Child skull fracture automatic detection method and device based on generated data augmentation
By combining generative data augmentation methods with 3D convolutional neural networks, the problems of missed and false detections in the detection of pediatric skull fractures have been solved, improving detection efficiency and accuracy.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- TSINGHUA UNIVERSITY
- Filing Date
- 2023-08-18
- Publication Date
- 2026-04-14
AI Technical Summary
Existing technologies are prone to missing or misdetecting skull fractures in children, especially small fractures and cranial sutures, which are difficult to distinguish. Furthermore, traditional methods are time-consuming and labor-intensive, and CT scans are inefficient.
Data augmentation was achieved by designing a data synthesis method for fracture lesions and cranial sutures. A generative model was constructed using generator and discriminator networks to expand the sample size. A binary classification network based on a 3D convolutional neural network was then trained for automatic detection.
It improved the detection rate of skull fractures in children, reduced the false positive rate, and enhanced detection efficiency and accuracy.
Smart Images
Figure CN117197052B_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of machine learning technology, and in particular to an automatic detection method and device for pediatric skull fractures based on generated data augmentation. Background Technology
[0002] Skull fractures are a common brain injury that can lead to many critical and life-threatening complications. Fragments of the skull can contusion the brain or damage blood vessels, so timely and accurate screening for skull fractures is crucial for the diagnosis and treatment of head injuries.
[0003] Computed tomography (CT) is the primary method for detecting skull fractures. Accurate diagnosis requires accumulated experience; conventional two-dimensional methods necessitate manually tracing multiple slices, while traditional three-dimensional visualization methods require rotating the image during diagnosis to avoid occlusion. Therefore, examining skull CT images is quite time-consuming and labor-intensive, and due to the skull's unique structure and location, CT can easily miss minor skull lesions.
[0004] Furthermore, small fracture lesions on the skull of children are easily missed. In addition, the cranial sutures of children have not yet closed, which increases the difficulty of distinguishing cranial sutures from fractures, and easily leads to missed or false diagnoses. Summary of the Invention
[0005] To address the aforementioned issues, an automatic detection method and device for pediatric skull fractures based on generated data augmentation is proposed. By designing a data synthesis method for fracture lesions and cranial sutures, data augmentation is performed to effectively distinguish between fractures and cranial sutures.
[0006] The first aspect of this application proposes an automatic detection method for pediatric skull fractures based on generated data augmentation, comprising:
[0007] A first set of pediatric brain CT images was acquired, and pixel annotation and preprocessing were performed on the first set of pediatric brain CT images to obtain fracture data, cranial suture data and normal data;
[0008] A candidate fracture lesion generation model and a candidate cranial suture generation model are constructed, and trained based on the fracture data, cranial suture data and normal data to obtain a trained fracture lesion generation model and cranial suture generation model.
[0009] Gaussian noise is randomly sampled multiple times and input into the fracture lesion generation model and cranial suture generation model to obtain fracture CT patch data and cranial suture CT patch data.
[0010] Construct a candidate fracture detection model, and train the candidate fracture detection model based on the fracture data, cranial suture data, normal data, generated fracture CT patch data and generated cranial suture CT patch data until the training is completed, and output the fracture detection model.
[0011] A second pediatric brain CT image was acquired, and the second pediatric brain CT image was detected according to the fracture detection model.
[0012] Optionally, the step of performing pixel annotation and preprocessing on the first set of pediatric brain CT images to obtain fracture data, cranial suture data, and normal data includes:
[0013] On the transverse images of the first set of pediatric brain CT images, the fracture site and cranial sutures were segmented and labeled pixel by pixel.
[0014] The first set of pediatric brain CT images, annotated with bone window width and window level processing, transforms the pixel value of each voxel to a grayscale image range of 0-255.
[0015] Optionally, both the candidate fracture lesion generation model and the candidate cranial suture generation model are composed of a generator network and a discriminator network, wherein,
[0016] The generator network is used to convert the input Gaussian noise into a 3D CT patch of a preset size;
[0017] The discriminator network is used to determine whether the input 3D CT patch is real data or generated data.
[0018] Optionally, the step of training the model based on the fracture data, cranial suture data, and normal data to obtain a trained fracture lesion generation model and cranial suture generation model includes:
[0019] The fracture data, cranial suture data, and normal data were randomly sampled multiple times to obtain CT patch positive samples containing fractures, CT patch negative samples not containing fractures, CT patch positive samples containing cranial sutures, and CT patch negative samples not containing cranial sutures of a preset size.
[0020] The candidate fracture lesion generation model is trained using the stochastic gradient descent method with positive CT patch samples containing fractures and negative CT patch samples without fractures to obtain a trained fracture lesion generation model.
[0021] Furthermore, the candidate cranial suture generation model is trained using the stochastic gradient descent method with positive CT patch samples containing cranial sutures and negative CT patch samples without cranial sutures to obtain a trained cranial suture generation model.
[0022] Optionally, the candidate fracture detection model is a binary classification network based on a 3D convolutional neural network.
[0023] Optionally, training the candidate fracture detection model based on the fracture data, cranial suture data, normal data, generated fracture CT patch data, and generated cranial suture CT patch data includes:
[0024] A first training set is constructed based on the CT patch positive samples containing fractures and the generated fracture CT patch data; a second training set is constructed based on the CT patch positive samples containing cranial sutures and the generated cranial suture CT patch data; and a third training set is constructed based on the normal data. The sampling ratio of the first training set, the second training set, and the third training set is 1:1:1.
[0025] The candidate fracture detection model is trained using the cross-entropy loss function as the optimization function and the stochastic gradient descent method with the first training set, the second training set, and the third training set.
[0026] Optionally, before training the candidate fracture detection model using the first training set, the second training set, and the third training set, the method further includes:
[0027] Sample 3D CTpatch of a preset size from the first training set, the second training set, and the third training set to a preset resolution.
[0028] Optionally, acquiring the second pediatric brain CT image and detecting the second pediatric brain CT image according to the fracture detection model includes:
[0029] For the second pediatric brain CT image to be interpreted, the skull regions with HU values greater than 200 were selected.
[0030] Starting from the upper left corner of the skull region, select a test patch of a preset size according to preset rules;
[0031] The test patch is input into the fracture detection model for detection, and the detection results are generated.
[0032] Optionally, the preset rules include:
[0033] Select the topmost and leftmost skull region pixels that are not covered by existing patches as the center of the new patch.
[0034] The second aspect of this application proposes an automatic detection device for pediatric skull fractures based on generated data augmentation, comprising:
[0035] The acquisition module is used to acquire a first set of pediatric brain CT images, and to perform pixel annotation and preprocessing on the first set of pediatric brain CT images to obtain fracture data, cranial suture data and normal data.
[0036] The first training module is used to construct a candidate fracture lesion generation model and a candidate cranial suture generation model, and to train them based on the fracture data, cranial suture data and normal data to obtain a trained fracture lesion generation model and cranial suture generation model.
[0037] The data generation module is used to randomly sample Gaussian noise multiple times, input the Gaussian noise into the fracture lesion generation model and the cranial suture generation model, and obtain the fracture CT patch data and the cranial suture CT patch data.
[0038] The second training module is used to construct a candidate fracture detection model. It trains the candidate fracture detection model based on the fracture data, cranial suture data, normal data, generated fracture CT patch data, and generated cranial suture CT patch data until the training is completed and the fracture detection model is output.
[0039] The detection module is used to acquire a second pediatric brain CT image and to detect the second pediatric brain CT image according to the fracture detection model.
[0040] The technical solutions provided by the embodiments of this application have at least the following beneficial effects:
[0041] By augmenting fracture and cranial suture data using a generative model, the sample size was expanded. Fractures and cranial sutures were distinguished based on a binary classification network, which improved the ability to distinguish between fractures and cranial sutures, increased the detection rate of skull fractures, and reduced the false positive rate.
[0042] Additional aspects and advantages of this application will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of this application. Attached Figure Description
[0043] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein:
[0044] Figure 1 This is a flowchart illustrating an automatic detection method for pediatric skull fractures based on generated data augmentation, according to an embodiment of this application.
[0045] Figure 2 This is a block diagram illustrating an automatic detection device for pediatric skull fractures based on generated data augmentation, according to an embodiment of this application. Detailed Implementation
[0046] The embodiments of this application are described in detail below. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this application, and should not be construed as limiting this application.
[0047] In existing technologies, convolutional neural networks are widely used in the field of medical image analysis for multiple tasks such as disease classification, lesion detection, and modality transformation.
[0048] Figure 1 This is a flowchart illustrating an automated method for detecting skull fractures in children based on augmented generated data, according to an embodiment of this application, comprising the following steps:
[0049] Step 101: Obtain the first set of pediatric brain CT images, and perform pixel annotation and preprocessing on the first set of pediatric brain CT images to obtain fracture data, cranial suture data and normal data.
[0050] In this embodiment, the first set of pediatric brain CT images is obtained from a local hospital or from a public dataset, such as the 1000 Functional Connectomes Project public dataset, the Brainmap dataset, etc. This application does not specifically limit the source of the first set of pediatric brain CT images.
[0051] In the first set of pediatric brain CT images, the fracture sites and cranial sutures were segmented and labeled pixel by pixel on the transverse images of the first set of pediatric brain CT images by manual and software methods. The CT images were preprocessed by adjusting the window width and window level of the bone window, and the pixel value of each voxel in the CT images was transformed to the grayscale image range of 0-255.
[0052] In this embodiment, the window width refers to the range of CT values displayed in the CT image. Within this CT value range, tissue structures are divided into 16 gray levels from white to black according to their density for observation and comparison. For example, if the window width is selected as 100 HU, then the CT value that the human eye can distinguish is 100 / 16 = 6.25 HU, meaning that two tissues with CT values differing by more than 6.25 HU can be recognized by the human eye. Therefore, the width of the window directly affects the image's clarity and contrast. If a narrow window width is used, the displayed CT value range is small, each gray level represents a small CT value amplitude, resulting in strong contrast, suitable for observing tissue structures with similar densities (such as brain tissue). Conversely, if a wide window width is used, the displayed CT value range is large, each gray level represents a large CT value amplitude, resulting in poor image contrast but uniform density, suitable for observing structures with large density differences (such as bone and soft tissue).
[0053] Similarly, the window level (window center) refers to the average or central value within the window width. For example, in a CT image with a window width of 100 HU and a window level of 0 HU, the area above the window level (0 HU) and below it (including +50 HU) will display all tissues within this 100 HU range and be recognizable by the human eye. Tissues larger than +50 HU will appear white; tissues smaller than -50 HU will appear black, and their density differences will not be visible. The human eye can only recognize CT values within ±50 HU, and the CT value range for each grayscale level is 100 / 16 = 6.25 HU. In principle, the window level should be equal to or close to the CT value to be observed; the window width should reflect the range of CT value changes for the tissue or lesion.
[0054] In this embodiment, the CT images are preprocessed using the width and level of a bone window. The bone window has a width of 800 HU to 2000 HU and a level of 250 HU to 500 HU.
[0055] Step 102: Construct candidate fracture lesion generation model and candidate cranial suture generation model, and train them based on fracture data, cranial suture data and normal data to obtain trained fracture lesion generation model and cranial suture generation model.
[0056] In this embodiment of the application, both the candidate fracture lesion generation model and the candidate cranial suture generation model are composed of a generator network and a discriminator network.
[0057] In existing technologies, generative adversarial networks (GANs) are a commonly used data generation method, consisting of a generator network and a discriminator network. The generator network is responsible for generating target domain data based on noisy input, while the discriminator network is responsible for determining whether the data is authentic.
[0058] The generator network in this application is used to convert the input Gaussian noise into a 3D CT patch of a preset size, and the discriminator network is used to determine whether the input 3D CT patch is real data or generated data.
[0059] In one possible embodiment, the generator network is used to transform the input Gaussian noise into a 3DCT patch of size 32*32*32.
[0060] Regarding the training process, in this embodiment of the application, fracture data, cranial suture data and normal data are randomly sampled multiple times to obtain CT patch positive samples containing fractures, CT patch negative samples not containing fractures, CT patch positive samples containing cranial sutures and CT patch negative samples not containing cranial sutures of a preset size.
[0061] Specifically, for each CT image in the training dataset, a voxel is randomly selected from all voxels labeled as fracture lesions each time as the center point of a 32*32*32 3D CT patch. This random sampling is repeated to obtain a large number of positive CT patch samples containing fractures. Each time, a 32*32*32 3D CT patch is randomly selected from all regions that do not contain fracture lesions. This random sampling is repeated to obtain a large number of negative CT patch samples that do not contain fractures. Similar to the process of generating the fracture lesion model, positive CT patch samples containing cranial sutures and negative CT patch samples that do not contain cranial sutures are randomly sampled.
[0062] Then, the stochastic gradient descent method was used to train the candidate fracture lesion generation model with CT patch positive samples containing fractures and CT patch negative samples without fractures, and the candidate cranial suture generation model with CT patch positive samples containing cranial sutures and CT patch negative samples without cranial sutures, so as to obtain the trained fracture lesion generation model and cranial suture generation model.
[0063] It is understandable that model training is an iterative process. The model is trained by continuously adjusting the network parameters until the overall loss function value of the model is less than the preset value, or the overall loss function value of the model no longer changes or changes slowly, and the model converges, resulting in a well-trained model.
[0064] Alternatively, training can be considered complete once a preset number of training sessions has been completed.
[0065] Alternatively, training can be considered complete once a preset training time has been reached.
[0066] In existing technologies, stochastic gradient descent uses random sampling instead of complete samples. Its main purpose is to improve iteration speed and avoid getting bogged down in computationally intensive tasks. The name "stochastic gradient descent" itself reflects its core idea: randomly selecting a point for gradient descent, rather than iterating through all samples before parameter iteration. This is because calculating the cost function of gradient descent requires traversing all samples in each iteration until a local optimum is reached. With a large sample size, this results in slow convergence and extremely high computational cost. Stochastic gradient descent, by finding the minimum value only at the current sample point, typically cannot reach a true local optimum, but it can get very close. It represents a compromise between large sample sizes and computational cost.
[0067] Step 103: Randomly sample Gaussian noise multiple times, and input the Gaussian noise into the fracture lesion generation model and the cranial suture generation model to obtain the generated fracture CT patch data and the generated cranial suture CT patch data.
[0068] In this embodiment, a trained fracture lesion generation model and cranial suture generation model are used to generate new samples.
[0069] Step 104: Construct a candidate fracture detection model. Train the candidate fracture detection model based on fracture data, cranial suture data, normal data, generated fracture CT patch data, and generated cranial suture CT patch data until training is complete, and output the fracture detection model.
[0070] In this embodiment, the candidate fracture detection model is a binary classification network based on a 3D convolutional neural network, which is responsible for determining whether the input CT patch contains a skull fracture lesion. Each 3D CT patch of size 32*32*32 to be judged is first upsampled to a resolution of 128*128*128, and then input into the candidate fracture detection model. The output of the model is a probability value in the range of 0-1, representing the probability that the input patch contains a skull fracture lesion.
[0071] Regarding the training process, in this embodiment of the application, a first training set is constructed based on CT patch positive samples containing fractures and generated fracture CT patch data, a second training set is constructed based on CT patch positive samples containing cranial sutures and generated cranial suture CT patch data, and a third training set is constructed based on normal data. The sampling ratio of the first training set, the second training set and the third training set is 1:1:1.
[0072] It should be noted that the sampling ratio of 1:1:1 in this application is to improve the ability of the fracture detection model to distinguish between fractures and cranial sutures and to reduce the false positive rate of prediction. The process of obtaining the fracture detection model by training with other sampling ratios is still within the scope of protection of this application.
[0073] For example, similar to the training process type in step 102, this application uses the cross-entropy loss function as the optimization function and employs the stochastic gradient descent method to train and generate a fracture detection model.
[0074] Step 105: Obtain the second child's brain CT image and perform detection on the second child's brain CT image according to the fracture detection model.
[0075] In this embodiment of the application, a fracture detection model is used to detect newly input brain CT images.
[0076] Specifically, for the second pediatric brain CT image to be interpreted, the skull region with a HU value greater than 200 is selected. Starting from the upper left corner of the skull region, a test patch of a preset size is selected according to preset rules. The test patch is then input into the fracture detection model for detection, and the detection results are generated.
[0077] In this embodiment, the preset rule refers to selecting the topmost and leftmost skull region pixels that are not covered by existing patches as the center of the new patch. In this way, all selected patches can just cover all skull regions, and the position in each skull region will definitely be at the non-edge position of a certain patch.
[0078] In one possible embodiment, if the probability of a test patch being detected is greater than 0.5, then the test patch is considered to potentially contain a fracture lesion.
[0079] This application's embodiments augment fracture and cranial suture data using a generative model, thereby expanding the sample size. Based on a binary classification network, fractures and cranial sutures are distinguished, improving the ability to differentiate between fractures and cranial sutures, increasing the detection rate of skull fractures, and reducing the false positive rate.
[0080] Figure 2 This is a block diagram illustrating an automatic pediatric skull fracture detection device 200 based on generated data augmentation, according to an embodiment of this application, comprising:
[0081] The acquisition module 210 is used to acquire the first set of pediatric brain CT images, and to perform pixel annotation and preprocessing on the first set of pediatric brain CT images to obtain fracture data, cranial suture data and normal data.
[0082] The first training module 220 is used to construct a candidate fracture lesion generation model and a candidate cranial suture generation model, and to train them based on fracture data, cranial suture data and normal data to obtain a trained fracture lesion generation model and cranial suture generation model.
[0083] The data generation module 230 is used to randomly sample Gaussian noise multiple times, input the Gaussian noise into the fracture lesion generation model and the cranial suture generation model, and obtain the fracture CT patch data and the cranial suture CT patch data.
[0084] The second training module 240 is used to construct a candidate fracture detection model. It trains the candidate fracture detection model based on fracture data, cranial suture data, normal data, generated fracture CT patch data, and generated cranial suture CT patch data until the training is completed and the fracture detection model is output.
[0085] The detection module 250 is used to acquire the second child's brain CT image and to detect the second child's brain CT image according to the fracture detection model.
[0086] Regarding the apparatus in the above embodiments, the specific manner in which each module performs its operation has been described in detail in the embodiments related to the method, and will not be elaborated upon here.
[0087] It should be understood that the various forms of processes shown above can be used to rearrange, add, or delete steps. For example, the steps described in this disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution of this disclosure can be achieved, and this is not limited herein.
[0088] The specific embodiments described above do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this disclosure should be included within the scope of protection of this disclosure.
Claims
1. A method for automatic detection of skull fracture in children based on generative data augmentation, characterized in that, include: A first set of pediatric brain CT images was acquired, and pixel annotation and preprocessing were performed on the first set of pediatric brain CT images to obtain fracture data, cranial suture data and normal data; Constructing candidate fracture lesion generation models and candidate cranial suture generation models, and training them based on fracture data, cranial suture data, and normal data to obtain trained fracture lesion generation models and cranial suture generation models, specifically includes: performing multiple random sampling operations on the fracture data, cranial suture data, and normal data to obtain CT patch positive samples containing fractures, CT patch negative samples not containing fractures, CT patch positive samples containing cranial sutures, and CT patch negative samples not containing cranial sutures of a preset size; using the stochastic gradient descent method, training the candidate fracture lesion generation models with the CT patch positive samples containing fractures and the CT patch negative samples not containing fractures to obtain trained fracture lesion generation models; and using the stochastic gradient descent method, training the candidate cranial suture generation models with the CT patch positive samples containing cranial sutures and the CT patch negative samples not containing cranial sutures to obtain trained cranial suture generation models. Gaussian noise is randomly sampled multiple times and input into the fracture lesion generation model and cranial suture generation model to obtain fracture CT patch data and cranial suture CT patch data. Construct a candidate fracture detection model, and train the candidate fracture detection model based on the fracture data, cranial suture data, normal data, generated fracture CT patch data and generated cranial suture CT patch data until the training is completed, and output the fracture detection model; Acquire a second pediatric brain CT image, and perform detection on the second pediatric brain CT image according to the fracture detection model, including: For the second pediatric brain CT image to be interpreted, screen out skull regions with HU values greater than 200; starting from the upper left corner of the skull region, select a test patch of a preset size according to preset rules; input the test patch into the fracture detection model for detection, and generate detection results; the preset rules include: selecting the topmost and leftmost skull region pixels not covered by existing patches as the center of the new patch.
2. The method of claim 1, wherein, The pixel annotation and preprocessing of the first set of pediatric brain CT images yields fracture data, cranial suture data, and normal data, including: On the transverse images of the first set of pediatric brain CT images, the fracture site and cranial sutures were segmented and labeled pixel by pixel. The first set of pediatric brain CT images, annotated with bone window width and window level processing, transforms the pixel value of each voxel to a grayscale image range of 0-255.
3. The method according to claim 1, characterized in that, Both the candidate fracture lesion generation model and the candidate cranial suture generation model are composed of a generator network and a discriminator network, wherein... The generator network is used to convert the input Gaussian noise into a 3D CT patch of a preset size; The discriminator network is used to determine whether the input 3D CT patch is real data or generated data.
4. The method according to claim 1, characterized in that, The candidate fracture detection model is a binary classification network based on a 3D convolutional neural network.
5. The method according to claim 1, characterized in that, The process of training the candidate fracture detection model based on the fracture data, cranial suture data, normal data, generated fracture CT patch data, and generated cranial suture CT patch data includes: A first training set is constructed based on the CT patch positive samples containing fractures and the generated fracture CT patch data; a second training set is constructed based on the CT patch positive samples containing cranial sutures and the generated cranial suture CT patch data; and a third training set is constructed based on the normal data. The sampling ratio of the first training set, the second training set, and the third training set is 1:1:
1. The candidate fracture detection model is trained using the cross-entropy loss function as the optimization function and the stochastic gradient descent method with the first training set, the second training set, and the third training set.
6. The method according to claim 5, characterized in that, Before training the candidate fracture detection model using the first training set, the second training set, and the third training set, the method further includes: Sample 3D CT patches of a preset size from the first training set, the second training set, and the third training set to a preset resolution.
7. An automatic detection device for pediatric skull fractures based on generated data augmentation, characterized in that, include: The acquisition module is used to acquire a first set of pediatric brain CT images, and to perform pixel annotation and preprocessing on the first set of pediatric brain CT images to obtain fracture data, cranial suture data and normal data. The first training module is used to construct a candidate fracture lesion generation model and a candidate cranial suture generation model, and train them based on the fracture data, cranial suture data, and normal data to obtain trained fracture lesion generation models and cranial suture generation models. Specifically, it includes: performing multiple random sampling operations on the fracture data, cranial suture data, and normal data to obtain a preset size of CT patch positive samples containing fractures, CT patch negative samples not containing fractures, CT patch positive samples containing cranial sutures, and CT patch negative samples not containing cranial sutures; using the stochastic gradient descent method, training the candidate fracture lesion generation model with the CT patch positive samples containing fractures and the CT patch negative samples not containing fractures to obtain a trained fracture lesion generation model; and using the stochastic gradient descent method, training the candidate cranial suture generation model with the CT patch positive samples containing cranial sutures and the CT patch negative samples not containing cranial sutures to obtain a trained cranial suture generation model. The data generation module is used to randomly sample Gaussian noise multiple times, input the Gaussian noise into the fracture lesion generation model and the cranial suture generation model, and obtain the fracture CT patch data and the cranial suture CT patch data. The second training module is used to construct a candidate fracture detection model. It trains the candidate fracture detection model based on the fracture data, cranial suture data, normal data, generated fracture CT patch data, and generated cranial suture CT patch data until the training is completed and the fracture detection model is output. The detection module is used to acquire a second pediatric brain CT image and perform detection on the second pediatric brain CT image according to the fracture detection model, including: for the second pediatric brain CT image to be interpreted, screening out skull regions with HU values greater than 200; starting from the upper left corner of the skull region, selecting a test patch of a preset size according to preset rules; inputting the test patch into the fracture detection model for detection and generating detection results; the preset rules include: selecting the topmost and leftmost skull region pixels not covered by existing patches as the center of the new patch.
Citation Information
Patent Citations
Method for medical diagnosis by using neural network
KR102179090B1