segment brain images, precisely locating and extracting image data of glioma regions. This framework combines deep learning techniques with traditional image segmentation methods to accurately identify and segment glioma tissue. Using a well-trained convolutional neural network model, brain images are classified pixel by pixel, distinguishing glioma tissue from other normal brain tissue and generating corresponding segmentation masks. Applying the segmentation mask to the original image yields image data containing only the glioma region, providing a clear region for subsequent feature extraction. This ensures that the extracted features are highly targeted and effective, accurately reflecting the imaging characteristics of gliomas.

[0084] Radiomics features (CR) are high-dimensional features extracted from preprocessed magnetic resonance image data that can reflect the internal heterogeneity, texture features, shape features, and signal intensity changes of tumors.

[0085] In some embodiments, extracting the radiomics feature CR of the pre-processed magnetic resonance image data specifically includes:

[0086] 1) Data Preparation

[0087] We use rigorously preprocessed T1WI enhanced and FLAIR MRI data as input for feature extraction. These data have undergone preprocessing steps such as format conversion, spatial resampling, N4 bias field correction, signal intensity normalization, skull stripping, and tumor segmentation, ensuring high data quality and consistency, providing a reliable foundation for feature extraction.

[0088] The glioma region mask obtained by automatic segmentation of brain images using the CKD TransBTS framework accurately defines the boundaries of the tumor, ensures the targeted and effective feature extraction, and avoids interference from non-tumor areas on the features.

[0089] 2) Feature extraction tools and parameter settings

[0090] The pyradiomics open-source library was used for radiomics feature extraction. This library provides comprehensive, standardized feature extraction methods that can extract a variety of radiomics features from magnetic resonance images, including first-order statistical features, shape features, and texture features, meeting the requirements of this example for feature diversity and professionalism.

[0091] During the feature extraction process, reasonable parameters were set to ensure the accuracy and stability of the extracted features. For example, the voxel size was set to 1mm×1mm×1mm to ensure the spatial resolution of the image; the image was discretized using an appropriate number of grayscale levels to balance feature detail and computational complexity; and appropriate filters were selected for image preprocessing to enhance the robustness of the features.

[0092] 3) Feature extraction process

[0093] First-order statistical feature extraction: Statistical features reflecting the signal intensity distribution, such as mean, variance, skewness, and kurtosis, are extracted from the pixel values ​​of the tumor region. These features can quantify the central tendency, dispersion, and distribution of signal intensity within the tumor, providing a basis for assessing tumor heterogeneity.

[0094] Shape feature extraction: Calculate geometric features such as tumor volume, surface area, sphericity, and flatness to describe the tumor's morphology. Shape features help understand tumor growth patterns and invasiveness, providing a reference for clinical diagnosis and treatment planning.

[0095] Texture feature extraction: Using methods such as the gray-level co-occurrence matrix (GLCM), gray-level run-length matrix (GLRLM), and gray-level size-zone matrix (GLSZM), we extract texture features that reflect the grayscale distribution patterns and spatial relationships within the tumor. These features can reveal the microscopic structure and tissue characteristics of the tumor and have potential value in identifying the biological behavior of the tumor.

[0096] 4) Feature screening and optimization

[0097] Correlation analysis is performed on the large number of extracted radiomics features to remove highly correlated features to reduce data redundancy and model complexity. By calculating indicators such as the Pearson correlation coefficient between features, a subset of features with independent information is screened to improve model efficiency and stability.

[0098] Methods such as resampling and cross-validation are used to assess the stability of features and ensure that the selected features have good consistency and repeatability across different datasets and experimental conditions. This helps improve the generalization ability and reliability of the model, enabling it to be better applied in actual clinical scenarios.

[0099] Figure 2It is the feature name and correlation heat map of the traditional radiomics feature CR and deep learning feature DL provided in some embodiments of the present application; the heat map intuitively presents the strength and direction of the correlation between different features, where the depth of the color represents the strength of the correlation (from -1 to 1, the larger the absolute value, the higher the correlation), and the warm and cold colors represent the positive and negative correlation. Specific features include a variety of radiomics features extracted from FLAIR and T1CE sequences, such as large-area low-grayscale emphasis, energy and mean after wavelet transform, and deep learning features extracted by convolutional neural networks. The correlation analysis of these features helps to understand the relationship between features, evaluate the diversity and redundancy of feature sets, and provide a basis for subsequent feature fusion and model training.

[0100] In this embodiment, see Figure 2 As shown in the figure, the final radiomics features (CR) used for the unknown test dataset include the following 9:

[0101] Rad1: FLAIR_square_glszm_LargeAreaLowGrayLevelEmphasis (FLAIR sequence: large area low gray level emphasis of gray level size matrix (GLSZM)): This feature combines the image information of the FLAIR sequence and the texture analysis method of the gray level size zone matrix (GLSZM) to reflect the distribution of large areas of low gray level within the tumor area, which may be related to the degree of tumor edema and cell density.

[0102] Rad2: FLAIR_wavelet-HHH_firstorder_Energy (FLAIR sequence: first-order energy of the HHH subband of the wavelet transform): The energy value of the first-order statistical feature extracted by performing a wavelet transform (HHH direction) on the FLAIR sequence images. This feature reflects the concentration of the tumor's signal intensity distribution under multiresolution analysis and may be related to tumor angiogenesis and cell proliferation activity.

[0103] Rad3: FLAIR_wavelet-HHH_firstorder_Mean (FLAIR sequence: first-order mean of the HHH subband of the wavelet transform): Also based on the wavelet-transformed FLAIR image, this function calculates the average grayscale value of pixels within the tumor area. This value reflects the overall signal intensity level of the tumor at a specific scale and helps assess the metabolic activity of the tumor.

[0104] Rad4: FLAIR_wavelet-LHL_glcm_MCC (FLAIR sequence: Gray-Level Co-occurrence Matrix (GLCM) correlation coefficient of the wavelet-transformed LHL subband): Combines the wavelet transform (LHL direction) and the gray-level co-occurrence matrix (GLCM) method to extract the matrix correlation coefficient (MCC) feature, which is used to describe the correlation and spatial distribution pattern of grayscale values ​​within the tumor, which may be related to the cellular arrangement and tissue structure of the tumor.

[0105] Rad5: T1CE_exponential_firstorder_Energy (T1CE sequence: first-order energy after exponential transformation): The energy value of the first-order statistical features extracted after exponential transformation of the tumor region in the T1CE sequence can highlight the distribution of high-intensity signal areas in the tumor, which may be related to the degree of tumor enhancement and vascular density.

[0106] Rad6: T1CE_exponential_ngtdm_Busyness (T1CE sequence: Busyness of the neighborhood grayscale difference matrix (NGTDM) after exponential transformation): Based on the exponential transformation of T1CE sequence images, the busyness feature extracted using the neighborhood grayscale difference matrix (NGTDM) reflects the complexity of grayscale changes within the tumor and may be related to tumor heterogeneity and invasiveness.

[0107] Rad7: T1CE_gradient_ngtdm_Busyness (T1CE sequence: busyness of the neighborhood grayscale difference matrix (NGTDM) after gradient transformation): After calculating the gradient of T1CE sequence images, the busyness feature extracted by NGTDM emphasizes the grayscale changes and details of the tumor margin area, which helps to identify the boundary characteristics and invasive growth pattern of the tumor.

[0108] Rad8: T1CE_square_firstorder_Entropy (T1CE sequence: first-order entropy after square transformation): The entropy value of the first-order statistical features extracted after square transformation of the tumor region in the T1CE sequence images reflects the degree of disorder in the tumor signal intensity distribution and may be related to the malignancy and prognosis of the tumor.

[0109] Rad9: T1CE_wavelet-HLL_glcm_Correlation (T1CE sequence: correlation of gray-level co-occurrence matrices (GLCM) of wavelet-transformed HLL subbands): Combines the wavelet transform (HLL direction) and GLCM methods to extract gray-level co-occurrence matrix correlation features. This describes the linear relationship between grayscale values ​​in different directions within the tumor, which may be related to tumor cell-cell interactions and tissue structural stability.

[0110] These features are derived from FLAIR and T1CE sequences respectively, covering different feature types and analysis methods. They can comprehensively characterize the imaging characteristics of tumors from multiple angles, providing rich and representative data support for subsequent model training and prediction.

[0111] Deep learning (DL) features are high-level features obtained at the final convolutional layer by inputting preprocessed MRI image data into a convolutional neural network (CNN). These features, through the network's automatic learning and feature extraction process, capture complex, nonlinear patterns and detailed information in the image, providing a crucial foundation for the model's high-precision predictions. Compared to traditional radiomics features, deep learning features possess greater expressive power and sensitivity to subtle image changes, providing a richer source of information for the model.

[0112] In some embodiments, the pre-processed magnetic resonance image data is subjected to multi-angle and multi-slice 2D image analysis and processing, and the processed data is input into a convolutional neural network for training. Extracting the deep learning features DL of the last convolutional layer includes:

[0113] (1) Data preparation

[0114] The input is rigorously preprocessed T1WI enhanced and FLAIR sequence MRI data. These data have undergone format conversion, spatial resampling, N4 bias field correction, signal intensity normalization, skull stripping, and tumor segmentation. This ensures high data quality and consistency, providing a reliable foundation for subsequent feature extraction.

[0115] The glioma region mask obtained by automatic segmentation of brain images using the CKD TransBTS framework accurately defines the boundaries of the tumor, ensures the targeted and effective feature extraction, and avoids interference from non-tumor areas on the features.

[0116] (2) Image analysis and processing

[0117] The preprocessed T1CE and FLAIR sequences and their tumor masks were used to extract the tumor bounding box on the largest slice. Additionally, slices with tumor regions at the 20th, 30th, 40th, 50th, and 60th percentiles were extracted, forming a six-layer image set to capture local tumor features and contextual information. Furthermore, because conventional 2D image data augmentation cannot achieve 3D rotation, the registered sequences and their VOIs were rotated 60° to the right and forward, respectively, before bounding box extraction. This resulted in 18 representative slices from each of the two sequences for a given patient.

[0118] The obtained slice image data are preprocessed by resizing, normalization and tensor conversion to ensure the consistency of the input.

[0119] (3) Convolutional Neural Network Model Construction and Training

[0120] A convolutional neural network model was constructed, consisting of five convolutional layers (including multiple residual blocks), a pooling layer, a fully connected layer, and an output classification layer. The convolution kernels of the convolutional layers were 7x7 (the first layer) and 3x3 (the convolution layers in the residual blocks) to effectively extract both low-level and high-level features of the image.

[0121] The processed MRI image data and classification labels were fed into a convolutional neural network model for training. During training, a cross-entropy loss function was used to calculate the loss, and the weights and biases in the model were updated using a backpropagation algorithm. With this extensive training data, the model was able to learn the complex mapping relationship between image data and CSF1R infiltration labels, thereby optimizing network parameters and improving the model's predictive accuracy.

[0122] At the end of the convolutional neural network model, a fully connected layer is used for classification, and the classification results of the network are output through the Softmax function.

[0123] During the training phase, a comprehensive data augmentation strategy was implemented, including adding random flips, rotations, and Gaussian noise to increase data diversity and reduce the risk of overfitting. During the validation phase, only basic preprocessing steps were applied.

[0124] (IV) Deep Learning Feature Extraction

[0125] After the convolutional neural network model is trained, the output features of the last convolutional layer are extracted as deep learning features (DL). The feature map of the last convolutional layer undergoes multiple layers of convolution and nonlinear transformations, and has high abstraction and semantic information, which can effectively represent the key features in the image.

[0126] In this embodiment, the above steps are batch processed by the program and no manual operation is required.

[0127] In this embodiment, see Figure 2 , the deep learning features (DL) finally used for the unknown test dataset include the following 8:

[0128] Rad10: rad_Flair_379 (FLAIR sequence: Deep Learning-Based Feature 379): This feature is the eigenvalue of the 379th feature channel extracted from the last convolutional layer of the FLAIR sequence image after convolutional neural network processing. It reflects a specific texture or pattern of the tumor on the FLAIR image and may be related to the characteristics of the edematous area of ​​the tumor.

[0129] Rad11: rad_T1CE_826 (T1CE sequence: 826th feature based on deep learning): Based on the T1CE sequence images, the feature value of the 826th feature channel extracted from the last convolutional layer can reflect the specific characteristics of the tumor enhancement area and may be related to the angiogenesis and cell proliferation activity of the tumor.

[0130] Rad12: rad_Flair_1194 (FLAIR sequence: the 1194th feature based on deep learning): On the FLAIR sequence images, the feature value of the 1194th feature channel extracted from the last convolutional layer may be related to the cell density and tissue structure of the tumor.

[0131] Rad13:rad_T1CE_1660 (T1CE sequence: 1160th feature based on deep learning): The feature value of the 1660th feature channel extracted from the last convolutional layer of the T1CE sequence image may reflect the metabolic activity and invasiveness of the tumor.

[0132] Rad14: rad_T1CE_1920 (T1CE sequence: 1920th feature based on deep learning): The feature value of the 1920th feature channel extracted from the last convolutional layer in the T1CE sequence image may be correlated with the malignancy and prognosis of the tumor.

[0133] Rad15: rad_Flair_338 (FLAIR sequence: the 338th feature based on deep learning): Based on the FLAIR sequence image, the feature value of the 338th feature channel extracted from the last convolutional layer may be related to the tumor's inflammatory response and immune cell infiltration.

[0134] Rad16: rad_Flair_1552 (FLAIR sequence: 1552th feature based on deep learning): The feature value of the 1552th feature channel extracted from the last convolutional layer of the FLAIR sequence image may reflect the degree of tumor fibrosis and tissue hardness.

[0135] Rad17: rad_Flair_109 (FLAIR sequence: 109th feature based on deep learning): On the FLAIR sequence images, the feature value of the 109th feature channel extracted from the last convolutional layer may be related to the vascular permeability and fluid exchange of the tumor.

[0136] These deep learning features DL are obtained through the automatic learning process of convolutional neural networks, which can capture the complex and nonlinear feature information in the image and provide important data support for subsequent model predictions.

[0137] The radiomics features (CR) and deep learning features (DL) are integrated to form a comprehensive feature (S) to more comprehensively characterize the imaging characteristics of gliomas. This fusion combines the advantages of both features, improving the model's ability to express tumor characteristics and predict accuracy, and providing richer information for subsequent classification predictions.

[0138] In some embodiments, the radiomics feature CR and the deep learning feature DL are fused to obtain a feature S including:

[0139] Since radiomics features (CR) and deep learning features (DL) may have different numerical ranges and dimensions, in order to eliminate the impact of these differences on the fusion features, the two types of features need to be normalized separately.

[0140] The normalized radiomics features (CR) and deep learning features (DL) are spliced ​​in a certain order to form a new feature S.

[0141] In some embodiments, the feature S is input into a pre-trained machine learning ensemble learning strategy to predict the CSF1R label, and the final prediction result of CSF1R includes:

[0142] 1) Input and preprocessing of feature S

[0143] Before inputting feature S into the machine learning model, it is normalized to ensure that the numerical ranges of different features are consistent, preventing certain features from dominating the model due to their large numerical ranges. Min-Max normalization can be used to map feature values ​​to the range [0, 1].

[0144] Alternatively, if the dimension of feature S is high, you can consider using dimensionality reduction techniques such as principal component analysis (PCA) or linear discriminant analysis (LDA) to reduce the feature dimension, reduce computational complexity, and improve the generalization ability of the model.

[0145] 2) Machine Learning Ensemble Learning Strategy

[0146] Nine different machine learning algorithms were selected as base learners, including linear models (logistic regression LR, support vector machine SVM), probabilistic models (naive Bayes NB), distance-based models (K nearest neighbor KNN), tree-based models (random forest RF, AdaBoost, gradient boosted decision tree GBDT, LightGBM) and multi-perceptron (MLP).

[0147] The normalized (and dimensionally reduced) features S are fed into each training set for training. During the training process, each base learner learns different model parameters based on the training data to minimize the prediction error.

[0148] A voting mechanism is used for ensemble learning. Specifically, when more than four algorithms predict CSF1R as high or low, that prediction is finalized. This voting mechanism integrates the predictions of multiple base learners, improving model stability and prediction accuracy.

[0149] 3) Model evaluation and optimization

[0150] Cross-validation methods (such as five-fold cross-validation) are used to evaluate and optimize the ensemble learning model. The dataset is divided into training and validation sets, and the training and validation process is repeated multiple times to ensure the stability and generalization ability of the model.

[0151] The predictive performance of the model is evaluated using a variety of performance indicators, including accuracy, precision, recall, F1-score, and area under the receiver operating characteristic curve (AUC-ROC).

[0152] According to the performance of the model in cross-validation, the parameters of the base learner are adjusted and optimized to improve the prediction performance of the model.

[0153] Figure 3-aThis is a Receiver Operating Characteristic Curve (ROC) curve plot of a five-fold cross-validation of traditional radiomics features (CR) and deep learning features (DL) in the training set provided by some embodiments of the present application. It is used to evaluate the classification performance of various machine learning models. The horizontal axis represents specificity, which indicates the proportion of negative cases correctly identified by the model (actually negative and predicted as negative). Its value range is [0, 1]. The closer the specificity is to 1, the more accurate the model's judgment of negative cases is. The vertical axis represents sensitivity (also known as recall), which indicates the proportion of positive cases correctly identified by the model (actually positive and predicted as positive). Its value range is also [0, 1]. The closer the sensitivity is to 1, the stronger the model's ability to identify positive cases is. The closer the ROC curve is to the upper left corner, the better the model performance is (this interpretation applies to each ROC curve in Figures 3 and 4). This subfigure shows that the top seven machine learning classifiers achieved good performance in the training set of cohort 1 (including clinical risk factors and deep learning features, i.e., CR+DL), and also shows the median ± standard deviation of the five-fold cross-validation.

[0154] Figure 3-b This is the ROC curve result of the five-fold cross-validation of the traditional imaging genomics feature CR and the deep learning feature DL in the training set provided by some embodiments of the present application, showing the best performance of the three model classifiers in the training set, and the red text highlights the best classifier result based on accuracy. RSD is the abbreviation of the relative standard deviation of the AUC value (relative standard deviation of the AUC), which indicates the stability of the classifier performance. For the CR+DL model, its best classifier is MLP, and the overall performance after five-fold cross-validation is ACC=0.92±0.07, AUC=0.96±0.05, and RSD=5.56%. Compared with the separate CR and DL models, the CR+DL model shows higher classification efficiency and result stability.

[0155] Figure 4-a This figure compares the ROC curves of different models from a five-fold cross-validation analysis in an independent test set for the CSF1R prediction method for gliomas provided in some embodiments of this application. Similar to Figure 3-a, this figure shows the performance of the top seven machine learning classifiers by accuracy. It can be seen that the LR, SVM, and MLP classifiers still achieved good performance in the test set, with the highest ACU value of 0.72.

[0156] Figure 4-bThis is a performance comparison result diagram of different feature combinations in an independent test set in the CSF1R prediction method for brain glioma provided by some embodiments of the present application; Same as 3-b, the results show that the CR+DL model still achieved the highest classification efficiency and result stability in the test set, and the corresponding SVM classifier had ACC = 0.76 and AUC = 0.70.

[0157] 4) Generation of prediction results

[0158] The trained and optimized ensemble learning model takes the feature S as input and outputs the predicted CSF1R label. Based on the decision of the voting mechanism, the CSF1R infiltration status of each sample is determined to be high or low.

[0159] To better understand and apply predictions, model predictions can be visualized. For example, plotting receiver operating characteristic (ROC) curves and confusion matrices can provide intuitive insights into the model's classification performance and prediction effectiveness. Furthermore, methods such as feature importance analysis can be used to explain which features have a significant impact on predictions, providing clinicians with deeper insights.

[0160] Figure 5-a This is a KM survival curve plot showing the predictive performance of the CSF1R prediction results obtained by the CSF1R prediction method for gliomas provided in the examples of this application on an external dataset (cohort 4). The horizontal axis represents survival time, and the vertical axis represents survival probability. Each coordinate point represents a specific glioma patient. The closer the curve is to the lower left corner, the worse the prognosis for that group of patients, while the closer it is to the upper right corner, the better the prognosis. The categories predicted by this method shown in the figure are statistically significant in the glioma population (p < 0.0001). That is, when a glioma patient is predicted to be in the high CSF1R group, the prognosis is worse, and conversely, when predicted to be in the low CSF1R group, the prognosis is better.

[0161] Figure 5-b similar Figure 5-a , the categories predicted by this method are statistically significant in low-grade glioma (LGG) (p=0.003), that is, when LGG patients are predicted to be in the CSF1R high group, the prognosis is worse, and vice versa, when they are predicted to be in the CSF1R low group, the prognosis is better.

[0162] Figure 5-c similar Figure 5-a , the categories predicted by this method are statistically significant in glioblastoma (GBM) (p=0.031), that is, when GBM patients are predicted to be in the CSF1R high group, the prognosis is worse, and vice versa, when they are predicted to be in the CSF1R low group, the prognosis is better. Figure 5-a , Figure 5-b , Figure 5-cThis shows that the glioma CSF1R prediction method provided in the examples of the present application has the practical efficacy of predicting the prognosis of patients.

[0163] Figure 6-a This figure shows the statistical correlation between the CSF1R prediction results obtained by the CSF1R prediction method in gliomas provided by some embodiments of this application and the immunohistochemical staining (IHC) results for the macrophage marker CD163 in an external dataset. The IHC score is a comprehensive metric that quantifies staining intensity. Each point in the histogram on the left represents the value and classification of a random field. The results show that the CSF1R infiltration abundance predicted by the method of this embodiment can be verified by immunohistochemical staining, with the result being statistically significant (p=0.0003).

[0164] Figure 6-b This figure shows the CSF1R prediction results obtained by the CSF1R prediction method for gliomas provided in some embodiments of this application, based on an external dataset (using the representative case NJ01 as an example), along with imaging information (T1CE sequence on the left and FLAIR sequence on the right), immunohistochemical staining (IHC) results for the macrophage marker CD163, and case data. The results demonstrate that the method described herein predicted the patient's CSF1R-low group, consistent with the actual immunohistochemical staining results (low risk) and the patient's long survival time (39.1 months).

[0165] Figure 6-c similar Figure 6-b , showing the imaging information of the representative case NJ02, immunohistochemical staining of the macrophage marker CD163, and data information of the case. The results show that the prediction results of the method described in this application show that the patient was predicted to be in the CSF1R high group, which is consistent with the actual immunohistochemical staining results (high risk) and the shorter survival time (16.2 months).

[0166] Figure 6-d similar Figure 6-b , presenting the imaging information of the representative case NJ03, immunohistochemical staining of the macrophage marker CD163, and data information of the case. The results show that according to the prediction results of the method described in this application, this patient was predicted to be in the CSF1R low group. The actual immunohistochemical staining results (low risk) and survival time were 16.3 months, which is relatively consistent with the above results considering the patient's advanced age (71 years old).

[0167] Figure 6-e similar Figure 6-b, showing the imaging information of the representative case NJ04, the immunohistochemical staining of the macrophage marker CD163, and the case data. The results show that the prediction results of the method described in this application show that this patient was predicted to be in the CSF1R high group, which is consistent with the actual immunohistochemical staining results (high risk) and the shorter survival time (10.4 months).

[0168] Figure 7-a and Figure 7-b This is a schematic diagram of the correlation between the CSF1R prediction results obtained by the CSF1R prediction method in brain glioma provided by an embodiment of the present application and the macrophage single-cell sequencing results of different lesions (labeled as F1 and F2, respectively, where F1 is labeled in the T1CE sequence and F2 is labeled in the FLAIR sequence) in a multi-lesion glioma patient (NJ05) on an external dataset (the right side of the two sequences are tSNE dimensionality reduction clustering diagrams of single-cell sequencing, each point represents a cell, and red represents macrophages). The results show that for the F1 lesion, more than half of the 9 classifiers The F2 lesion was classified as belonging to the CSF1R-high group (the scores in parentheses represent the voting results of the remaining classifiers after excluding the three least accurate classifiers, which are closely related to prediction confidence). Single-cell sequencing results also showed that 2,001 of the 10,160 cells in the sequenced tissue block were macrophages, indicating a 19.7% macrophage infiltration. For the F2 lesion, more than half of the nine classifiers classified it as belonging to the CSF1R-low group. Single-cell sequencing results also showed that 1,191 of the 11,358 cells in the sequenced tissue block were macrophages, indicating a 10.5% macrophage infiltration. This demonstrates that the prediction results of the method described in this application are consistent with the macrophage infiltration observed after single-cell sequencing.

[0169] Figure 8-a and Figure 8-b This is a schematic diagram showing the correlation between the CSF1R prediction results obtained by the CSF1R prediction method in brain gliomas provided in another embodiment of the present application and the single-cell sequencing results of macrophages in lesions F1 and F2 of a multi-lesion glioma patient (NJ06) in an external dataset. The results show that for the F1 lesion, while predicted by the classifier as the CSF1R-high group, the single-cell sequencing results showed that 2109 of the 9767 cells in the sequenced tissue block were macrophages, indicating a 21.6% macrophage infiltration. For the F2 lesion, while predicted by the classifier as the CSF1R-low group, the single-cell sequencing results showed that 913 of the 11948 cells in the sequenced tissue block were macrophages, indicating a 7.6% macrophage infiltration. This demonstrates that the prediction results of the method described in this application are consistent with the macrophage infiltration observed after single-cell sequencing.

[0170] The implementation of this application is mainly based on two programming languages, Python and R. On the back end, Python calls R programs to complete specific tasks, such as the extraction of imaging genomics features. Then, these programs are completely encapsulated to form an efficient overall system. In this way, when using it, the user only needs to input data on the front end, and the system can automatically complete the entire process from data processing, feature extraction, model prediction to result output, truly realizing end-to-end result output, greatly improving the convenience and efficiency of use.

[0171] A second aspect of the present application provides a storage medium storing a computer program, which is executed by a processor to implement the above-mentioned CSF1R prediction method in brain glioma.

[0172] The third aspect of the present application provides a device for predicting CSF1R infiltration in brain gliomas, comprising:

[0173] processor; and

[0174] The memory stores a computer program storage medium therein, which, when executed by the processor, enables the processor to execute the above-mentioned CSF1R prediction method in brain glioma.

[0175] It should be noted that the technical solutions in the various embodiments of the present application can be combined with each other, but the basis for the mutual combination is that it can be implemented by ordinary technicians in this field; when the combination of technical solutions is mutually contradictory or cannot be implemented, it should be deemed that such a combination of technical solutions does not exist, that is, it does not fall within the scope of protection of this application.

[0176] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for predicting CSF1R in brain glioma, characterized in that: include: Preprocessing magnetic resonance image data; extracting radiomics features CR of the preprocessed magnetic resonance image data; Performing multi-angle and multi-slice 2D image analysis and processing on the pre-processed magnetic resonance image data, inputting the processed data into a convolutional neural network for training, and extracting the deep learning features DL of the last convolutional layer; Fusing the radiomics feature CR and the deep learning feature DL to obtain feature S; Inputting the feature S into the pre-trained machine learning ensemble learning strategy to predict the CSF1R label, and finally obtaining the prediction result of CSF1R; Wherein, the magnetic resonance image data includes T1WI enhanced sequence and FLAIR sequence; Extracting radiomics features CR of the pre-processed magnetic resonance image data includes: Extract first-order statistical features, shape features, and texture features from preprocessed T1WI enhanced sequences and FLAIR sequences; The radiomics feature CR and the deep learning feature DL are fused to obtain the feature S including: The radiomics feature CR and the deep learning feature DL are fused through one or more methods including feature concatenation, weighted averaging, feature selection and dimensionality reduction, and feature standardization and normalization to form feature S; The feature S is input into the pre-trained machine learning ensemble learning strategy to predict the CSF1R label, and the final prediction result of CSF1R includes: Perform ensemble learning using a variety of machine learning algorithms including but not limited to logistic regression, support vector machines, naive Bayes, K-nearest neighbors, random forests, AdaBoost, gradient boosted decision trees, LightGBM, and multi-perceptrons; When more than half of the algorithms predict that CSF1R is high or low, the result is determined to be the final prediction result, thereby improving the accuracy and reliability of the prediction.

2. The method for predicting CSF1R in brain glioma according to claim 1, characterized in that: The preprocessing of the magnetic resonance image data includes: converting the magnetic resonance image data into a format; The format-converted magnetic resonance image data are resampled and registered in standard space; Perform N4 bias field correction on magnetic resonance image data; The magnetic resonance image data were normalized for signal intensity; The MRI data was skull-stripped to extract brain images; Glioma region segmentation is performed on magnetic resonance image data.

3. The method for predicting CSF1R in brain glioma according to claim 2, characterized in that: The pre-processed magnetic resonance image data is subjected to multi-angle and multi-slice 2D image analysis and processing, and the processed data is input into a convolutional neural network for training. The deep learning features DL of the last convolutional layer are extracted, including: Based on the preprocessed magnetic resonance image data, extracting a bounding box of the tumor on the largest slice, and additionally extracting slices where the tumor area is located at multiple percentiles, to form a multi-layer image set to capture local tumor features and contextual information; The registered FLAIR and T1CE sequences and their VOIs were rotated in different directions and then bounding boxes were extracted, so that each sequence had several typical slices. The extracted slice images are resized, normalized, and converted into tensors.

4. The method for predicting CSF1R in brain glioma according to claim 3, characterized in that: The convolutional neural network model includes five convolutional layers, a pooling layer, a fully connected layer and an output classification layer.

5. The method for predicting CSF1R in brain glioma according to claim 3, characterized in that: During the training process, the cross entropy loss function is used to calculate the loss, and the weights and biases in the model are updated through the error back propagation algorithm. After the training is completed, the features of the last layer are extracted to obtain the deep learning features DL.

6. A storage medium, characterized in that The storage medium stores a computer program, and the computer program executes the method for predicting CSF1R in glioma according to any one of claims 1 to 5.

7. A CSF1R prediction device in brain glioma, characterized in that: The device comprises: processor; and A memory for storing the storage medium according to claim 6, wherein when the program in the storage medium is executed by the processor, the program executes the method for predicting CSF1R in glioma according to any one of claims 1 to 5.

Images