An automatic segmentation method based on morphological features of the frontal visual pathway

Abstract

This automatic segmentation method incorporates the morphological characteristics of the anterior visual pathway. Compared to commonly used atlas segmentation methods and deep learning-based segmentation methods, this method effectively combines the advantages of both methods, overcoming the limitations of existing anterior visual pathway segmentation algorithms. While maintaining segmentation accuracy, it can also effectively display the path details of the anterior visual pathway, reduce the impact of tumor compression on nerve fiber segmentation accuracy, and more intuitively display the location information of the anterior visual pathway. It can provide an accurate and efficient method for the anterior visual pathway segmentation process.

Description

Technical Field

[0001] The present invention belongs to the fields of artificial intelligence and information technology, and relates to medical imaging and medical image segmentation technology under computer graphics, in particular to an automatic segmentation method based on morphological features of the frontal visual pathway. Background Art

[0002] With the advancement of medical imaging technology and computer equipment, the resolution and accuracy of medical images have continued to improve, leading to their widespread use in clinical practice and medical research. Medical image segmentation technology is an essential and important technology in many medical imaging applications. It can isolate key structures of interest to physicians before surgery, facilitating disease analysis, diagnosis, and surgical planning. The anterior visual pathway plays a crucial role in the visual system, transmitting visual information between the brain and the retina via the optic nerve, optic chiasm, optic tracts, and lateral geniculate body. The small size and tortuous geometry of the anterior visual pathway, as well as the challenging magnetic resonance imaging (MRI) environment surrounding it, make identifying the correct fiber tracts during dMRI imaging more challenging than identifying other fiber tracts in the forebrain. Due to limitations in imaging technology and other factors, dMRI imaging can produce a large number of false-positive fibers. Traditional segmentation methods require manual segmentation of the optic nerve, which consumes significant labor and resources. Furthermore, the accuracy of segmentation is highly limited by the physician's expertise, making large-scale optic nerve segmentation difficult to achieve.

[0003] In recent years, with the increasing attention paid to the neuroimaging of the anterior visual pathway, voxel-based automatic segmentation methods such as model-based and learning-based methods have been proposed for segmenting the anterior visual pathway on anatomical MRI (e.g., T1 / T2 weighted images). However, these methods cannot directly distinguish between nerve fibers that pass through the optic chiasm in the visual pathway or not. On the other hand, researchers use the results of voxel-based methods as ROIs for dMRI tracing imaging to identify optic nerve fibers, but their application is limited due to the inability to segment the anterior visual pathway in tumors. Similarly, a U-Net network-based anterior visual pathway segmentation method automatically segments the anterior visual pathway by fusing features from multiple modal data. Although the overall segmentation accuracy has increased, when interfered by tumor compression, the signal intensity of the nerve fibers in the anterior visual pathway almost merges with the tumor, resulting in a decrease in the stability of its segmentation accuracy.

[0004] Another approach for automatic cranial nerve (CN) identification is based on fiber clustering, which relies on the geometric structure and spatial location of fiber tracts and aims to group adjacent fibers with similar trajectories into clusters. Building on this approach, researchers have proposed an automatically annotated fiber clustering (AAFC) method that can identify anatomically significant white matter structures from whole-brain tracing images. This approach avoids the extensive labor and effort required to select and label the correct fiber clusters during neural atlas creation. By using cortical parcellation information across multiple subjects, each fiber tract is automatically annotated to segment potential nerve fibers. However, the anterior visual pathway does not directly connect to the cerebral cortex, making the AAFC method unsuitable for automatic segmentation of the optic nerve. While clustering methods can effectively reduce the impact of tumors on segmentation accuracy, since the anterior visual pathways vary from person to person, using the same template for segmentation registration results in lower segmentation accuracy under normal conditions than methods trained on deep learning networks. Summary of the Invention

[0005] To overcome the limitations of existing anterior visual pathway segmentation algorithms, the present invention proposes an automatic segmentation method that combines the morphological characteristics of the anterior visual pathway. While ensuring segmentation accuracy, it can also effectively display the direction details of the anterior visual pathway, reduce the impact of tumor compression on the accuracy of nerve fiber segmentation, and more intuitively display the position information of the anterior visual pathway.

[0006] The technical solution adopted by the present invention to solve its technical problem is:

[0007] An automatic segmentation method based on morphological features of the front visual pathway comprises the following steps:

[0008] Step 1: Image preprocessing: Obtain T1 images and DWI images from human brain MRI data, normalize and equalize the image grayscale values, and then generate the corresponding FA images;

[0009] Step 2: Generate data samples: Combine the T1 and FA images generated in step 1 and use manual labeling to combine the two types of images to obtain the morphological characteristics of the optic nerve, including the optic disc, optic chiasm, and lateral geniculate body;

[0010] Step 3: Create a neural fiber template: The dataset from Step 1 is subjected to whole-brain fiber tracking imaging. All data is then manually segmented by experts and aligned to the same common space using a similarity transformation. With expert participation in identification, a neural fiber template with the most geometric characteristics of the anterior visual pathway is extracted, ensuring that this template accurately contains the neural fibers of the anterior visual pathway to a higher degree of accuracy and is also universally applicable.

[0011] Step 4: Train the morphological feature extraction network: Use the labeled data from step 2 combined with DTI and FA images to train and generate a morphological feature extraction model for the anterior visual pathway;

[0012] Step 5: Automatic segmentation: First, the nerve fibers to be segmented are matched with the corresponding template to screen out the nerve fibers of the front visual pathway that are similar to the template; on this basis, the network trained in step 4 is used to extract the features of the front visual pathway, and the features are used as the morphological constraints of the front visual pathway to eliminate those nerve fibers that do not meet the morphological characteristics, thereby obtaining accurate segmentation results.

[0013] Furthermore, in step 2, the feature points marked for training the optic nerve fiber morphology feature extraction network are: the optic disc area, the optic chiasm area, and the lateral geniculate body area.

[0014] Furthermore, in step 3, the data is first imaged for whole-brain nerve fibers, and then, based on this, experts manually screen the data to obtain the accurate front visual pathway nerve fibers of each human brain, and then match them to a common space to obtain an overall universal nerve fiber template.

[0015] Furthermore, in step 4, the T1 image and the FA image are combined as the input of the network.

[0016] In step 5, the whole-brain nerve fibers are subjected to two constraint screenings. First, the whole-brain nerve fibers are matched with the nerve fiber template generated in step 3 to screen out the matching nerve fibers. On this basis, the morphological features generated by the network are used for secondary constraints to obtain more accurate automatic segmentation results.

[0017] The beneficial effects of the present invention are: while ensuring segmentation accuracy, it can also effectively display the direction details in the anterior visual pathway, reduce the impact of tumor compression on the accuracy of nerve fiber segmentation, and more intuitively display the position information of the anterior visual pathway. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 Flowchart of an automatic segmentation method based on morphological features of the anterior visual pathway. DETAILED DESCRIPTION

[0019] The present invention will be further described below with reference to the accompanying drawings.

[0020] Reference Figure 1 , an automatic segmentation method based on morphological features of the front visual pathway, comprising the following steps:

[0021] Step 1: Image preprocessing: Obtain T1 images and DWI images from human brain MRI data, normalize and equalize the image grayscale values, and then generate the corresponding FA images;

[0022] Step 2: Generate data samples: Combine the T1 and FA images generated in step 1 and use manual labeling to combine the two types of images to obtain the morphological characteristics of the optic nerve, including the optic disc, optic chiasm, and lateral geniculate body;

[0023] Feature points marked for training the optic nerve fiber morphology feature extraction network: optic disc area, optic chiasm area, and lateral geniculate body area;

[0024] Step 3: Create a neural fiber template: The dataset from Step 1 is subjected to whole-brain fiber tracking imaging. All data is then manually segmented by experts and aligned to the same common space using a similarity transformation. With expert participation in identification, a neural fiber template with the most geometric characteristics of the anterior visual pathway is extracted, ensuring that this template accurately contains the neural fibers of the anterior visual pathway to a higher degree of accuracy and is also universally applicable.

[0025] First, the data is imaged for the entire brain's neural fibers. Experts then manually screen the data to obtain the accurate neural fibers of each person's frontal visual pathway. These fibers are then mapped to a common space to create a universal neural fiber template.

[0026] Step 4: Train the morphological feature extraction network: Use the labeled data from step 2, combined with DTI and FA images, to train and generate a morphological feature extraction model for the anterior visual pathway; use the combination of T1 images and FA images as input to the network;

[0027] Step 5: Automatic segmentation: First, the nerve fibers to be segmented are matched with the corresponding template to screen out the nerve fibers of the front visual pathway that are similar to the template; on this basis, the network trained in step 4 is used to extract the features of the front visual pathway, and the features are used as the morphological constraints of the front visual pathway to eliminate those nerve fibers that do not meet the morphological characteristics, thereby obtaining accurate segmentation results.

[0028] Preferably, the whole-brain nerve fibers are subjected to two constraint screenings. First, the whole-brain nerve fibers are matched with the nerve fiber template generated in step 3 to screen out the matching nerve fibers. Then, on this basis, the morphological features generated by the network are used for secondary constraints to obtain more accurate automatic segmentation results.

[0029] The implementation process of this embodiment includes the following steps:

[0030] Step 1: Dataset Preparation: We screened high-quality dMRI data from the Human Connectome Project (HCP), covering multiple age groups and a balanced male-female ratio. We selected T1 images, DWI images, and corresponding FA images from the MRI images as experimental data.

[0031] Step 2: Data preprocessing: Preprocessing of the FA image and T1 image in the selected data includes: image cropping, histogram equalization, image normalization and training data extraction.

[0032] (1) Image cropping: The original data was cropped into 128 × 160 × 128 three-dimensional MRI data and saved in nii.gz format.

[0033] (2) Histogram equalization: In order to enhance the contrast of the image and reduce the differences between different data, histogram equalization is performed on the MRI image.

[0034] (3) Image normalization. Normalize the grayscale value of the image after the above histogram equalization to the range of 0-255. This operation does not change the image information, but can accelerate the network convergence in the subsequent training process.

[0035] (4) Generate training samples: Experts label the morphological features of the anterior visual pathway on this data and save each labeled label separately as a nii.gz file. At the same time, data augmentation is performed on the training dataset, and the original image data is mirrored and expanded, so that the final training data is expanded by four times the original size.

[0036] Step 3: Create a neural fiber template for the anterior visual pathway: First, generate a whole-brain neural fiber tracking image of all the data. Then, experts manually segment the neural fibers of the anterior visual pathway from the imaging results. These are then aligned to the same common space through a similarity transformation, ultimately obtaining a neural fiber template with the geometric characteristics of the anterior visual pathway.

[0037] Step 4, pre-training visual pathway morphological feature extraction network: Construct a U-Net network model, use the training samples generated in step 2 to train the constructed network model, and generate probability maps and segmentation masks for T1 and FA respectively; the 2D-Unet framework applied by the present invention, the encoder module contains 5 convolutional layers and maximum pooling layers, respectively containing 32, 64, 128, 256, and 512 feature maps; the decoder module contains 5 deconvolution layers and convolution layers, respectively containing 512, 256, 128, 64, and 32 feature maps. In the convolution layer, the size of all convolution kernels is 3×3×3. For all maximum pooling layers, the pool size is 2×2×2, and the step size is 2; for all deconvolution layers, the deconvolution feature map is combined with the corresponding features in the encoder module. After decoding, the Softmax classifier is used to generate voxel-level probability maps and predictions, and the Dice coefficient is used as the loss function of the network.

[0038] Step 5: Automatic Segmentation of the Anterior Visual Pathway: First, perform whole-brain fiber tracking imaging on the test data to obtain the data to be segmented. This data is then matched against the template generated in Step 3 to obtain a coarse segmentation. Additionally, the T1 and FA images of the test data are fed into the trained network in Step 4 to obtain relevant morphological features of the anterior visual pathway. Finally, the morphological features output by the trained network are used to constrain the coarse segmentation results, eliminating fibers that do not meet the morphological characteristics to obtain the final segmentation result.

[0039] Due to the adoption of the above technical solution, the present invention has the following advantages: it can effectively improve the accuracy of anterior visual pathway segmentation; reduce the impact of tumor compression on the automatic segmentation of the anterior visual pathway; and provide a stable, efficient and repeatable analysis method for the segmentation research of cranial nerves other than the optic nerve.

[0040] The specific implementation described above is only an optimal implementation method of the present invention and is not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made by utilizing the spirit and principles of the present invention and the contents of the drawings should be included in the patent protection scope of the present invention.

Claims

1. An automatic segmentation method based on morphological features of the front visual pathway, characterized in that: The segmentation method includes the following steps: Step 1: Image preprocessing: Obtain T1 images and DWI images from human brain MRI data, normalize and equalize the image grayscale values, and then generate the corresponding FA images; Step 2: Generate data samples: Combine the T1 and FA images generated in step 1 and use manual labeling to combine the two types of images to obtain the morphological characteristics of the optic nerve, including the optic disc, optic chiasm, and lateral geniculate body; Step 3: Create a neural fiber template: The dataset from Step 1 is subjected to whole-brain fiber tracking imaging. All data is then manually segmented by experts and aligned to the same common space using a similarity transformation. With expert participation in identification, a neural fiber template with the most geometric characteristics of the anterior visual pathway is extracted, ensuring that this template accurately contains the neural fibers of the anterior visual pathway to a higher degree of accuracy and is also universally applicable. Step 4: Train the morphological feature extraction network: Use the labeled data from step 2 combined with DTI and FA images to train and generate a morphological feature extraction model for the anterior visual pathway; Step 5: Automatic segmentation: First, the nerve fibers to be segmented are matched with the corresponding template to screen out the nerve fibers of the front visual pathway that are similar to the template; on this basis, the network trained in step 4 is used to extract the features of the front visual pathway, and the features are used as the morphological constraints of the front visual pathway to eliminate those nerve fibers that do not meet the morphological characteristics, thereby obtaining accurate segmentation results.

2. The automatic segmentation method based on morphological features of the front visual pathway according to claim 1, characterized in that: In step 2, the feature points marked for training the optic nerve fiber morphology feature extraction network are: the optic disc area, the optic chiasm area, and the lateral geniculate body area.

3. The automatic segmentation method based on morphological features of the front visual pathway according to claim 1 or 2, characterized in that: In step 3, the data is first imaged for whole-brain nerve fibers, and then manually screened by experts to obtain accurate anterior visual pathway nerve fibers of each human brain, which are then matched to a common space to obtain an overall universal nerve fiber template.

4. The automatic segmentation method based on morphological features of the front visual pathway according to claim 1 or 2, characterized in that: In step 4, a combination of T1 image and FA image is used as the input of the network.

5. The automatic segmentation method based on morphological features of the front visual pathway according to claim 1 or 2, characterized in that: In step 5, the whole-brain nerve fibers are subjected to two constraint screenings. First, the whole-brain nerve fibers are matched with the nerve fiber template generated in step 3 to screen out the matching nerve fibers. On this basis, the morphological features generated by the network are used for secondary constraints to obtain more accurate automatic segmentation results.

Images