Method for predicting multiple molecular pathologies of glioma based on deep learning of multi-modal whole-brain MRI (Magnetic Resonance Imaging)

Abstract

The invention relates to a method for predicting multiple molecular pathologies of glioma based on deep learning of multi-modal whole-brain MRI. The method comprises the following steps: inputting multi-modal whole-brain MRI imaging data and DT I imaging data; extracting a multi-dimensional and multi-slice two-dimensional CNN block; analyzing the multi-dimensional and multi-slice image features obtained after preprocessing, respectively extracting an axial plane feature Esag, a coronal plane feature Ecor and a sagittal plane feature Eax, and calculating a multi-plane and multi-slice feature S; and carrying out multi-plane multi-slice image marking on the S, adding a converter coding block with a classification head, and carrying out I DH mutation, 1p / 19q co-deletion and MGMT promoter methylation multi-molecular pathology prediction. MRI imaging data and DT I imaging data of a multi-modal whole brain are comprehensively trained, features are extracted from the MR I imaging data and the DT I imaging data for training, I DH mutation, 1p / 19q co-deletion and MGMT promoter methylation multi-molecular pathology prediction is achieved, the DT I imaging data are considered in the prediction process, and MRI of a glioma area does not need to be manually marked.

Description

Technical Field

[0001] The present invention relates to the field of medical technology, and in particular to a method for predicting multiple molecular pathologies of gliomas through deep learning based on multimodal whole-brain MRI. Background Art

[0002] The fifth edition of the World Health Organization Classification of Central Nervous System Tumors, released in 2021, integrates the molecular pathology of gliomas, including IDH mutations, 1p / 19q co-deletion, and MGMT promoter methylation. Non-invasive prediction of the molecular pathology of gliomas will help with the next medical decision. Using deep learning to learn and predict MRI (magnetic resonance imaging) is considered a promising method. However, current methods often only use conventional MRI for prediction, ignoring the potential predictive benefits of advanced imaging such as DTI (diffusion tensor imaging); and current technologies usually require MRI glioma segmentation areas as training data, which is very time-consuming and limits further development; in addition, previous methods are limited to predicting only one type of molecular pathology and cannot provide multiple molecular pathology information. Summary of the invention

[0003] The technical problem to be solved by the present invention is to provide a method for predicting multiple molecular pathologies of glioma based on deep learning of multimodal whole-brain MRI, so as to overcome the deficiencies in the above-mentioned prior art.

[0004] The technical solution of the present invention to solve the above technical problems is as follows: A method for predicting multiple molecular pathologies of glioma based on deep learning of multimodal whole-brain MRI, comprising the following steps:

[0005] Step S01: input multi-modal whole-brain MRI imaging data and DTI imaging data;

[0006] Step S02: extracting multi-dimensional, multi-slice two-dimensional CNN blocks from the MRI imaging data and the DTI imaging data;

[0007] Step S03: preprocessing the two-dimensional CNN block;

[0008] Step S04: Analyze the multi-dimensional and multi-slice image features obtained after preprocessing, extract the axial plane feature Esag, the coronal plane feature Ecor, and the sagittal plane feature Eax, and calculate the multi-plane multi-slice feature S based on Esag, Ecor, and Eax;

[0009] Step S05: perform multi-plane and multi-slice image labeling on S, add a transformer coding block with a classification head, and perform multi-molecular pathology prediction of IDH mutation, 1p / 19q co-deletion, and MGMT promoter methylation.

[0010] The beneficial effects of the present invention are as follows: the present invention realizes multi-molecular pathology prediction of IDH mutation, 1p / 19q co-deletion and MGMT promoter methylation by comprehensively training multi-modal whole-brain MRI imaging data and DTI imaging data, extracting features from MRI imaging data and DTI imaging data for training, and taking DTI imaging data into consideration in the prediction process, and MRI of glioma areas does not need to be manually marked.

[0011] Based on the above technical solution, the present invention can also be improved as follows.

[0012] Furthermore, the extracted multi-dimensional and multi-slice image features in step S03 are calculated by the extraction operator E, and the calculation formula of the operator E is:

[0013] R^C3d×L×W×H→R^C3d×N×W×H.

[0014] Further, the calculation of the multi-plane multi-slice feature S in step S04 specifically includes using E=[Ecor, Esag, Eax] of the 3D representation feature X; the calculated S=[Scor, Ssag, Sax].

[0015] Further, the transformer coding block with classification header in step S05 includes:

[0016] Non-linear layer Non-L i near, which is used to reduce the dimension of the original image input;

[0017] Multi-dimensional, multi-faceted embedding layer, where the multi-dimensional embedding layer is used to preserve position information;

[0018] A multi-plane embedding layer for providing one-dimensional labels to indicate to which plane these labels belong;

[0019] A multi-dimensional, multi-faceted encoder that embeds the position information in the image into the feature representation and handles position and scale changes in the image;

[0020] A multi-classification hierarchical structure was used to simultaneously predict multiple molecular pathological features of IDH mutation, 1p / 19q co-deletion, and MGMT promoter methylation.

[0021] Furthermore, the multi-classification layer structure is specifically used for:

[0022] IDH mutation status: wild type and mutant;

[0023] 1p / 19q deletion status: deletion and no deletion;

[0024] MGMT methylation status: unmethylated and methylated.

[0025] Furthermore, the quantitative parameters of the MRI imaging data include T1w, T1wC, T2w and FLAIR.

[0026] Furthermore, the quantitative parameters of DTI imaging data include ADC and FA.

[0027] Furthermore, the preprocessing in step S03 includes normalization, denoising and enhancement. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 The present invention is a flow chart of the method. DETAILED DESCRIPTION

[0029] The principles and features of the present invention are described below in conjunction with the accompanying drawings. The examples given are only used to explain the present invention and are not used to limit the scope of the present invention.

[0030] like Figure 1 As shown in Example 1, a method for predicting multiple molecular pathologies of glioma based on deep learning of multimodal whole-brain MRI comprises the following steps:

[0031] Step S01: input multi-modal whole-brain MRI imaging data and DTI imaging data;

[0032] Step S02: extracting multi-dimensional, multi-slice two-dimensional CNN blocks from the MRI imaging data and the DTI imaging data; these blocks contain key image information and are ready for subsequent processing and analysis;

[0033] Step S03: preprocessing the two-dimensional CNN block;

[0034] Step S04: analyzing the multi-dimensional, multi-slice image features obtained after the preprocessing, respectively extracting the axial plane feature Esag, the coronal plane feature Ecor, and the sagittal plane feature Eax, and calculating the multi-plane multi-slice feature S according to the Esag, the Ecor, and the Eax;

[0035] Step S05: perform multi-plane and multi-slice image labeling on the S, add a transformer coding block with a classification head, and perform multi-molecular pathology prediction of IDH mutation, 1p / 19q co-deletion and MGMT promoter methylation.

[0036] The present invention realizes the prediction of multi-molecular pathology of IDH mutation, 1p / 19q co-deletion and MGMT promoter methylation by comprehensively training multimodal whole-brain MRI imaging data and DTI imaging data, extracting features from MRI imaging data and DTI imaging data for training, and taking DTI imaging data into consideration in the prediction process, and eliminating the need for manual marking of MRI of glioma areas.

[0037] Embodiment 2: This embodiment is a further improvement on the basis of embodiment 1, and its details are as follows:

[0038] The extracted multi-dimensional and multi-slice image features in step S03 are calculated by the extraction operator E. The calculation formula of the operator E is:

[0039] R^C3d×L×W×H→R^C3d×N×W×H.

[0040] This extraction operator can efficiently extract key information from 3D images and convert it into image features suitable for processing by deep learning models, providing strong data support for subsequent predictions.

[0041] Embodiment 3, this embodiment is a further improvement on the basis of embodiment 2, and its details are as follows:

[0042] The calculation of the multi-plane multi-slice feature S in step S04 specifically includes using E=[Ecor, Esag, Eax] of the 3D representation feature X; the calculated S=[Scor, Ssag, Sax].

[0043] Embodiment 4: This embodiment is a further improvement on the basis of any one of Embodiments 1 to 3, and its details are as follows:

[0044] The transformer coding block with classification header in step S05 includes:

[0045] Non-linear layer Non-L i near, which is used to reduce the dimension of the original image input;

[0046] Multi-dimensional, multi-faceted embedding layer, where the multi-dimensional embedding layer is used to preserve position information;

[0047] A multi-plane embedding layer for providing one-dimensional labels to indicate to which plane these labels belong;

[0048] A multi-dimensional, multi-faceted encoder that embeds the position information in the image into the feature representation and handles position and scale changes in the image;

[0049] A multi-classification layer structure, these markers are supplemented with classification markers that can be learned in multiple tasks, which are used to simultaneously predict multiple molecular pathological features such as IDH (isocitrate dehydrogenase) mutation, 1p / 19q (short arm of chromosome 1 and long arm of chromosome 19) co-deletion, and MGMT (O6-methylguanine-DNA methyltransferase) promoter methylation.

[0050] Embodiment 5, this embodiment is a further improvement on the basis of embodiment 4, and its details are as follows:

[0051] The multi-classification layer structure is specifically used for:

[0052] IDH mutation status: wild type and mutant;

[0053] 1p / 19q deletion status: deletion and no deletion;

[0054] MGMT methylation status: unmethylated and methylated.

[0055] Embodiment 6: This embodiment is a further improvement on the basis of embodiment 1, and its details are as follows:

[0056] The quantitative parameters of MRI imaging data include T1w (T1-weighted imaging), T1wC (T1-weighted imaging + contrast agent), T2w (T2-weighted imaging), and FLAIR (magnetic resonance imaging fluid attenuated inversion recovery sequence).

[0057] Embodiment 7, this embodiment is a further improvement on the basis of embodiment 1, and its details are as follows:

[0058] The quantitative parameters of DTI imaging data include ADC (mean diffusion coefficient) and FA (fractional anisotropy).

[0059] Embodiment 8: This embodiment is a further improvement on the basis of Embodiment 1, and its details are as follows:

[0060] The preprocessing in step S03 includes normalization, denoising and enhancement, which effectively preserves important features in the image and ensures the quality of the image data.

[0061] Although the embodiments of the present invention have been shown and described above, it is to be understood that the above embodiments are exemplary and are not to be construed as limitations of the present invention. A person skilled in the art may change, modify, replace and vary the above embodiments within the scope of the present invention.

Claims

1. A method for predicting multiple molecular pathologies of glioma based on deep learning of multimodal whole-brain MRI, characterized in that: The steps include: Step S01: input multi-modal whole-brain MRI imaging data and DTI imaging data; Step S02: extracting multi-dimensional, multi-slice two-dimensional CNN blocks from the MRI imaging data and the DTI imaging data; Step S03: preprocessing the two-dimensional CNN block; Step S04: analyzing the multi-dimensional, multi-slice image features obtained after the preprocessing, respectively extracting the axial plane feature Esag, the coronal plane feature Ecor, and the sagittal plane feature Eax, and calculating the multi-plane multi-slice feature S according to the Esag, the Ecor, and the Eax; Step S05: perform multi-plane and multi-slice image labeling on the S, add a transformer coding block with a classification head, and perform multi-molecular pathology prediction of IDH mutation, 1p / 19q co-deletion and MGMT promoter methylation.

2. According to claim 1, a method for predicting multiple molecular pathologies of glioma based on deep learning of multimodal whole-brain MRI is characterized in that: The extracted multi-dimensional, multi-slice image features in step S03 are calculated by an extraction operator E, and the calculation formula of the operator E is: R^C3d×L×W×H→R^C3d×N×W×H.

3. The method of predicting multiple molecular pathologies of glioma based on deep learning of multimodal whole-brain MRI according to claim 2, characterized in that: The calculation of the multi-plane multi-slice feature S in step S04 specifically includes using E=[Ecor, Esag, Eax] of the 3D representation feature X; the calculated S=[Scor, Ssag, Sax].

4. A method for predicting multiple molecular pathologies of glioma based on deep learning of multimodal whole-brain MRI according to any one of claims 1 to 3, characterized in that: The transformer coding block with classification header in step S05 includes: Non-Linear layer, which is used to reduce the dimension of the original image input; Multi-dimensional, multi-faceted embedding layer, where the multi-dimensional embedding layer is used to preserve position information; A multi-plane embedding layer for providing one-dimensional labels to indicate to which plane these labels belong; A multi-dimensional, multi-faceted encoder that embeds the position information in the image into the feature representation and handles position and scale changes in the image; A multi-classification hierarchical structure was used to simultaneously predict multiple molecular pathological features of IDH mutation, 1p / 19q co-deletion, and MGMT promoter methylation.

5. The method for predicting multiple molecular pathologies of glioma based on deep learning of multimodal whole-brain MRI according to claim 4, characterized in that: The multi-classification layer structure is specifically used for: IDH mutation status: wild type and mutant; 1p / 19q deletion status: deletion and no deletion; MGMT methylation status: unmethylated and methylated.

6. The method of predicting multiple molecular pathologies of glioma based on deep learning of multimodal whole brain MRI according to claim 1, characterized in that: The quantitative parameters of the MRI imaging data include T1w, T1wC, T2w and FLAIR.

7. The method of predicting multiple molecular pathologies of glioma by deep learning based on multimodal whole-brain MRI according to claim 1, characterized in that: The quantitative parameters of the DTI imaging data include ADC and FA.

8. The method of predicting multiple molecular pathologies of glioma based on deep learning of multimodal whole-brain MRI according to claim 1, characterized in that: The preprocessing in step S03 includes normalization, denoising and enhancement.

Images