Baby brain age prediction method, system and equipment based on T2 weighted imaging

Abstract

The invention relates to the technical field of image analysis, in particular to a baby brain age prediction method, system and equipment based on T2 weighted imaging, and aims to solve the technical problem of low baby brain age prediction accuracy in the prior art. The baby brain age prediction method comprises the following steps: firstly, based on T2 weighted imaging, extracting age-related brain local features and brain myelin sheath features under different scales; obtaining a multi-scale local feature and a multi-scale myelin sheath feature; then, the multi-scale local features and the multi-scale myelin sheath features are subjected to corresponding-scale element-by-element addition, complementarity and characterization capacity of the multi-scale local features and the multi-scale myelin sheath features in a brain age prediction task are enhanced, self-adaptive multi-branch feature fusion processing is performed after multi-scale myelin sheath enhanced features are obtained, and multi-scale brain fusion features are obtained; performing element-by-element multiplication on the channel attention information of the multi-scale brain fusion feature and the deep myelin sheath feature to obtain a whole brain enhancement feature; and finally, carrying out linear and nonlinear processing on the whole-brain enhancement features to obtain an infant brain age prediction result.

Description

Technical Field

[0001] The present invention relates to the technical field of image analysis, and in particular to a method, system and device for predicting infant brain age based on T2-weighted imaging. Background Art

[0002] Structural magnetic resonance imaging (MRI) is a non-invasive imaging technology that can provide high-resolution image information of brain anatomical structures. The richness and high quality of MRI data support the development of brain age prediction algorithms, giving it significant advantages in early infant brain development research, abnormal development screening and intervention strategies.

[0003] At present, convolutional neural networks (CNNs) are often used to extract MRI features and predict brain age. They also include the use of image information in single-modality or multi-modality neuroimaging technology to predict brain age. However, these technologies have weak expression capabilities for image features obtained on infant brain images, and the accuracy of infant brain age prediction results is low. Some studies have focused on the fact that myelin information in neuroimaging is an important biological feature of rapid development in infancy. Myelin information can be used as a neurodevelopmental indicator to predict infant brain age. However, myelin information currently often relies on doctors to manually outline the myelin area. This process is not only time-consuming and labor-intensive, but also results in poor final brain age prediction results and unstable prediction quality. Summary of the invention

[0004] The purpose of the present invention is to provide a method, system and device for predicting infant brain age based on T2-weighted imaging.

[0005] The technical solution of the present invention is as follows: A method for predicting infant brain age based on T2-weighted imaging, comprising the following operations: S1. Based on T2-weighted imaging, whole-brain images and white matter images are obtained; local features of different scales of whole-brain images are obtained to obtain shallow local features, middle local features and deep local features; specifically: down-sampling and dense feature stacking processing are performed on the whole-brain images respectively, and the down-sampling and dense feature stacking processing are repeated twice, and the down-sampling and dense feature stacking processing are repeated three times to obtain shallow local features, middle local features and deep local features; myelin information of different scales of white matter images is obtained to obtain shallow myelin features, middle myelin features and deep myelin features; S2. Add the shallow local features, middle local features, and deep local features to the shallow myelin features, middle myelin features, and deep myelin features element by element, respectively, to obtain the shallow myelin enhancement features, middle myelin enhancement features, and deep myelin enhancement features; the shallow myelin enhancement features, middle myelin enhancement features, and deep myelin enhancement features are processed by adaptive multi-branch feature fusion to obtain multi-scale brain fusion features; the deep myelin features are processed by convolution and channel attention to obtain myelin channel attention features; the myelin channel attention features and the multi-scale brain fusion features are multiplied element by element to obtain the whole brain enhancement features; S3. The whole-brain enhanced features are processed linearly and nonlinearly to obtain the infant brain age prediction results.

[0006] The specific operation of obtaining the shallow myelin sheath feature in S1 is as follows: the white matter image is processed by local detail enhancement in different directions to obtain multi-directional myelin sheath detail enhancement features; the multi-directional myelin sheath detail enhancement features are convolved and average pooled to obtain the shallow myelin sheath features.

[0007] The specific operation of obtaining the middle layer myelin sheath features in S1 is as follows: the shallow layer myelin sheath features are processed by spatial attention to obtain the myelin sheath spatial features; the myelin sheath spatial features and the shallow layer myelin sheath features are multiplied element by element, and then convolution and average pooling are performed to obtain the middle layer myelin sheath features.

[0008] The deep myelin features in S1 are obtained by performing local detail enhancement, convolution and average pooling on the middle myelin features in different directions.

[0009] The specific operations of the adaptive multi-branch feature fusion processing in S2 are as follows: the shallow myelin enhancement features, the middle myelin enhancement features and the deep myelin enhancement features are respectively subjected to maximum pooling and depthwise separable convolution, and then spliced ​​to obtain multi-scale brain fusion features.

[0010] The operations for obtaining the whole brain image and white matter image in S1 are as follows: the T2-weighted image is subjected to voxel normalization and then skull information is stripped to obtain an initial whole brain image; the initial whole brain image is subjected to white matter segmentation to obtain an initial white matter image; the initial whole brain image is subjected to deviation correction and then registered with the initial white matter image to obtain a whole brain image and a white matter image.

[0011] The specific operations of channel attention processing in S1 are as follows: the deep myelin features are processed by convolution to obtain the deep myelin convolution features; the deep myelin convolution features are processed by average pooling to obtain the deep myelin pooling features; the deep myelin pooling features are fully connected and then multiplied element-by-element with the deep myelin pooling features to obtain the initial myelin attention features; the initial myelin attention features are processed nonlinearly to obtain the myelin channel attention features.

[0012] A system for predicting infant brain age based on T2-weighted imaging, used to implement the above-mentioned method for predicting infant brain age based on T2-weighted imaging, comprising: The multi-scale local feature and multi-scale myelin feature generation module is used to obtain the whole brain image and white matter image based on T2-weighted imaging; obtain the local features of the whole brain image at different scales to obtain the shallow local features, middle local features and deep local features; specifically: perform downsampling and dense feature stacking processing on the whole brain image, repeat the downsampling and dense feature stacking processing twice, and repeat the downsampling and dense feature stacking processing three times to obtain the shallow local features, middle local features and deep local features; obtain the myelin information of the white matter image at different scales to obtain the shallow myelin features, middle myelin features and deep myelin features; The whole-brain enhancement feature generation module is used to add the shallow local features, the middle local features, and the deep local features to the shallow myelin features, the middle myelin features, and the deep myelin features element by element, respectively, to obtain the shallow myelin enhancement features, the middle myelin enhancement features, and the deep myelin enhancement features; the shallow myelin enhancement features, the middle myelin enhancement features, and the deep myelin enhancement features are processed by adaptive multi-branch feature fusion to obtain multi-scale brain fusion features; the deep myelin features are processed by convolution and channel attention to obtain myelin channel attention features; the myelin channel attention features and the multi-scale brain fusion features are multiplied element by element to obtain the whole-brain enhancement features; The module for generating the prediction results of infant brain age is used to obtain the prediction results of infant brain age by linear and nonlinear processing of the whole brain enhancement features.

[0013] A device for predicting the brain age of an infant based on T2 weighted imaging comprises a processor and a memory, wherein the processor implements the above-mentioned method for predicting the brain age of an infant based on T2 weighted imaging when executing a computer program stored in the memory.

[0014] A computer-readable storage medium is used to store a computer program, wherein when the computer program is executed by a processor, the above-mentioned method for predicting infant brain age based on T2-weighted imaging is implemented.

[0015] The beneficial effects of the present invention are: The present invention provides a method for predicting infant brain age based on T2-weighted imaging. First, based on T2-weighted imaging, whole-brain images and white matter images that can respectively reflect the global structure and myelination characteristics of the whole brain are obtained, and local brain features and brain myelin features related to age at different scales are extracted to obtain multi-scale local features and multi-scale myelin features; then, the multi-scale local features and multi-scale myelin features are added element by element at corresponding scales, the myelination information is integrated with the age-related features in the whole-brain channel, the complementarity and representation ability of the two in the brain age prediction task are enhanced, and after the multi-scale myelin enhancement features are obtained, adaptive multi-branch feature fusion processing is performed to obtain multi-scale brain fusion features; the multi-scale brain fusion features are multiplied element by element with the channel attention information of the deep myelin features to obtain the whole-brain enhancement features; finally, the whole-brain enhancement features are linearly and nonlinearly processed to obtain the infant brain age prediction results. DETAILED DESCRIPTION

[0016] This embodiment provides a method for predicting infant brain age based on T2-weighted imaging, including the following operations: S1. Based on T2-weighted imaging, whole-brain images and white matter images are obtained; local features of different scales of whole-brain images are obtained to obtain shallow local features, middle local features and deep local features; specifically: down-sampling and dense feature stacking processing are performed on the whole-brain images respectively, and the down-sampling and dense feature stacking processing are repeated twice, and the down-sampling and dense feature stacking processing are repeated three times to obtain shallow local features, middle local features and deep local features; myelin information of different scales of white matter images is obtained to obtain shallow myelin features, middle myelin features and deep myelin features; S2. Add the shallow local features, middle local features, and deep local features to the shallow myelin features, middle myelin features, and deep myelin features element by element, respectively, to obtain the shallow myelin enhancement features, middle myelin enhancement features, and deep myelin enhancement features; the shallow myelin enhancement features, middle myelin enhancement features, and deep myelin enhancement features are processed by adaptive multi-branch feature fusion to obtain multi-scale brain fusion features; the deep myelin features are processed by convolution and channel attention to obtain myelin channel attention features; the myelin channel attention features and the multi-scale brain fusion features are multiplied element by element to obtain the whole brain enhancement features; S3. The whole-brain enhanced features are processed linearly and nonlinearly to obtain the infant brain age prediction results.

[0017] S1. Based on T2-weighted imaging, whole brain images and white matter images are obtained; local features of whole brain images at different scales are obtained to obtain shallow local features, middle local features and deep local features; myelin information of white matter images at different scales is obtained to obtain shallow myelin features, middle myelin features and deep myelin features.

[0018] Based on T2-weighted imaging, whole-brain images and white matter images are obtained, which can reflect the global structure and myelination characteristics of the whole brain respectively. The age-related local brain features and brain myelin features at different scales are extracted to obtain multi-scale local features and multi-scale myelin features.

[0019] First, based on T2-weighted imaging, whole-brain images and white matter images were obtained. Specifically, the original data collected in the hospital were T2-weighted images in the Digital Imaging and Communications in Medicine (DICOM) format. The T2-weighted images were converted into three-dimensional images in the Neuroimaging Informatics Technology Initiative (NIFTI) format, and the patient privacy data was deleted. Then, the T2-weighted images were voxel-normalized to 1×1×1 mm, and then the skull information was stripped using FSL software to obtain the initial whole-brain image. The initial whole-brain image was processed by white matter segmentation using FSL software to ensure accurate extraction of the white matter area to obtain the initial white matter image. The initial whole-brain image was processed by bias correction to eliminate low-frequency brightness unevenness in the image, and was registered with the initial white matter image, and registered to the UNC Infant 0-1-2 atlas, respectively. After ensuring spatial consistency, it was cropped to 150×198×156 to obtain the whole-brain image and white matter image.

[0020] Then, local features of the whole brain image at different scales are obtained, structural features related to age at different scales are extracted, local features of the whole brain are enriched, and shallow local features, middle local features, and deep local features are obtained. Specifically, the whole brain image is downsampled and dense feature stacking is performed, the downsampling and dense feature stacking are repeated twice, and the downsampling and dense feature stacking are repeated three times to achieve feature reuse and information transfer at different levels, enhance the expression ability of local features, and obtain shallow local features, middle local features, and deep local features. Dense feature stacking includes but is not limited to being realized through dense convolutional connection network DenseNet, and downsampling includes but is not limited to being realized through convolution blocks.

[0021] At the same time, myelin information at different scales of white matter images is obtained to enrich the myelin features, and obtain shallow myelin features, middle myelin features and deep myelin features.

[0022] The operation of obtaining the shallow myelin feature is as follows: the white matter image is subjected to local detail enhancement processing in different directions to obtain multi-directional myelin detail enhancement features; the multi-directional myelin detail enhancement features are subjected to convolution and average pooling to obtain the shallow myelin feature. The local detail enhancement processing in different directions can be achieved by performing conventional convolution, central difference convolution, horizontal difference convolution and vertical difference convolution on the white matter image and then adding them element by element to highlight the central pixel features, horizontal edge features and vertical edge features of the white matter image, and enhance the details of the myelin area in the white matter image.

[0023] The specific operation of obtaining the middle layer myelin features is as follows: the shallow layer myelin features are processed by spatial attention, the characteristic distribution of the myelin area is learned, and the myelin spatial features are obtained; the myelin spatial features and the shallow layer myelin features are multiplied element by element, and then convolution and average pooling are performed to automatically focus on the detailed information of the myelin in the white matter of the brain, strengthen the expression of the myelin area, and obtain the middle layer myelin features. The specific operation of spatial attention processing is as follows: the shallow layer myelin features are processed by average pooling and maximum pooling, and then channel dimension splicing, three-dimensional convolution and Sigmoid function are performed to obtain the myelin spatial features.

[0024] The deep myelin features are obtained by performing local detail enhancement, convolution and average pooling on the middle-layer myelin features in different directions.

[0025] S2. Add the shallow local features, middle local features, and deep local features to the shallow myelin features, middle myelin features, and deep myelin features element by element, respectively, to obtain shallow myelin enhancement features, middle myelin enhancement features, and deep myelin enhancement features; the shallow myelin enhancement features, middle myelin enhancement features, and deep myelin enhancement features are processed by adaptive multi-branch feature fusion to obtain multi-scale brain fusion features; the deep myelin features are processed by convolution and channel attention to obtain myelin channel attention features; the myelin channel attention features and the multi-scale brain fusion features are multiplied element by element to obtain the whole brain enhancement features.

[0026] The multi-scale local features and multi-scale myelin features are added element by element at the corresponding scales, and the myelination information is integrated with the age-related features in the whole-brain channel to enhance the complementarity and representation ability of the two in the brain age prediction task. After obtaining the multi-scale myelin enhancement features, adaptive multi-branch feature fusion processing is performed to obtain the multi-scale brain fusion features; the multi-scale brain fusion features are multiplied element by element with the channel attention information of the deep myelin features to obtain the whole-brain enhancement features.

[0027] Firstly, the shallow local features, middle local features, and deep local features are added element by element with the shallow myelin features, middle myelin features, and deep myelin features, respectively, to achieve the fusion of local features and myelin features of brain regions at corresponding scales, and integrate the myelination information extracted from the white matter myelin channel with the age-related features in the whole brain channel to enhance their complementarity and representation ability in the brain age prediction task, and obtain the shallow myelin enhancement features, middle myelin enhancement features, and deep myelin enhancement features, which reflect the local detail features and myelin information of brain regions at different levels, respectively.

[0028] Then, the shallow myelin enhancement features, middle myelin enhancement features and deep myelin enhancement features are processed by adaptive multi-branch feature fusion. On the basis of extracting features of different scales, multi-scale information fusion is realized to obtain multi-scale brain fusion features with rich features and prominent details. The specific operation of adaptive multi-branch feature fusion processing is as follows: the shallow myelin enhancement features, middle myelin enhancement features and deep myelin enhancement features are spliced ​​after maximum pooling and depthwise separable convolution respectively to obtain multi-scale brain fusion features. In the depthwise separable convolution processing, the input is convolved channel by channel, only focusing on the spatial information in each channel, and then point-by-point convolution is performed to integrate the outputs of each channel.

[0029] At the same time, the deep myelin features are processed by convolution and channel attention to better focus on the myelin information and obtain the myelin channel attention features. The specific operation of channel attention processing is as follows: the deep myelin features are processed by convolution to obtain the deep myelin convolution features; the deep myelin convolution features are processed by average pooling to obtain the deep myelin pooling features; the deep myelin pooling features are processed by full connection (composed of 2 fully connected layers) and multiplied element by element with the deep myelin pooling features to obtain the initial myelin attention features; the initial myelin attention features are processed by nonlinearity (which can be achieved by Sigmoid function) to obtain the myelin channel attention features.

[0030] Finally, the myelin channel attention features and the multi-scale brain fusion features are multiplied element by element to further enhance the complementarity and representation ability of the whole-brain structural information and myelin information, and obtain the whole-brain enhanced features.

[0031] S3. The whole-brain enhanced features are processed linearly and nonlinearly to obtain the infant brain age prediction results.

[0032] The whole-brain enhanced features are processed linearly and nonlinearly (which can be achieved through Linear function and Softmax function respectively) to obtain the infant brain age prediction results.

[0033] To verify the effect of the method of this embodiment, the following experiment was conducted.

[0034] Experimental settings: The dataset collected in the hospital was randomly divided into training set, test set and validation set in a ratio of 8:1:1. The experimental development environment used was pytorch2.1.0, and training was performed on the NVIDIA RTX 3090 graphics processor. A total of 100 rounds were performed during the model training period. The Adam optimizer was used, the initial learning rate was, the batch size was set to 5, and the model was trained for about 12 hours. The experimental environment and specific settings are shown in Table 1.

[0035] Table 1 Summary of experimental parameter settings

[0036] The experimental results are shown in Table 2. It can be seen that the prediction method of this embodiment has a small error in predicting the brain age of infants and a high prediction accuracy.

[0037] Table 2 Prediction results of infant brain age in this example

[0038] This embodiment further provides a system for predicting infant brain age based on T2-weighted imaging, which is used to implement the above-mentioned method for predicting infant brain age based on T2-weighted imaging, including: The multi-scale local feature and multi-scale myelin feature generation module is used to obtain the whole brain image and white matter image based on T2-weighted imaging; obtain the local features of the whole brain image at different scales to obtain the shallow local features, middle local features and deep local features; specifically: perform downsampling and dense feature stacking processing on the whole brain image, repeat the downsampling and dense feature stacking processing twice, and repeat the downsampling and dense feature stacking processing three times to obtain the shallow local features, middle local features and deep local features; obtain the myelin information of the white matter image at different scales to obtain the shallow myelin features, middle myelin features and deep myelin features; The whole-brain enhancement feature generation module is used to add the shallow local features, the middle local features, and the deep local features to the shallow myelin features, the middle myelin features, and the deep myelin features element by element, respectively, to obtain the shallow myelin enhancement features, the middle myelin enhancement features, and the deep myelin enhancement features; the shallow myelin enhancement features, the middle myelin enhancement features, and the deep myelin enhancement features are processed by adaptive multi-branch feature fusion to obtain multi-scale brain fusion features; the deep myelin features are processed by convolution and channel attention to obtain myelin channel attention features; the myelin channel attention features and the multi-scale brain fusion features are multiplied element by element to obtain the whole-brain enhancement features; The module for generating the prediction results of infant brain age is used to obtain the prediction results of infant brain age by linear and nonlinear processing of the whole brain enhancement features.

[0039] This embodiment also provides a device for predicting infant brain age based on T2-weighted imaging, comprising a processor and a memory, wherein the processor implements the above-mentioned method for predicting infant brain age based on T2-weighted imaging when executing a computer program stored in the memory.

[0040] This embodiment also provides a computer-readable storage medium for storing a computer program, wherein the computer program, when executed by a processor, implements the above-mentioned method for predicting infant brain age based on T2-weighted imaging.

[0041] The present embodiment provides a method for predicting the brain age of infants based on T2-weighted imaging. First, based on T2-weighted imaging, a whole-brain image and a white matter image that can respectively reflect the global structure and myelination characteristics of the whole brain are obtained, and local brain features and brain myelin features related to age at different scales are extracted to obtain multi-scale local features and multi-scale myelin features; then, the multi-scale local features and multi-scale myelin features are added element by element at corresponding scales, the myelination information is integrated with the age-related features in the whole-brain channel, and the complementarity and representation ability of the two in the brain age prediction task are enhanced, and after the multi-scale myelin enhancement features are obtained, adaptive multi-branch feature fusion processing is performed to obtain multi-scale brain fusion features; the multi-scale brain fusion features are multiplied element by element with the channel attention information of the deep myelin features to obtain the whole-brain enhancement features; finally, the whole-brain enhancement features are linearly and nonlinearly processed to obtain the infant brain age prediction results.

[0042] This embodiment provides a method for predicting infant brain age based on T2-weighted imaging, which significantly improves the accuracy and robustness of prediction by collaboratively extracting myelin information from whole-brain images and white matter images. The fusion mechanism effectively integrates myelination characteristics with age-related information, providing a reliable biomarker for early neurodevelopmental assessment of infants, and has important clinical application potential.

Claims

1. A method for predicting infant brain age based on T2-weighted imaging, characterized in that: The following operations are included: S1, based on T2-weighted imaging, obtain whole brain images and white matter images; Acquire local features of different scales of the whole brain image to obtain shallow local features, middle local features and deep local features; specifically: perform downsampling and dense feature stacking processing on the whole brain image, repeat the downsampling and dense feature stacking processing twice, and repeat the downsampling and dense feature stacking processing three times to obtain shallow local features, middle local features and deep local features; Obtain myelin information of different scales in white matter images to obtain shallow myelin features, middle myelin features and deep myelin features; S2, adding the shallow local features, the middle local features, and the deep local features to the shallow myelin sheath features, the middle myelin sheath features, and the deep myelin sheath features, element by element, to obtain the shallow myelin sheath enhancement features, the middle myelin sheath enhancement features, and the deep myelin sheath enhancement features; The superficial myelin enhancement features, the middle myelin enhancement features and the deep myelin enhancement features are processed by adaptive multi-branch feature fusion to obtain multi-scale brain fusion features; The deep myelin features are processed by convolution and channel attention to obtain the myelin channel attention features; the myelin channel attention features are multiplied element by element with the multi-scale brain fusion features to obtain the whole brain enhancement features; S3. The whole-brain enhanced features are processed linearly and nonlinearly to obtain the infant brain age prediction results.

2. The method for predicting infant brain age based on T2-weighted imaging according to claim 1, characterized in that: The operation of obtaining the superficial myelin sheath characteristics in S1 is specifically as follows: The white matter image is processed by local detail enhancement in different directions to obtain multi-directional myelin detail enhancement features; the multi-directional myelin detail enhancement features are processed by convolution and average pooling to obtain shallow myelin features.

3. The method for predicting infant brain age based on T2-weighted imaging according to claim 1, characterized in that: The operation of obtaining the middle layer myelin sheath characteristics in S1 is specifically as follows: The shallow myelin features are processed by spatial attention to obtain the myelin spatial features; the myelin spatial features and the shallow myelin features are multiplied element by element, and then convolution and average pooling are performed to obtain the middle myelin features.

4. The method for predicting infant brain age based on T2-weighted imaging according to claim 1, characterized in that: In S1, the deep myelin sheath features are obtained by performing local detail enhancement processing, convolution and average pooling processing on the middle myelin sheath features in different directions.

5. The method for predicting infant brain age based on T2-weighted imaging according to claim 1, characterized in that: In S2, the operation of the adaptive multi-branch feature fusion processing is specifically as follows: The shallow myelin enhancement features, middle myelin enhancement features, and deep myelin enhancement features are respectively subjected to maximum pooling and depthwise separable convolution, and then spliced ​​to obtain multi-scale brain fusion features.

6. The method for predicting infant brain age based on T2-weighted imaging according to claim 1, characterized in that: In S1, the operations of obtaining the whole brain image and the white matter image are specifically as follows: The T2-weighted image was processed by voxel normalization and skull information was stripped to obtain the initial whole-brain image; the initial whole-brain image was processed by white matter segmentation to obtain the initial white matter image; the initial whole-brain image was processed by deviation correction and then registered with the initial white matter image to obtain the whole-brain image and white matter image.

7. The method for predicting infant brain age based on T2-weighted imaging according to claim 1, characterized in that: In S1, the operation of channel attention processing is specifically as follows: The deep myelin sheath features are processed by convolution to obtain the deep myelin sheath convolution features; the deep myelin sheath convolution features are processed by average pooling to obtain the deep myelin sheath pooling features; the deep myelin sheath pooling features are fully connected and then multiplied element by element with the deep myelin sheath pooling features to obtain the initial myelin sheath attention features; the initial myelin sheath attention features are processed by nonlinearity to obtain the myelin channel attention features.

8. A system for predicting infant brain age based on T2-weighted imaging, used to implement the method for predicting infant brain age based on T2-weighted imaging according to claim 1, characterized in that: include: The multi-scale local feature and multi-scale myelin feature generation module is used to obtain the whole brain image and white matter image based on T2-weighted imaging; obtain the local features of the whole brain image at different scales to obtain the shallow local features, middle local features and deep local features; specifically: perform downsampling and dense feature stacking processing on the whole brain image, repeat the downsampling and dense feature stacking processing twice, and repeat the downsampling and dense feature stacking processing three times to obtain the shallow local features, middle local features and deep local features; obtain the myelin information of the white matter image at different scales to obtain the shallow myelin features, middle myelin features and deep myelin features; The whole brain enhancement feature generation module is used to add the shallow local features, the middle local features, and the deep local features to the shallow myelin features, the middle myelin features, and the deep myelin features element by element, respectively, to obtain the shallow myelin enhancement features, the middle myelin enhancement features, and the deep myelin enhancement features; The shallow myelin enhancement features, middle myelin enhancement features and deep myelin enhancement features are processed by adaptive multi-branch feature fusion to obtain multi-scale brain fusion features; the deep myelin features are processed by convolution and channel attention to obtain myelin channel attention features; the myelin channel attention features and the multi-scale brain fusion features are multiplied element by element to obtain the whole brain enhancement features; The module for generating the prediction results of infant brain age is used to obtain the prediction results of infant brain age by linear and nonlinear processing of the whole brain enhancement features.

9. A device for predicting infant brain age based on T2-weighted imaging, characterized in that: The method comprises a processor and a memory, wherein when the processor executes the computer program stored in the memory, the method for predicting the infant brain age based on T2-weighted imaging as described in any one of claims 1 to 7 is implemented.

10. A computer-readable storage medium, characterized in that: Used to store a computer program, wherein when the computer program is executed by a processor, the method for predicting the infant brain age based on T2-weighted imaging as described in any one of claims 1 to 7 is implemented.

Images