Nasopharyngeal carcinoma prognosis prediction method based on MRI habitat space interaction characteristics

Abstract

The invention discloses a nasopharyngeal carcinoma prognosis prediction method based on MRI habitat and habitat space interaction characteristics. The method comprises the following steps: selecting a batch of nasopharyngeal carcinoma image data and prognosis data information corresponding to the nasopharyngeal carcinoma image data; the method comprises the steps of preprocessing data, extracting gray features of a nasopharyngeal carcinoma focus area as input of a K-Means algorithm, deconstructing a tumor area into habitat subareas with different biological characteristics, constructing a multi-area space interaction matrix based on the habitat subareas, extracting multi-area space interaction features, and taking obtained features and prognosis information as training data; a nasopharyngeal carcinoma prognosis prediction model based on multi-region space interaction features is established, training data is used as feature input of a machine learning classifier algorithm to serve as a learning sample for training, a corresponding nasopharyngeal carcinoma prognosis prediction model is obtained, the death risk of nasopharyngeal carcinoma patients can be predicted, and the patients are divided into high and low risk groups. The method does not need hypothesis, and is small in model scale, high in speed and high in accuracy.

Description

Technical Field

[0001] This invention relates to medical imaging, specifically a method for predicting the prognosis of nasopharyngeal carcinoma based on MRI habitat spatial interaction features. Background Technology

[0002] Nasopharyngeal carcinoma (NPC) is a common head and neck malignancy. According to data from the International Agency for Research on Cancer (IARC), NPC exhibits significant regional variations, occurring more frequently in Southeast Asia and East Asia. Due to its insidious early onset, lack of specific clinical symptoms, and low positive predictive value of conventional screening methods, over 70% of NPC patients are diagnosed at a locally advanced stage with poor treatment outcomes, significantly impacting treatment effectiveness. Furthermore, modern tumor biology research has confirmed significant spatial heterogeneity within tumors, with cells in different tumor regions potentially exhibiting different gene mutations. The distribution of these gene mutations within the tumor shows clear spatial differences, affecting tumor proliferation rate, invasiveness, and sensitivity to drugs and radiotherapy. Traditional radiomics assumes the entire tumor region is homogeneous, which is inconsistent with reality. Therefore, intratumoral spatial heterogeneity not only influences treatment selection and efficacy assessment but also directly relates to patient prognosis and survival. More precise, non-invasive identification of intratumoral spatial heterogeneity in NPC is crucial for predicting the prognosis of NPC patients. In recent years, habitat imaging, as an emerging image analysis strategy, can divide the interior of a tumor into image habitats with the same biological properties at the voxel level by combining multi-parameter clustering and partitioning. This habitat analysis provides a new perspective and method for analyzing the spatial heterogeneity within a tumor. Summary of the Invention

[0003] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method for predicting the prognosis of nasopharyngeal carcinoma based on MRI habitat spatial interaction features. This method has a small model size, fast speed, and high accuracy.

[0004] The technical solution to achieve the objective of this invention is:

[0005] A method for predicting the prognosis of nasopharyngeal carcinoma based on MRI habitat spatial interaction features includes the following steps:

[0006] 1) Data preparation: Select a batch of MRI images, corresponding labeled images and survival data of nasopharyngeal carcinoma patients as training samples, and set the data range for step 5);

[0007] 2) Data preprocessing: The data obtained in step 1) is preprocessed by normalizing the image data and mapping it to the range of 0 to 1;

[0008] 3) Habitat region segmentation: Cluster the normalized image data obtained in step 2) to divide the tumor region into multiple habitat regions, and generate pseudo-color labels for each image;

[0009] 4) Extracting multi-region spatial interaction features: Generate a multi-region spatial interaction matrix from the pseudo-color labels obtained in step 3), and extract multi-region spatial interaction features from the matrix;

[0010] 5) Establish a prognostic prediction model for nasopharyngeal carcinoma based on multi-regional spatial interaction features: After normalization and feature selection, the multi-regional spatial interaction features obtained in step 4) are concatenated with the prognostic data and used as feature inputs for a machine learning classifier as training samples to obtain the prognostic prediction model for nasopharyngeal carcinoma.

[0011] 6) Classify patients into high-risk and low-risk groups: Use the nasopharyngeal carcinoma prognostic prediction model established in step 5) to predict the prognosis of nasopharyngeal carcinoma patients and classify them into high-risk and low-risk groups.

[0012] Step 3) specifically refers to extracting the gray values ​​of all central voxel points in the lesion region from the normalized image data obtained in step 2), and their differences from the gray values ​​of four adjacent voxel points in the same layer, forming a global feature matrix. This matrix serves as the feature input for K-Means clustering, and the output is a global label matrix, which segments the target region into spatially unique regions with uniform signal intensity. In this way, individual tumors can be divided into different sub-regions based on the voxel features of the population. The optimal number of clusters is determined by calculating the Calinski-Harabasz (CH) index within a range of 2-10 clusters.

[0013] Step 4), specifically the multi-region spatial interaction matrix generation method, involves assigning each tumor voxel to a habitat region after clustering, resulting in a cluster label. We explore the neighborhood of each tumor voxel, identifying clustering pairs between that voxel and its neighboring voxels, and updating these pairs in the corresponding positions of the multi-region spatial interaction matrix. This process is repeated until all tumor voxels are traversed, ultimately quantifying spatial heterogeneity in the multi-region spatial interaction matrix. Intuitively, the diagonal elements in the multi-region spatial interaction matrix represent the size of a single subregion, while the off-diagonal elements reflect the size of the intersection boundaries between different subregions.

[0014] Beneficial effects: The present invention provides a prognostic prediction method for nasopharyngeal carcinoma based on MRI habitat spatial interaction features. The method has a small model size, fast speed and high accuracy, providing technical support for the prognostic prediction of nasopharyngeal carcinoma. Attached Figure Description

[0015] Figure 1This is a technical roadmap for a prognostic prediction method for nasopharyngeal carcinoma based on MRI habitat spatial interaction features.

[0016] Figure 2 A schematic diagram illustrating the characteristics of multi-regional spatial interaction; Detailed Implementation

[0017] The present invention will be further described below with reference to the accompanying drawings and embodiments, but this is not intended to limit the invention.

[0018] Example:

[0019] like Figure 1 As shown, a method for predicting the prognosis of nasopharyngeal carcinoma based on MRI habitat spatial interaction features specifically includes the following steps:

[0020] 1) MRI images, corresponding labeled images, and survival data of nasopharyngeal carcinoma patients were selected as training samples;

[0021] 2) Preprocessing the data obtained in step 1) involves normalizing the image data and mapping it to the range of 0 to 1;

[0022] 3) Habitat region segmentation: Clustering is performed on the normalized image data obtained in step 2). Specifically, the gray values ​​of all central voxel points in all image lesion regions and their differences from the gray values ​​of four adjacent voxel points in the same layer are extracted to form a global feature matrix. This matrix serves as the feature input for K-Means clustering, and the output is a global label matrix, segmenting the target region into spatially unique regions with uniform signal intensity. In this way, individual tumors can be divided into different sub-regions based on the voxel features of the population. The optimal number of clusters is determined by calculating the Calinski-Harabasz (CH) index within a range of 2-10 clusters. Based on this, the tumor region is divided into multiple different habitat regions, and a pseudo-color label is generated for each image.

[0023] 4) Extracting Multi-Region Spatial Interaction Features: A multi-region spatial interaction matrix is ​​generated from the pseudo-color labels obtained in step 3). Specifically, each tumor voxel is assigned to a habitat region after clustering, resulting in a cluster label. We explore the neighborhood of each tumor voxel, finding clustering pairs between that voxel and its neighboring voxels. These clustering pairs are then updated to the corresponding positions in the multi-region spatial interaction matrix. This process is repeated until all tumor voxels are traversed, and spatial heterogeneity is finally quantified in the multi-region spatial interaction matrix. Intuitively, the diagonal elements in the multi-region spatial interaction matrix represent the size of a single subregion, while the off-diagonal elements reflect the size of the intersection boundaries between different subregions. Multi-region spatial interaction features are then extracted from the matrix.

[0024] 5) Establish a prognostic prediction model for nasopharyngeal carcinoma based on multi-regional spatial interaction features: After normalization and feature selection, the multi-regional spatial interaction features obtained in step 4) are used as feature inputs for a machine learning classifier as training samples to obtain a prognostic prediction model for nasopharyngeal carcinoma.

[0025] 6) Classifying patients into high- and low-risk groups: The nasopharyngeal carcinoma prognostic prediction model established in step 5) can be used to classify patients into high- and low-risk groups.

[0026] The innovation of this invention lies in proposing a method based on habitat analysis: first, K-Means clustering is used to divide the tumor into subregions, and then a multi-region spatial interaction matrix is ​​constructed to extract the spatial interaction features of multiple regions within the nasopharyngeal carcinoma tumor. This method can more accurately quantify the spatial heterogeneity within the nasopharyngeal carcinoma tumor. Feature selection and machine learning classifier modeling are performed on the multi-region spatial interaction features to construct a mortality prediction model for nasopharyngeal carcinoma patients. This model shows good performance in prognostic prediction of nasopharyngeal carcinoma and improves the accuracy of prognostic prediction.

[0027] The process described above mainly involves the extraction of multi-region spatial interaction features. In practice, only MRI image data from nasopharyngeal carcinoma patients are needed. The algorithm automatically extracts features and models them, and then the model can predict the final prediction result.

[0028] The above describes the detailed technical principles of this invention, which details the clustering process, multi-region spatial interaction feature extraction algorithm, and prognostic prediction model construction method.

Claims

1. A method for predicting the prognosis of nasopharyngeal carcinoma based on MRI habitat spatial interaction features, specifically including the following steps: 1) Data preparation: Select a batch of MRI images, corresponding labeled images and survival data of nasopharyngeal carcinoma patients as training samples, and use them as the data range for step 5); 2) Data preprocessing: The data obtained in step 1) is preprocessed by normalizing the image data and mapping it to the range of 0 to 1; 3) Habitat region segmentation: Cluster the normalized image data obtained in step 2) to divide the tumor region into multiple habitat regions, and generate pseudo-color labels for each image; 4) Extracting multi-region spatial interaction features: Generate a multi-region spatial interaction matrix from the pseudo-color labels obtained in step 3), and extract multi-region spatial interaction features from the matrix; 5) Establish a prognostic prediction model for nasopharyngeal carcinoma based on multi-regional spatial interaction features: After normalization and feature selection, the multi-regional spatial interaction features obtained in step 4) are concatenated with the prognostic data and used as feature inputs for a machine learning classifier as training samples to obtain the prognostic prediction model for nasopharyngeal carcinoma. 6) Classify patients into high-risk and low-risk groups: Use the nasopharyngeal carcinoma prognostic prediction model established in step 5) to predict the prognosis of nasopharyngeal carcinoma patients and classify them into high-risk and low-risk groups.

2. The method for predicting the prognosis of nasopharyngeal carcinoma based on MRI habitat spatial interaction features according to claim 1, characterized in that, The clustering described in step 3) specifically refers to extracting the gray values ​​of all central voxel points in the lesion region from the normalized image data obtained in step 2), and their differences from the gray values ​​of four adjacent voxel points in the same layer, to form a global feature matrix. This matrix serves as the feature input for K-Means clustering, and the output is a global label matrix, which segments the target region into spatially unique regions with uniform signal intensity. In this way, individual tumors can be divided into different sub-regions based on the voxel features of the population. The optimal number of clusters is determined by calculating the Calinski-Harabasz (CH) index within a range of 2-10 clusters.

3. The method for predicting the prognosis of nasopharyngeal carcinoma based on MRI habitat spatial interaction features according to claim 1, characterized in that, Step 4), specifically the multi-region spatial interaction matrix generation method, involves assigning each tumor voxel to a habitat region after clustering, resulting in a cluster label. We explore the neighborhood of each tumor voxel, identifying clustering pairs between that voxel and its neighboring voxels, and updating these pairs in the corresponding positions of the multi-region spatial interaction matrix. This process is repeated until all tumor voxels are traversed, ultimately quantifying spatial heterogeneity in the multi-region spatial interaction matrix. Intuitively, the diagonal elements in the multi-region spatial interaction matrix represent the size of a single subregion, while the off-diagonal elements reflect the size of the intersection boundaries between different subregions.

Images