Audible ossicle automatic positioning and segmentation method and system suitable for three-dimensional CT (Computed Tomography) image

Abstract

The invention provides an auditory ossicle automatic positioning and segmentation method and system suitable for a three-dimensional CT image, and belongs to the technical field of image processing, and the method comprises the steps: S1, carrying out the data preprocessing of an original three-dimensional CT image, and enabling the original three-dimensional CT image to serve as a target image; s2, performing ossicle positioning on the target image to obtain ossicle positioning points, and cutting out an ossicle target area through the ossicle positioning points; and S3, segmenting the ossicle target area to obtain an ossicle three-dimensional segmentation result. According to the method, the auditory ossicle can be automatically positioned on the three-dimensional CT image within a short time, and corresponding segmentation marks of the hammer bone, the incus and the stapes are generated.

Description

Technical Field

[0001] This invention belongs to the field of image processing technology, and specifically relates to a method and system for automatic localization and segmentation of ossicles in three-dimensional CT images. Background Technology

[0002] The ossicles are the smallest bones in the human body, located in the tympanic cavity of the middle ear. They consist of the malleus, incus, and stapes, connected by ligaments and joints to form the ossicular chain. The ossicular chain is responsible for transmitting sound waves sensed by the tympanic membrane to the inner ear, making it a crucial link in sound conduction. Ossicular chain damage is usually caused by otitis media or head trauma, and may lead to symptoms such as hearing loss, ear pain, and deafness. Otitis media can cause necrosis, sclerosis, and adhesion of the ossicles; head trauma can cause ossicular dislocation or fracture. Diagnosis of ossicular chain damage relies on ear CT images. High-quality three-dimensional imaging technology, with its advantages of good hard tissue imaging quality, low radiation dose, and fast reconstruction speed, is widely used in the diagnosis and treatment of oral and maxillofacial and temporal bones.

[0003] The location and morphological characteristics of the ossicles are crucial for diagnosing ossicular chain injuries. Currently, doctors need to manually annotate the ossicular structure on multiple two-dimensional images to obtain a three-dimensional model of the ossicles for further observation and measurement. This process is highly dependent on the doctor's experience, and manual annotation is laborious and time-consuming. Because two-dimensional images cannot fully display the three-dimensional structure of the ossicles, and due to the small size of the ossicles and the decreased imaging quality caused by otitis media, the accuracy of annotation is difficult to guarantee, leading to low diagnostic efficiency and misdiagnosis. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides an automatic localization and segmentation method for ossicles in three-dimensional CT images, comprising the following steps:

[0005] Step S1: Perform data preprocessing on the original 3D CT image to obtain the target image;

[0006] Step S2: Locate the ossicles in the target image, obtain the ossicle positioning points, and crop out the target area of ​​the ossicles using the ossicle positioning points;

[0007] Step S3: Segment the target region of the ossicles to obtain the three-dimensional segmentation result of the ossicles.

[0008] Beneficial effects:

[0009] 1. This invention proposes an automated ossicle localization and segmentation network, which obtains the ossicle localization points, target areas of the ossicles, and segmentation marks of the malleus, incus, and stapes on 3D CT images. This saves doctors time in marking the ossicles and improves diagnostic efficiency.

[0010] 2. The ossicle localization and segmentation network of the present invention meets the needs of doctors in terms of both computation time and accuracy, and can provide fast and accurate ossicle marking results, helping doctors to diagnose the ossicular chain and reducing the risk of misdiagnosis. Attached Figure Description

[0011] Figure 1 This is a schematic diagram of the automatic ossicle localization and segmentation method applicable to three-dimensional CT images according to the present invention;

[0012] Figure 2 A schematic diagram of the neural network for ossicle localization;

[0013] Figure 3 A schematic diagram of the neural network structure for segmenting the ossicles;

[0014] Figure 4 This is a structural block diagram of an automatic ossicle localization and segmentation system for three-dimensional CT images according to the present invention. Detailed Implementation

[0015] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0016] Example 1

[0017] like Figure 1 As shown in the figure, an embodiment of the present invention provides an automatic localization and segmentation method for ossicles in three-dimensional CT images, comprising the following steps:

[0018] Step S1: Perform data preprocessing on the original 3D CT image to obtain the target image;

[0019] Step S2: Locate the ossicles in the target image, obtain the ossicle positioning points, and crop out the target area of ​​the ossicles using the ossicle positioning points;

[0020] Step S3: Segment the target region of the ossicles to obtain the three-dimensional segmentation result of the ossicles.

[0021] In one embodiment, step S1 above: preprocessing the original three-dimensional CT image into a target image specifically includes: preprocessing the original three-dimensional CT image into a target image, including downsampling to a preset size and image z-score normalization, to obtain the target image.

[0022] In one embodiment, step S2 above: locating the ossicles of the target image, obtaining ossicle positioning points, and cropping the target ossicle region using the ossicle positioning points, specifically includes:

[0023] Step S21: Input the target image into the ossicle localization neural network to obtain the ossicle localization points;

[0024] First, a dataset for ossicular localization was constructed. Three-dimensional CT images containing the ossicles were collected, with the same number of samples from different age groups and genders. The malleus, incus, and stapes regions were labeled on the images, with all ossicular regions set to 1 and all background regions set to 0. This data served as training data for the ossicular localization algorithm.

[0025] Secondly, an improved ossicle localization neural network based on 3D-Unet was constructed to segment the regions of background and foreground ossicles for localization.

[0026] To improve the accuracy and performance of the algorithm, improvements can be made to the conventional 3D-Unet:

[0027] 1) Set the number of downsampling layers to 4 to fit the size of the input image.

[0028] 2) At the jump connection between the encoder and decoder, convolution is used to increase the dimension, which is directly added to the upper layer features. This reduces the model's memory usage and inference time without reducing accuracy.

[0029] Next, fixed-size voxel blocks are cropped from the images in the ossicle localization dataset and used as training data to train the ossicle localization neural network. During training, data augmentation methods such as random image flipping and random occlusion are used. The Dice similarity coefficient (DSC) is used as the loss function.

[0030] Finally, the trained ossicle localization neural network is obtained.

[0031] Figure 2 A schematic diagram of the structure of the ossicle localization neural network is shown.

[0032] Step S22: Using the ossicle positioning point as the center, crop out a fixed-size image from the original 3D CT image as the target area for the ossicles.

[0033] To minimize computational resource consumption, a block-based prediction method is employed, dividing the target image into multiple fixed-size voxel blocks with overlap. These blocks are sequentially input into a trained ossicle localization neural network. After obtaining the segmentation results of all voxel blocks, a voting and merging process is performed to form a complete segmentation result. Localization points (x, y, z) are constructed from the largest areas in the axial, coronal, and sagittal planes using the complete segmentation results. These localization points are then used to form the final ossicle localization points, centered on the original image dimensions. A fixed-size image is then cropped from the original 3D CT image using these ossicle localization points as the target region for the ossicles.

[0034] In one embodiment, step S3 above, which involves segmenting the target region of the ossicles to obtain a three-dimensional segmentation result of the ossicles, specifically includes:

[0035] Step S31: Construct the ossicle segmentation dataset. Input the ossicle localization dataset into the ossicle localization algorithm to obtain images of all ossicle target regions. Distinguish between the malleus, incus, stapes, and background regions. The labels are represented by different numerical values ​​to complete the construction of the ossicle segmentation dataset.

[0036] Step S32: Construct an improved residual 3D-UNet network as a neural network for ossicle segmentation. The specific steps are as follows:

[0037] 1) Use deconvolution operations in the 3D-UNet decoder to concatenate low-level and high-level features.

[0038] 2) Each layer of the entire network uses a residual convolution module for feature extraction;

[0039] 3) Finally, the Softmax activation function is used to convert the obtained feature output into predicted probabilities of the malleus, incus, and stapes.

[0040] Step S33: Train the ossicle segmentation neural network, using data augmentation methods such as random image flipping and random angle rotation during training; use a combination of Dice similarity coefficient (DSC) and multi-class cross-entropy (CE) as the loss function, as shown in the following formula:

[0041]

[0042] Where N is the number of samples, specifically the number of voxels output by the segmentation network; C is the number of categories, and in this embodiment, N is 4; y ic This is the true label of sample i, encoded using one-hot encoding; p ic ε is the predicted probability that sample i belongs to category c, which is obtained from the output of the segmentation model; ε is a smoothing factor to prevent the denominator from being 0.

[0043] Finally, a well-trained ossicle segmentation neural network was obtained.

[0044] Figure 3 A schematic diagram of the structure of the ossicle segmentation neural network is shown.

[0045] Step S34: Input the target region of the ossicles from step S22 into the trained ossicle segmentation neural network, outputting voxel-level prediction probabilities. Set a threshold; voxels greater than the threshold are considered foreground results, and those less than the threshold are considered background results, thus obtaining segmentation results for the three ossicles. In post-processing, analyze the connected components of the prediction results for each ossicle, remove noise and fill holes, and finally generate a labeled model for each ossicle.

[0046] Example 2

[0047] like Figure 4 As shown, this embodiment of the invention provides an automatic ossicle localization and segmentation system suitable for three-dimensional CT images, comprising the following modules:

[0048] The target image acquisition module 41 performs data preprocessing on the original 3D CT image to obtain the target image;

[0049] The ossicle automatic positioning module 42 is used to locate the ossicles in the target image, obtain the ossicle positioning points, and crop out the target area of ​​the ossicles based on the ossicle positioning points.

[0050] The ossicle automatic segmentation module 43 is used to segment the target region of the ossicles and obtain the three-dimensional segmentation result of the ossicles.

Claims

1. A method for automatic localization and segmentation of auditory ossicles in three-dimensional CT images, characterized in that: include: Step S1: preprocessing the original 3D CT image to obtain the target image; Step S2: locating the ossicles in the target image to obtain ossicle positioning points; and cropping the ossicle target area based on the ossicle positioning points; Step S3: Segment the auditory ossicle target area to obtain a three-dimensional segmentation result of the auditory ossicle.

2. The method for automatic ossicle localization and segmentation for three-dimensional CT images according to claim 1, characterized in that: The step S1: pre-processing the original three-dimensional CT image as a target image, specifically including down-sampling to a preset size and image normalization.

3. The method for automatic ossicle localization and segmentation for three-dimensional CT images according to claim 2, characterized in that: The step S2: locating the ossicles of the target image, obtaining ossicle positioning points, and cropping the ossicle target area based on the ossicle positioning points, specifically includes: Step S21: inputting the target image into the auditory ossicle positioning neural network to obtain auditory ossicle positioning points; Step S22: cropping an image of a fixed size from the original three-dimensional CT image with the auditory ossicle positioning point as the center as the auditory ossicle target area.

4. The method for automatic ossicle localization and segmentation for three-dimensional CT images according to claim 1, characterized in that: In step S3, an improved residual 3D-UNet network is constructed as an auditory ossicle segmentation neural network for segmenting the auditory ossicle target area.

5. An automatic ossicle localization and segmentation system for three-dimensional CT images, characterized in that: Includes the following modules: A target image acquisition module is used to perform data preprocessing on the original three-dimensional CT image to obtain the target image; an auditory ossicle automatic positioning module, configured to perform auditory ossicle positioning on the target image, obtain auditory ossicle positioning points, and crop an auditory ossicle target area based on the auditory ossicle positioning points; The auditory ossicle automatic segmentation module is used to segment the auditory ossicle target area to obtain a three-dimensional segmentation result of the auditory ossicle.

Images