Semi-automatic head shadow survey mark point detection method based on deep learning

Through deep learning technology, using deep convolutional neural networks and other machine learning models, semi-automated head shadow measurement mark point detection is realized, solving the problems of time-consuming, labor-intensive and subjective error in manual tracing, and improving detection accuracy and efficiency. Doctors provide more accurate information on bone and tooth structure.

CN119067949APending Publication Date: 2024-12-03SHANDONG UNIV
View PDF 1 Cites 0 Cited by

Patent Information

Application Number
CN202411174855.6
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-26
Publication Date
2024-12-03

AI Technical Summary

Technical Problem

Existing cephalogram measurement analysis relies on manual tracing, which is time-consuming and labor-intensive and subjective errors, making it difficult to achieve high-precision mark point detection.

Method used

The semi-automatic head shadow measurement mark point detection method based on deep learning is adopted, and the local block characteristics of the X-ray head shadow image are detected by deep convolutional neural network (such as Resnet50) combined with SVR, random forest or fully connected neural network. Reduce manual intervention and improve detection accuracy.

Benefits of technology

It significantly improves the accuracy and efficiency of landmark detection, can provide more precise information on teeth, bone structure and soft tissue, and reduces subjective errors.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119067949A_ABST
    Figure CN119067949A_ABST
Patent Text Reader

Abstract

The invention relates to the technical field of head shadow measurement analysis, and provides a semi-automatic head shadow measurement mark point detection method based on deep learning, and the method comprises the following steps: S1, obtaining a certain number of X-ray head shadow measurement images as a sample set; s2, cutting the image into small blocks, wherein each local image block only contains one mark point as input; s3, randomly selecting 200 local image blocks of the region of interest for each mark point, wherein the size of each local image block is 128 * 128 pixels; and S4, taking the extracted local image blocks of the 200 regions of interest as input images of the deep convolutional neural network for mark point detection, and independently training the network for each mark point. According to the invention, semi-automatic head shadow measurement can be well carried out.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of cephalometric analysis, and in particular to a semi-automatic cephalometric landmark point detection method based on deep learning. Background Art

[0002] X-ray cephalometric measurement mainly measures X-ray skull positioning films, draws certain line angles on the dental and maxillary, craniofacial landmarks for measurement and analysis, thereby understanding the structure of the dental and maxillary, craniofacial soft and hard tissues, and enabling the examination and diagnosis of the dental and maxillary, craniofacial area to go from surface morphology to the internal bone structure.

[0003] Currently, cephalometric analysis mainly relies on manual tracing, positioning and analysis by clinical physicians, which is not only time-consuming and labor-intensive, but also inevitably leads to subjective errors. Summary of the invention

[0004] The present invention provides a semi-automatic cephalometric landmark detection method based on deep learning, which solves the problem that manual annotation of cephalometric landmarks is time-consuming and labor-intensive and has subjective errors, and improves recognition accuracy.

[0005] The semi-automatic cephalometric landmark detection method based on deep learning according to the present invention comprises the following steps:

[0006] Step S1: obtaining a certain number of X-ray cephalometric images as a sample set;

[0007] Step S2: crop the image into small blocks, each local image block contains only one landmark point as input, and a local image block is a region of interest;

[0008] Step S3: randomly select 200 local image blocks of the region of interest for each landmark point, and the size of each local image block is 128 x 128 pixels;

[0009] Step S4: The extracted 200 local image blocks of the region of interest are used as input images of the deep convolutional neural network for landmark point detection, and the network is trained separately for each landmark point.

[0010] Preferably, in step S3, the step of extracting the local image block of the region of interest is:

[0011] Step 1: Find the actual position of the landmark point A: (x, y);

[0012] Step 2: Randomly select a 128x128 local image block in the area [x-128:x+128,y-128:y+128] and add the vertex positions of the selected local image block;

[0013] Step 3: Create a list V and check whether there are four vertices in the list V. If not, append the four vertices to V. Otherwise, repeat step 2.

[0014] Step 4: Repeat steps 1 to 3 until there are 200 sets of local image patch vertices in V;

[0015] Step 5: For each landmark point, use the vertices in V to crop 200 local image patches of the region of interest.

[0016] Preferably, in step S4, the deep convolutional neural network is Resnet50.

[0017] Preferably, in step S4, Resnet50 automatically extracts useful features from the input image, and based on these features, SVR, random forest or fully connected neural network is used to estimate the position of the landmark points.

[0018] Preferably, in step S4, a Titan X GPU is used to accelerate the training process, and Python and Keras are used to implement it; the learning rate is set to 0.001; the Adam optimizer is used; 10% of the training data is used for validation, and if the validate MSE value remains unchanged within 5 cycles, early-stop is set.

[0019] Compared with the existing methods, the present invention achieves better detection effect, can better perform semi-automatic cephalometric measurement, and can more accurately provide doctors with necessary information such as bone structure, soft tissue, and tooth structure. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 is a flow chart of a semi-automatic cephalometric landmark point detection method based on deep learning in an embodiment;

[0021] Figure 2 is a schematic diagram of a local image block of a region of interest in an embodiment;

[0022] Figure 3 Schematic diagram of the residual structure in Resnet50 in the embodiment. DETAILED DESCRIPTION

[0023] In order to further understand the content of the present invention, the present invention is described in detail in conjunction with the accompanying drawings and embodiments. It should be understood that the embodiments are only for explaining the present invention and are not intended to limit it.

[0024] Example

[0025] like Figure 1 As shown, this embodiment provides a semi-automatic cephalometric landmark detection method based on deep learning, which includes the following steps:

[0026] Step S1: Obtain a certain number of X-ray cephalometric images from the dataset as a sample set; the dataset used is the 2015 ISBI grand challenge data, which contains 150 images for training, 150 images as testset1, and 100 images as testset2.

[0027] Step S2: Crop the image into small blocks. Each local image block contains only one landmark point as input. A local image block is a region of interest. Figure 2 As shown;

[0028] Step S3: randomly select 200 local image blocks of the region of interest for each landmark point, and the size of each local image block is 128 x 128 pixels;

[0029] Steps to extract local image blocks of the region of interest:

[0030] Step 1: Find the actual position of the landmark point A: (x, y);

[0031] Step 2: Randomly select a 128x128 local image block in the area [x-128:x+128,y-128:y+128] and add the vertex positions of the selected local image block;

[0032] Step 3: Create a list V and check whether there are four vertices in the list V. If not, append the four vertices to V. Otherwise, repeat step 2.

[0033] Step 4: Repeat steps 1 to 3 until there are 200 sets of local image patch vertices in V;

[0034] Step 5: For each landmark point, use the vertices in V to crop 200 local image patches of the region of interest.

[0035] Step S4: The extracted 200 local image blocks of the region of interest are used as input images of the deep convolutional neural network for landmark point detection, and the network is trained separately for each landmark point.

[0036] In step S4, the deep convolutional neural network is Resnet50. The previous feature map is added to the later feature map. In this way, the feature information is passed to the later layers, which works well even with deeper architectures.

[0037] As the depth of a deep convolutional neural network increases, the neural network will learn different features. If there are too many convolutional layers, gradient disappearance is likely to occur. The residual structure in Resnet50 adds a short cut branch structure, also known as a highway. Figure 3 In the figure, the left side is a normal convolution layer, which is passed down layer by layer. A line is added on the right side, so that the entire network structure forms a residual structure. In this way, the output of the network is no longer the output of the simple convolution F(x), but the superposition of the convolution output and the previous input F(x)+X, where X and F(x) represent the previous and next feature maps respectively.

[0038] In step S4, Resnet50 automatically extracts useful features from the input image, and based on these features, SVR, random forest or fully connected neural network is used to estimate the location of the landmarks.

[0039] In step S4, Titan X GPU is used to accelerate the training process, which is implemented in Python and Keras. The learning rate is set to 0.001. The Adam optimizer is used. 10% of the training data is used for validation. If the validate MSE value remains unchanged within 5 cycles, early-stop is set.

[0040] This method can provide doctors with necessary information about bone structure, soft tissue, tooth structure, etc. more accurately.

[0041] The present invention and its embodiments are described schematically above, and the description is not restrictive. The drawings show only one embodiment of the present invention, and the actual structure is not limited thereto. Therefore, if a person skilled in the art is inspired by it and designs a structural method and an embodiment similar to the technical solution without creativity without departing from the purpose of the invention, they shall all fall within the protection scope of the present invention.

Claims

1. A semi-automatic cephalometric landmark detection method based on deep learning, characterized by: The following steps are involved: Step S1: obtaining a certain number of X-ray cephalometric images as a sample set; Step S2: crop the image into small blocks, each local image block contains only one landmark point as input, and a local image block is a region of interest; Step S3: randomly select 200 local image blocks of the region of interest for each landmark point, and the size of each local image block is 128 x 128 pixels; Step S4: The extracted 200 local image blocks of the region of interest are used as input images of the deep convolutional neural network for landmark point detection, and the network is trained separately for each landmark point.

2. The semi-automatic cephalometric landmark detection method based on deep learning according to claim 1, characterized in that: In step S3, the step of extracting the local image block of the region of interest is as follows: Step 1: Find the actual position of the landmark point A: (x, y); Step 2: Randomly select a 128x128 local image block in the area [x-128:x+128,y-128:y+128] and add the vertex positions of the selected local image block; Step 3: Create a list V and check whether there are four vertices in the list V. If not, append the four vertices to V. Otherwise, repeat step 2. Step 4: Repeat steps 1 to 3 until there are 200 sets of local image patch vertices in V; Step 5: For each landmark point, use the vertices in V to crop 200 local image patches of the region of interest.

3. The semi-automatic cephalometric landmark detection method based on deep learning according to claim 2, characterized in that: In step S4, the deep convolutional neural network is Resnet50.

4. The semi-automatic cephalometric landmark detection method based on deep learning according to claim 3, characterized in that: In step S4, Resnet50 automatically extracts useful features from the input image, and based on these features, SVR, random forest or fully connected neural network is used to estimate the location of the landmarks.

5. The method for semi-automatic cephalometric landmark detection based on deep learning according to claim 4, characterized in that: In step S4, TitanX GPU is used to accelerate the training process, which is implemented in python and Keras; the learning rate is set to 0.001; the Adam optimizer is used; 10% of the training data is used for validation, and early-stop is set if the validate MSE value remains unchanged within 5 cycles.

Citation Information

Patent Citations

  • X-ray head shadow survey mark point automatic positioning method based on reinforcement learning

    CN114947902A