Unsupervised fetal cerebral hemorrhage detection and focus segmentation method and device and electronic equipment

Abstract

The invention relates to the technical field of medical image processing, in particular to an unsupervised fetal cerebral hemorrhage detection and focus segmentation method and device and electronic equipment, and the method comprises the steps: obtaining the magnetic resonance imaging data of a normal fetal brain; the method comprises the following steps: synthesizing a plurality of pseudo fetal focus and intraventricular hemorrhage images from magnetic resonance imaging data of a normal fetal brain by using medical priori knowledge, and generating training data according to the plurality of pseudo fetal focus and intraventricular hemorrhage images; training a segmentation model by using the training data, inputting the magnetic resonance imaging data of the brain of the target fetus into the trained segmentation model, and outputting a segmentation result of the focus of the brain of the target fetus and the intraventricular hemorrhage by the segmentation model. Therefore, the problems that in the prior art, training data is scarce, and a large amount of manual annotation data needs to be relied on for model training are solved.

Description

Technical Field

[0001] The present application relates to the field of medical image processing technology, and in particular to a method, device and electronic equipment for unsupervised fetal cerebral hemorrhage detection and lesion segmentation. Background Art

[0002] GMH-IVH (germinal matrix and intraventricular hemorrhage) is the most common type of fetal cerebral hemorrhage and has important clinical significance.

[0003] Traditional prenatal diagnosis relies primarily on transabdominal ultrasound, but due to its limited ability to identify subtle lesions and low signal-to-noise ratio, it is often prone to missed diagnoses. Therefore, deep learning artificial intelligence methods can be used for automatic segmentation and auxiliary diagnosis of GMH-IVH.

[0004] Related technologies primarily employ deep learning segmentation methods such as U-Net (U-shaped Convolutional Neural Network) and VM-Unet (Variational Mode–Unet). However, these segmentation methods rely heavily on labeled lesion imaging data as a training basis, requiring a large amount of labeled data for the model to learn lesion characteristics for accurate segmentation and diagnosis. However, the incidence of fetal GMH-IVH is extremely low, and data collection is limited by the complexity and ethical requirements of prenatal testing, making it difficult to obtain sufficient high-quality labeled data. This data scarcity leads to insufficient model training, limiting the model's generalization and application. Summary of the Invention

[0005] The present application provides a method, device and electronic device for unsupervised fetal cerebral hemorrhage detection and lesion segmentation to solve the problems in related technologies such as scarcity of training data and the need to rely on a large amount of manually labeled data for model training.

[0006] The first aspect of the present application provides an unsupervised fetal brain hemorrhage detection and lesion segmentation method, comprising the following steps: acquiring magnetic resonance imaging data of a normal fetal brain; synthesizing multiple pseudo-fetal lesions and images of intraventricular hemorrhage from the magnetic resonance imaging data of the normal fetal brain using medical prior knowledge, and generating training data based on the multiple pseudo-fetal lesions and images of intraventricular hemorrhage; training a segmentation model using the training data, inputting the magnetic resonance imaging data of the target fetal brain into the trained segmentation model, and the segmentation model outputting segmentation results of the target fetal brain lesions and intraventricular hemorrhage.

[0007] Optionally, medical prior knowledge is used to synthesize multiple pseudo fetal lesion and intraventricular hemorrhage images from magnetic resonance imaging data of a normal fetal brain, including: obtaining tissue segmentation results from magnetic resonance imaging data of a normal fetal brain; using medical prior knowledge to probabilistically select the bleeding location in the tissue segmentation result during each synthesis, generate a noise image of the same dimension as the input image, and generate a random shape mask based on the noise image; combining the random shape mask with the bleeding area to obtain a bleeding area mask, and masking the bleeding area during each synthesis to generate pseudo fetal lesion and intraventricular hemorrhage images.

[0008] Optionally, tissue segmentation results are obtained from magnetic resonance imaging data of a normal fetal brain, including: extracting slices from the magnetic resonance imaging data of the normal fetal brain along the axial, sagittal and coronal planes of the brain volume, and generating tissue segmentation results based on the slices of the brain volume along the axial, sagittal and coronal planes.

[0009] Optionally, generating a random shape mask based on the noise image includes: performing Gaussian blur processing on the noise image; scaling the Gaussian blurred image to obtain an image with different light and dark areas; performing threshold operations, opening operations, and closing operations on the image with different light and dark areas to obtain a mask image, and obtaining the random shape mask by element-by-element multiplication of the mask image and the brain mask.

[0010] Optionally, the bleeding area during each synthesis is masked into an image of a pseudo fetal lesion and intraventricular hemorrhage, including: scaling the pixel values within the mask area of the bleeding area during each synthesis; and performing Gaussian blurring to soften the boundary of the mask area of the bleeding area to obtain an image of the pseudo fetal lesion and intraventricular hemorrhage.

[0011] Optionally, the segmentation model includes an image segmentation layer, stacked downsampling and encoding blocks, an upsampling path consisting of an upsampling layer and a convolution block, and a downsampling path consisting of a convolution block and a maximum pooling layer, and the jump connection of the UNet variant structure adopts an addition operation.

[0012] Optionally, in the process of training the segmentation model using training data, it also includes: optimizing the training process of the segmentation model using a segmentation optimization model, wherein the segmentation optimization model includes an adjustment module, a convolutional neural network branch, a cross-branch attention mechanism and an optimizer, the adjustment module adjusts the image encoder input size according to the medical image resolution, and when the convolutional neural network branch and the cross-branch attention mechanism are trained, they receive images and point prompts randomly selected based on the segmentation mask as input, and the optimizer optimizes the training setting parameters of the segmentation model.

[0013] The second aspect of the present application provides an unsupervised fetal brain hemorrhage detection and lesion segmentation device, including: an acquisition module for acquiring magnetic resonance imaging data of a normal fetal brain; a synthesis module for synthesizing multiple pseudo fetal lesions and intraventricular hemorrhage images from the magnetic resonance imaging data of the normal fetal brain using medical prior knowledge, and generating training data based on the multiple pseudo fetal lesions and intraventricular hemorrhage images; a training module for training a segmentation model using the training data, inputting the magnetic resonance imaging data of the target fetal brain into the trained segmentation model, and the segmentation model outputting the segmentation results of the target fetal brain lesions and intraventricular hemorrhage.

[0014] Optionally, the synthesis module is further used to: obtain tissue segmentation results from magnetic resonance imaging data of a normal fetal brain; use medical prior knowledge to probabilistically select the bleeding location in the tissue segmentation result during each synthesis, generate a noise image of the same dimension as the input image, and generate a random shape mask based on the noise image; combine the random shape mask with the bleeding area to obtain a bleeding area mask, and generate an image of a pseudo fetal lesion and intraventricular hemorrhage based on the bleeding area mask during each synthesis.

[0015] Optionally, the synthesis module is further used to extract slices along the axial, sagittal and coronal planes of the brain volume from the magnetic resonance imaging data of the normal fetal brain, and generate tissue segmentation results based on the slices along the axial, sagittal and coronal planes of the brain volume.

[0016] Optionally, the synthesis module is further used to: perform Gaussian blur processing on the noise image; scale the Gaussian blurred image to obtain an image with different light and dark areas; perform threshold operations, opening operations and closing operations on the image with different light and dark areas to obtain a mask image, and obtain a random shape mask by element-by-element multiplication of the mask image and the brain mask.

[0017] Optionally, the synthesis module is further used to: scale the pixel values within the hemorrhage area mask region during each synthesis; perform Gaussian blurring to soften the boundary of the hemorrhage area mask region to obtain images of pseudo fetal lesions and intraventricular hemorrhage.

[0018] Optionally, the segmentation model includes an image segmentation layer, stacked downsampling and encoding blocks, an upsampling path consisting of an upsampling layer and a convolution block, and a downsampling path consisting of a convolution block and a maximum pooling layer, and the jump connection of the UNet variant structure adopts an addition operation.

[0019] Optionally, the training module is further used to: optimize the training process of the segmentation model using a segmentation optimization model, wherein the segmentation optimization model includes an adjustment module, a convolutional neural network branch, a cross-branch attention mechanism and an optimizer, the adjustment module adjusts the image encoder input size according to the medical image resolution, and when the convolutional neural network branch and the cross-branch attention mechanism are trained, they receive images and point prompts randomly selected based on the segmentation mask as input, and the optimizer optimizes the training setting parameters of the segmentation model.

[0020] The third aspect of the present application provides an electronic device, including: a memory, a processor, and a computer program stored in the memory and runnable on the processor. The processor executes the program to implement the unsupervised fetal cerebral hemorrhage detection and lesion segmentation method as described in the above embodiment.

[0021] The fourth aspect of the present application provides a computer-readable storage medium having a computer program or instructions stored thereon, which is executed by a processor to implement the unsupervised fetal cerebral hemorrhage detection and lesion segmentation method as described in the above embodiment.

[0022] Therefore, this application has the following beneficial effects:

[0023] The embodiment of the present application can use medical prior knowledge to synthesize diverse, high-quality images of pseudo fetal lesions and intraventricular hemorrhage from magnetic resonance imaging data of normal fetal brains, and generate training data based on multiple pseudo fetal lesions and intraventricular hemorrhage images. The training data is used to train a segmentation model, eliminating the need for real case labeled data for segmentation model training, reducing manual participation and data acquisition thresholds, thereby improving the feasibility of model development and application. The trained segmentation model can achieve segmentation of target fetal brain lesions and intraventricular hemorrhage with higher accuracy than manual clinical subjective judgment, providing more reliable technical support for subsequent prenatal diagnosis and risk assessment. This solves the technical problems in related technologies such as the scarcity of training data and the need to rely on a large amount of manually labeled data for model training.

[0024] Additional aspects and advantages of the present application will be given in part in the description below, and in part will become apparent from the description below, or will be learned through practice of the present application. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] The above and / or additional aspects and advantages of the present application will become apparent and easily understood from the following description of the embodiments in conjunction with the accompanying drawings, in which:

[0026] Figure 1 A flowchart of an unsupervised fetal cerebral hemorrhage detection and lesion segmentation method according to an embodiment of the present application;

[0027] Figure 2Schematic diagram of the FreeHemoSeg framework provided according to an embodiment of the present application;

[0028] Figure 3 A performance diagram of a segmentation model provided according to an embodiment of the present application;

[0029] Figure 4 A schematic diagram of a slice-level segmentation result of cerebral hemorrhage provided according to an embodiment of the present application;

[0030] Figure 5 A schematic diagram of an automated process for three-dimensional reconstruction and tissue segmentation of fetal brain MRI (Magnetic Resonance Imaging) according to an embodiment of the present application;

[0031] Figure 6 A comparison chart of the outputs of the segmentation model provided according to an embodiment of the present application and other anomaly detection models;

[0032] Figure 7 A schematic diagram showing a visual comparison of slice-level false positive segmentation results provided according to an embodiment of the present application;

[0033] Figure 8 A schematic diagram of a visualization result of enhancing the slice-level false-negative segmentation effect provided according to an embodiment of the present application;

[0034] Figure 9 This is an example diagram of an unsupervised fetal cerebral hemorrhage detection and lesion segmentation device provided according to an embodiment of the present application;

[0035] Figure 10 Schematic diagram of the structure of an electronic device according to an embodiment of the present application. DETAILED DESCRIPTION

[0036] The following describes in detail embodiments of the present application. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to be used to explain the present application, and should not be construed as limiting the present application.

[0037] Before describing the solution of this application, we first introduce the relevant technologies and existing technical problems to assist in understanding the solution of this application.

[0038] GMH-IVH, the most common type of fetal intracerebral hemorrhage, has important clinical implications. The germinal matrix is a transitional area surrounding the lateral ventricles, composed of fragile, thin-walled blood vessels and migrating neurons. Because premature infants have poor cerebral autoregulation, perinatal stress often leads to germinal matrix hemorrhage that spreads into the ventricles, resulting in GMH-IVH. Although the overall incidence of fetal GMH-IVH is low (approximately 0.5–0.9 per 1000 pregnancies), it is a significant cause of neonatal mortality and disability.

[0039] Traditional prenatal diagnosis mainly relies on transabdominal ultrasound, but due to its limited ability to identify subtle lesions and low image signal-to-noise ratio, it is often easy to miss the diagnosis. In recent years, MRI has gradually replaced ultrasound as the first choice for identifying fetal brain structures and lesions due to its excellent resolution of soft tissue. Studies have shown that MRI can detect more subtle lesions and improve diagnostic accuracy. However, even with MRI, the detection and localization of GMH-IVH lesions currently still mainly rely on manual reading by radiologists, which has shortcomings such as strong subjectivity, easy to miss small lesions, and long reading time. In T2-weighted imaging, both hemorrhage and germinal matrix signals appear as low signals, and the two are easily confused. Artifacts caused by fetal movement also affect identification.

[0040] Deep learning artificial intelligence methods have been widely used for the automated segmentation and auxiliary diagnosis of various brain lesions, such as brain tumors and intracerebral hemorrhage. Most of these methods use supervised learning, requiring large amounts of carefully labeled lesion data to train the models. However, GMH-IVH imaging data itself is extremely scarce, severely limiting the generalization ability and practical application value of deep learning models.

[0041] To overcome the challenge of scarce labeled data, unlabeled deep learning methods have been proposed, such as unsupervised anomaly detection, including knowledge distillation and reconstruction-based anomaly detection methods. However, these methods have limited success in achieving high-precision detection and segmentation. Another rapidly developing approach is "pseudo-anomaly synthesis," which leverages prior medical knowledge and existing normal images to artificially synthesize new images resembling the target lesion as training data. However, there is currently no precedent for the use of unsupervised, unlabeled methods for fetal GMH-IVH.

[0042] The related technologies have the following technical problems:

[0043] 1. Reliance on extensive manual annotation and difficulty in generalization: Current mainstream deep learning segmentation methods, such as U-Net and VM-Unet, require a large amount of labeled lesion image data as a training basis. However, due to the extreme rarity of fetal GMH-IVH and the difficulty in data collection, it is difficult to obtain sufficient high-quality labeled data, which limits the generalization and application of the model.

[0044] 2. Limited accuracy of unsupervised anomaly detection methods: Although there are unsupervised anomaly detection methods such as knowledge distillation and reconstruction, studies have shown that these algorithms are generally less accurate than traditional labeled methods for segmenting fine-grained lesion areas and are prone to missed detection or over-detection.

[0045] 3. Existing "pseudo-abnormality synthesis" methods are not specifically targeted at GMH-IVH: Various medical fields have begun to explore pseudo-abnormality synthesis, which can simulate lesion training models without relying on labels. However, published literature reports mostly focus on tumors (such as liver tumors, lung lesions, and cardiovascular and cerebrovascular events). There is no specific synthesis strategy and training framework for fetal GMH-IVH. Moreover, these methods often fail to accurately simulate the unique bleeding morphology, anatomical background, and signal characteristics of the fetus, and their application in this scenario is limited.

[0046] 4. The clinical application process is complex and relies on manual interpretation: Although MRI has high sensitivity, the clinical diagnosis of fetal brain hemorrhage and germinal matrix relies on expert interpretation, which is highly subjective and time-consuming, and is easily affected by fetal movement and adjacent tissue interference, resulting in misjudgment or missed diagnosis.

[0047] 5. The generalization ability of the segmentation model trained with labels is weak: The segmentation model with labels (such as VM-Unet) is significantly inferior to the model trained based on large-scale and diverse pseudo-synthetic lesion data in terms of lesion boundary and detail segmentation. In addition, the traditional model is prone to misclassifying normal tissue as lesions or missing tiny lesions.

[0048] To this end, this application provides an unsupervised fetal intracerebral hemorrhage detection and lesion segmentation method. This method does not rely on real lesion annotation data. By performing medically driven synthetic lesion simulation on normal fetal brain MRI images, it generates diverse, high-quality pseudo-abnormal data. Combined with an advanced deep learning segmentation framework, it achieves efficient and accurate automatic detection and segmentation of fetal GMH-IVH lesion areas. The aim is to improve the model's generalization ability and segmentation accuracy, reduce manual participation and data acquisition barriers, and provide more intelligent and reliable technical support for prenatal diagnosis and risk assessment.

[0049] Specifically, Figure 1 A flowchart of an unsupervised fetal cerebral hemorrhage detection and lesion segmentation method provided in an embodiment of the present application.

[0050] like Figure 1 As shown in FIG, the unsupervised fetal cerebral hemorrhage detection and lesion segmentation method includes the following steps:

[0051] In step S101 , magnetic resonance imaging data of a normal fetal brain is acquired.

[0052] In step S102, a plurality of pseudo fetal lesions and intraventricular hemorrhage images are synthesized from the magnetic resonance imaging data of the normal fetal brain using medical prior knowledge, and training data is generated based on the plurality of pseudo fetal lesions and intraventricular hemorrhage images.

[0053] It can be understood that the embodiments of the present application do not need to rely on real lesion annotation data. They use medical prior knowledge to synthesize multiple pseudo fetal lesions and intraventricular hemorrhage images from the magnetic resonance imaging data of the normal fetal brain to generate diverse and high-quality pseudo abnormality data, and generate training data based on the multiple pseudo fetal lesions and intraventricular hemorrhage images.

[0054] In an embodiment of the present application, medical prior knowledge is used to synthesize multiple pseudo fetal lesions and intraventricular hemorrhage images from magnetic resonance imaging data of a normal fetal brain, including: obtaining a tissue segmentation result from the magnetic resonance imaging data of a normal fetal brain; using medical prior knowledge to probabilistically select the bleeding location in the tissue segmentation result at each synthesis, generate a noise image of the same dimension as the input image, and generate a random shape mask based on the noise image; combining the random shape mask with the bleeding area to obtain a bleeding area mask, and generating an image of the pseudo fetal lesion and intraventricular hemorrhage based on the bleeding area mask at each synthesis.

[0055] It can be understood that the process of synthesizing the graphics of pseudo fetal lesions and intraventricular hemorrhage in the embodiment of the present application includes: obtaining tissue segmentation results from magnetic resonance imaging data of a normal fetal brain, and when synthesizing using medical prior knowledge, selecting the bleeding location in the tissue segmentation result by probability, and generating a noise image with the same circumference as the input image, generating a random shape mask based on the noise image, combining the random shape mask with the hemorrhage area to obtain a hemorrhage area mask, and then masking the hemorrhage area at each synthesis to form an image of pseudo fetal lesions and intraventricular hemorrhage.

[0056] Among them, when using medical prior knowledge to select the bleeding location, the following probability distribution can be used: 30% in the ventricles and deep gray matter areas (Grade I-III), 30% in the simple ventricles, 30% in the simple deep gray matter areas, and 10% in the white matter areas (Grade IV).

[0057] In an embodiment of the present application, tissue segmentation results are obtained from magnetic resonance imaging data of a normal fetal brain, including: extracting slices from the magnetic resonance imaging data of the normal fetal brain along the axial, sagittal, and coronal planes of the brain volume, and generating tissue segmentation results based on the slices of the brain volume along the axial, sagittal, and coronal planes.

[0058] It can be understood that the embodiments of the present application can extract slices from the magnetic resonance imaging data of the normal fetal brain along the axial, sagittal and coronal planes of the brain volume, generate tissue segmentation results based on the slices of the brain volume along the axial, sagittal and coronal planes, and ensure data diversity by extracting slices from multiple anatomical planes.

[0059] In an embodiment of the present application, a random shape mask is generated based on a noise image, including: performing Gaussian blur processing on the noise image; scaling the image after Gaussian blur processing to obtain an image with different light and dark areas; performing threshold operations, opening operations and closing operations on the image with different light and dark areas to obtain a mask image, and obtaining a random shape mask by element-by-element multiplication of the mask image and the brain mask.

[0060] It is understandable that the embodiment of the present application performs Gaussian blur processing on the noise image, and then applies threshold operation and morphological opening and closing operations to smooth the shape and fill small holes, thereby ensuring that the generated random shape is confined to the brain area. Specifically:

[0061] Given an input image I∈R H×W×C and threshold t∈R, the goal is to output a random shape mask M∈0,1 H×W First, generate random noise N∈R with the same dimension as I H×W , where the pixel values range from 0 to 255. The noisy image is then Gaussian blurred to obtain B∈R H×W , using a Gaussian kernel size of 15 in the X and Y directions. To maintain a reasonable pixel value range after blurring, B is rescaled to the range 0 to 255, resulting in S∈R H×W At this point, an image with different light and dark areas is obtained. Next, a threshold operation is applied to the image to obtain T, and then an opening and closing operation is performed on T to smooth the shape and fill the holes to obtain the optimized mask image R∈0,1 H×W The final output mask image M is obtained by combining R with the brain mask M b ∈0,1 H×W The result is element-wise multiplication, ensuring that the generated random shape is located within the overall brain area.

[0062] In an embodiment of the present application, the bleeding area is masked into an image of a pseudo fetal lesion and intraventricular hemorrhage according to each synthesis, including: scaling the pixel values within the bleeding area mask area at each synthesis; performing Gaussian blurring to soften the boundary of the bleeding area mask area to obtain an image of a pseudo fetal lesion and intraventricular hemorrhage.

[0063] It can be understood that the embodiment of the present application can scale the pixel values within the bleeding area mask area during each synthesis, and then perform Gaussian blurring to soften the boundaries of the bleeding area mask area to obtain images of pseudo fetal lesions and intraventricular hemorrhage, thereby efficiently generating a large amount of pseudo abnormality data for deep learning training without the need for real bleeding case data.

[0064] For the composite area of the pseudo fetal lesion and the intraventricular hemorrhage image, the pixel value in the hemorrhage mask area can be set to 0.3-0.5 times the original value to simulate the low signal feature of hemorrhage in T2-weighted MRI (Magnetic Resonance Imaging).

[0065] Specifically, this application uses medical prior knowledge to guide the generation of pseudo-fetal lesions and intraventricular hemorrhage images, and achieves high-quality unlabeled deep learning by embedding medical prior knowledge, including the following steps:

[0066] 1. First, slices were extracted from the reconstructed brain volume along the axial, sagittal, and coronal planes. Based on the tissue segmentation results, the slices containing the ventricles, deep gray matter, and surrounding white matter were retained as regions of interest, while the white matter regions were considered as potential surrogate regions (corresponding to Grade IV hemorrhage).

[0067] 2. Next, based on medical prior knowledge, the system determines the bleeding location by probabilistic selection during each synthesis, specifically including: ventricles and deep gray matter areas (applicable to Grades I-III, probability 30%), ventricles (excluding deep gray matter, probability 30%), deep gray matter areas (excluding ventricles, probability 30%), or white matter areas (applicable to Grade IV, probability 10%).

[0068] 3. Then, a uniform noise with the same dimension as the input image is generated and a smooth transition area is created by Gaussian blurring. Then, thresholding and morphological opening and closing operations are applied to smooth the shape and fill small holes, ensuring that the generated random shape is confined to the brain area. Specifically, given an input image I∈R H×W×C and threshold t∈R, the goal is to output a random shape mask M∈0,1 H×W First, generate random noise N∈R with the same dimension as I H×W , where the pixel values range from 0 to 255. The noisy image is then Gaussian blurred to obtain B∈R H×W , using a Gaussian kernel size of 15 in the X and Y directions. To maintain a reasonable pixel value range after blurring, B is rescaled to the range 0 to 255, resulting in S∈R H×WAt this point, an image with different light and dark areas is obtained. Next, a threshold operation is applied to the image to obtain T, and then an opening and closing operation is performed on T to smooth the shape and fill the holes to obtain the optimized mask image R∈0,1 H×W The final output mask image M is obtained by combining R with the brain mask M b ∈0,1 H×W The result is element-wise multiplication, ensuring that the generated random shape is located within the overall brain area.

[0069] 4. Finally, the random shape mask is combined with the selected hemorrhage area to create the final hemorrhage area mask. The pixel values in this area are multiplied by a scaling factor of 0.3-0.5 (because hemorrhage appears as a low signal in T2-weighted MRI). Gaussian blurring is then applied to soften the boundaries to obtain a medically reasonable pseudo GMH-IVH image. Through this process, a large number of pseudo abnormal samples for deep learning training can be efficiently generated without the need for real hemorrhage case data.

[0070] In step S103, the segmentation model is trained using the training data. The magnetic resonance imaging data of the target fetal brain is input into the trained segmentation model, and the segmentation model outputs the segmentation results of the target fetal brain lesions and intraventricular hemorrhage.

[0071] It can be understood that the embodiment of the present application can use training data to train the segmentation model, and input the magnetic resonance imaging data of the target fetal brain into the trained segmentation model. The segmentation model outputs the segmentation results of the target fetal brain lesions and intraventricular hemorrhage without relying on real lesion annotation data. By performing medical-driven synthetic lesion simulation on normal fetal brain MRI images, diversified and high-quality pseudo-abnormal data are generated, and the pseudo-abnormal data is used as training data to train the segmentation model, thereby achieving efficient and accurate automatic detection and segmentation of fetal brain lesions and intraventricular hemorrhage areas, improving the generalization ability and segmentation accuracy of the model, reducing manual participation and data acquisition thresholds, and thus providing more intelligent and reliable technical support for prenatal diagnosis and risk assessment.

[0072] In an embodiment of the present application, the segmentation model includes an image segmentation layer, stacked downsampling and encoding blocks, an upsampling path consisting of an upsampling layer and a convolution block, and a downsampling path consisting of a convolution block and a maximum pooling layer. The jump connection of the UNet variant structure adopts an addition operation.

[0073] It can be understood that the segmentation model of the embodiment of the present application includes an image segmentation layer, stacked downsampling and encoding blocks, an upsampling path consisting of an upsampling layer and a convolution block, and a downsampling path consisting of a convolution block and a maximum pooling layer, and the jump connection of the UNet variant structure adopts an addition operation.

[0074] Specifically, this application uses a new neural network architecture model (i.e., a segmentation model) to achieve unsupervised fetal intracerebral hemorrhage detection and lesion segmentation, specifically including:

[0075] First, in terms of the segmentation model, a variant UNet structure (VM-UNet) based on the Visual Mamba (Vmamba) technology was applied. This is a UNet variant that integrates VSS (Visual State Space) technology. It starts with a block embedding layer to divide the input image into multiple 4×4 blocks, which are then downsampled and encoded through stacked VSS modules. The block merging method is used to expand the feature channels while reducing the feature map size. The spatial resolution is restored through the block expansion method in the upsampling stage. Unlike the traditional UNet, the jump connection uses an addition operation instead of a splicing operation. The system also supports the classic UNet structure (UNet), which includes a downsampling path consisting of a convolution module and a maximum pooling layer, and an upsampling path consisting of an upsampling layer and a convolution module. The detail information is transferred between the two paths through a jump connection. The model training uses an adaptive weight optimizer with a learning rate of 0.001, a batch size of 8, and 10 rounds of training. When using the pseudo GMH-IVH dataset, only 1% of the images are selected in each round to prevent overfitting. The VSS module in VM-UNet is initialized with pre-trained visual Mamba (Vmamba) weights to accelerate convergence.

[0076] In an embodiment of the present application, in the process of training a segmentation model using training data, it also includes: optimizing the training process of the segmentation model using a segmentation optimization model, wherein the segmentation optimization model includes an adjustment module, a convolutional neural network branch, a cross-branch attention mechanism and an optimizer, the adjustment module adjusts the image encoder input size according to the medical image resolution, and when the convolutional neural network branch and the cross-branch attention mechanism are trained, the image and the point prompts randomly selected based on the segmentation mask are received as input, and the optimizer optimizes the training setting parameters of the segmentation model.

[0077] It can be understood that in the embodiment of the present application, in the process of using training data to train the segmentation model, the segmentation optimization model is also used to optimize the training process of the segmentation model to improve the detection and segmentation performance of the segmentation model. The segmentation optimization model includes an adjustment module, a convolutional neural network branch, a cross-branch attention mechanism and an optimizer. The adjustment module adjusts the image encoder input size according to the medical image resolution. When the convolutional neural network branch and the cross-branch attention mechanism are trained, the image and the point prompts randomly selected based on the segmentation mask are received as input, and the optimizer optimizes the training setting parameters of the segmentation model.

[0078] Specifically, this application also integrates the SAMUS segmentation optimization model based on SAM (Segment Anything Model), which retains the core architecture of SAM, but adjusts the input size of the image encoder to adapt to the medical image resolution, and introduces convolutional neural network branches and cross-branch attention mechanisms to promote cross-domain adaptation. During the training process, the model receives images and point prompts randomly selected based on segmentation masks as input, uses the Adam optimizer, and sets the learning rate to 0.0001 for 20 rounds of training. When using the pseudo GMH-IVH dataset, 0.5% of the images are selected each round to prevent overfitting. This dual-model collaborative architecture design greatly improves the detection and segmentation performance.

[0079] The following describes the supervised fetal cerebral hemorrhage detection and lesion segmentation method of the present application through a specific embodiment. The present application proposes a fetal germinal matrix and intraventricular hemorrhage automatic detection and segmentation method and system based on pseudo-abnormal synthesis guided by medical prior knowledge and unlabeled deep learning, named FreeHemoSeg framework, such as Figure 2 This method does not rely on real lesion annotation data and can accurately and efficiently achieve automatic diagnosis and precise segmentation of fetal GMH-IVH. The FreeHemoSeg framework consists of three stages: data synthesis, model training, and model inference.

[0080] 1. Data synthesis stage.

[0081] Medical prior knowledge is used to synthesize a large number of high-quality pseudo GMH-IVH images from normal fetal brain MRI images as training data for the deep learning model.

[0082] The process of generating pseudo data in this application includes the following steps:

[0083] 1. Obtain normal fetal brain MRI images and perform tissue segmentation;

[0084] 2. Select the bleeding location according to the probability distribution defined by medical a priori;

[0085] 3. Generate random shape masks and apply morphological operations to ensure medical plausibility;

[0086] 4. Combine the mask with the selected area and reduce the signal intensity to simulate bleeding characteristics;

[0087] 5. Apply Gaussian blur to soften the boundaries and generate the final pseudo GMH-IVH image.

[0088] Among them, the following probability distribution was used when selecting the bleeding location based on medical prior knowledge: 30% in the ventricles and deep gray matter areas (Grade I-III), 30% in the ventricles alone, 30% in the deep gray matter alone, and 10% in the white matter areas (Grade IV).

[0089] The random shape mask generation steps include:

[0090] 1. Generate and input image I∈R (H×W×C) Uniform random noise N∈R of the same dimension (H×W) ;

[0091] 2. Apply Gaussian blur (kernel size 15×15) to the noisy image to obtain B∈R (H×W) ;

[0092] 3. Rescale B to the range of 0-255 to get S∈R (H×W) ;

[0093] 4. Apply the threshold t∈R operation to S, and then perform the opening and closing operation to obtain R∈0,1 (H×W) ;

[0094] 5. Combine R with brain mask M b ∈0,1 (H×W) Element-by-element multiplication yields the final mask M. 2. Segmentation model training stage.

[0095] During the synthesis of pseudo-GMH-IVH images, the pixel values of the selected hemorrhage areas were adjusted to 0.3–0.5 times of the original values to simulate the low signal characteristics of hemorrhage in T2-weighted MRI.

[0096] The pseudo-abnormal synthesis method generates pseudo-GMH-IVH images from multiple anatomical planes (axial, sagittal, and coronal) to ensure data diversity.

[0097] 2. Segmentation model training stage.

[0098] Use the synthesized pseudo GMH-IVH images to train a segmentation model (such as VM-UNet or UNet) and fine-tune the SAM base model.

[0099] The training method for the segmentation model in this application includes the following steps:

[0100] 1. Use pseudo GMH-IVH images to train the VM-UNet segmentation model;

[0101] 2. Generate point hints based on the segmentation mask and fine-tune the SAM model to obtain SAMUS;

[0102] 3. Build a dual-model collaborative detection-segmentation process;

[0103] 4. Automatically generate point prompt mechanism through the maximum value of heat map.

[0104] Among them, the VM-UNet structure includes:

[0105] 1. The block embedding layer divides the input image into 4×4 blocks;

[0106] 2. Stacked VSS blocks are downsampled and encoded;

[0107] 3. Block merging expands feature channels and reduces feature map size;

[0108] 4. Block expansion for upsampling to restore spatial resolution;

[0109] 5. The addition operation realizes the jump connection to transmit detailed information.

[0110] The following loss function is used for model training:

[0111] L=λ1·L dice +λ2·L bce ;

[0112]

[0113] Among them, Y is the real mask, is the predicted mask, λ1 and λ2 are weight coefficients.

[0114] 3. Segmentation model inference stage.

[0115] The input fetal brain MRI is preprocessed and sliced, and the trained model is used to diagnose GMH-IVH and accurately segment the hemorrhage area. Specifically, the trained segmentation model generates a heat map, and the locations with the highest values in the heat map represent the areas with the highest hemorrhage probability, which is used for classification judgment. The maximum value of the heat map is then used as a point prompt input to further guide the fine-tuned SAM model to achieve accurate hemorrhage segmentation.

[0116] The embodiment of the present application provides a method and system for automatic detection and segmentation of fetal germinal matrix and intraventricular hemorrhage based on pseudo-anomaly synthesis guided by medical prior knowledge and unlabeled deep learning, including:

[0117] Data preprocessing module for format conversion and tissue segmentation;

[0118] Model training module, used to train VM-UNet detection model and SAMUS segmentation model;

[0119] GMH-IVH detection module, used to generate lesion heat maps and provide diagnostic results;

[0120] Point prompt generation module, used to generate the point prompt using the formula P = argmax(x,y) (H(x,y)) locates the lesion center;

[0121] Precise segmentation module for generating lesion area masks;

[0122] Performance evaluation module, used to evaluate the system diagnosis and segmentation accuracy.

[0123] Among them, the data preprocessing module includes:

[0124] Use DCM2NIIX software to convert DICOM format to NifTI format;

[0125] Use pre-trained Seg-Net to locate the brain;

[0126] Reconstruct high-resolution volumes using the GPU-accelerated NeSVoR method;

[0127] Tissue segmentation using a trusted AI framework based on Dempster-Shafer theory.

[0128] The model training module uses the AdamW optimizer (learning rate 0.001) for VM-UNet and the Adam optimizer (learning rate 0.0001) for SAMUS. When using the pseudo GMH-IVH dataset, VM-UNet selects 1% of images per round and SAMUS selects 0.5% of images per round to prevent overfitting.

[0129] The system uses multi-level evaluation indicators: AUROC and AUPR to evaluate diagnostic performance, DSC and mIOU to evaluate segmentation performance, and Point ACC to evaluate point cue generation accuracy.

[0130] The system can process multi-center heterogeneous data and maintain cross-center adaptability and performance stability; it automatically generates point prompts through the maximum value of the heat map, realizing a seamless transition from detection to segmentation; it can achieve high-precision GMH-IVH detection and segmentation without any real lesion annotation data, completely solving the problem of scarcity of fetal cerebral hemorrhage data.

[0131] In summary, the automated GMH-IVH diagnosis and segmentation process currently presented in this application includes a complete data processing chain. First, in the data preprocessing stage, the system uses DCM2NIIX software to convert the common clinical DICOM format into the easy-to-process NifTI format, and uses the pre-trained Seg-Net to locate the brain, and then uses the GPU-accelerated NeSVoR method to reconstruct high-resolution volumes from three brain-shielded thick-slice image stacks. Subsequently, a trusted AI framework based on the Dempster-Shafer theory is applied for tissue segmentation, dividing the fetal brain into eight regions: external cerebrospinal fluid, gray matter, white matter, ventricles, cerebellum, deep gray matter, brainstem / spinal cord, and corpus callosum. The system then generates slices from the reconstructed volume along three major anatomical planes (axial, sagittal, and coronal), and screens slices containing ventricles and deep gray matter as candidates based on the tissue segmentation results. During the GMH-IVH diagnosis phase, the system inputs the screened slices into a trained segmentation model to generate a heat map. High-value areas indicate a higher likelihood of bleeding. The system interprets the highest value in the heat map as the probability of bleeding, which serves as the basis for slice- and case-level diagnosis. Once bleeding is confirmed, the system uses the coordinates corresponding to the maximum value in the heat map as a point prompt and inputs them into the adapted SAM model for more precise slice-level segmentation, thereby accurately identifying the location and extent of bleeding. This highly automated process design significantly reduces manual intervention and improves diagnostic efficiency and accuracy.

[0132] In addition, this application uses a comprehensive scientific evaluation system to verify the performance of the system. In terms of diagnostic performance, the system uses the area under the receiver operating characteristic curve (AUROC) and the area under the precision-recall curve (AUPR) to evaluate the overall performance, and quantifies the clinical practicality by calculating the sensitivity at a fixed specificity (such as 80%) and the specificity at a fixed sensitivity (such as 80%). In terms of segmentation performance, the system uses the slice-level dice similarity coefficient (DSC) and the mean intersection-over-union (mIOU) to evaluate segmentation accuracy, and introduces the point accuracy (Point ACC) indicator to evaluate the accuracy of point prompt generation, that is, whether the point prompt is located in the real bleeding area.

[0133] Table 1 compares the diagnostic accuracy of our unsupervised fetal intracerebral hemorrhage detection and lesion segmentation methods on a test dataset. Table 1 categorizes the model training data into three groups: a synthetic training set developed in this application, a manually labeled training set, and a normal image training set (used for training the anomaly detection model). The AUROC, AUPR, sensitivity, and specificity of each model at both the case and slice levels are listed, along with 95% confidence intervals.

[0134] Table 1

[0135]

[0136]

[0137] The case-level and slice-level diagnostic performance of the segmentation model of this application on the test dataset is as follows: Figure 3 As shown, (A) is the case-level receiver operating characteristic curve (ROC curve) analysis of GMH-IVH diagnosis; (B) is the case-level precision-recall curve (PR curve) of GMH-IVH diagnosis; (C) is the confusion matrix result of the case-level classification of the VM-Unet (A) model (i.e., the model trained with data of the data synthesis method proposed in this application) at the threshold determined by maximizing the Youden index; (D) is the slice-level ROC curve analysis of GMH-IVH diagnosis; (E) is the slice-level PR curve of GMH-IVH diagnosis; (F) is the confusion matrix result of the slice-level classification of the VM-Unet (A) model at the threshold determined by maximizing the Youden index.

[0138] The slice-level segmentation results of cerebral hemorrhage in this application are as follows Figure 4 As shown, Figure 4 RS represents the reference standard, VM-Unet (C) (i.e., the model trained with the traditional manually annotated data set) and VM-Unet (A) (i.e., the model trained with the data synthesis method proposed in this application) refer to the models trained with different training data sets C and A, respectively. This application uses the VM-Unet model to perform slice-level automatic segmentation of cerebral hemorrhage cases of different grades (grades I-IV), and intuitively displays the location and range of blood in low-signal areas in T2-weighted magnetic resonance images. (a) In grade I, the hemorrhage is confined to the germinal matrix area; (b) In grade II, the hemorrhage extends into the ventricle; (c) In grade III, the hemorrhage has penetrated the ventricle and caused obvious ventricular enlargement; (d) In grade IV, the hemorrhage further spreads and infiltrates into the brain parenchyma area. The results fully demonstrate the fine segmentation capability and wide applicability of this application in cases of hemorrhage of different severities.

[0139] The comparison results of slice-level segmentation accuracy using the Segment Everything Model (SAM) in this application are shown in Table 2. Table 2 mainly statistics the Dice similarity coefficient (DSC), mean intersection-over-union (mIOU) and point cue generation accuracy (Point ACC) of each method. Through horizontal comparison, the system evaluated the changes in segmentation performance before and after the introduction of the SAM model. The results show that the segmentation scheme using the SAM model is superior to the traditional method without the introduction of SAM in terms of DSC, mIOU and point cue generation accuracy, which fully demonstrates the significant advantages of SAM in improving the segmentation accuracy of cerebral hemorrhage areas and the efficiency of point cue guidance.

[0140] Table 2

[0141]

[0142] This application uses the automated process of fetal brain MRI 3D reconstruction and tissue segmentation as a preprocessing method, such as Figure 5 As shown in the figure, the process first uses the pre-trained neural network model Seg-Net to process the raw data and automatically extract the fetal brain mask. Subsequently, the NeSVoR method based on the implicit neural network aligns and reconstructs the multi-sequence slices to generate high-resolution three-dimensional brain volume data. Finally, this application uses the pre-trained nn-Unet model to achieve fine segmentation of three-dimensional brain tissue, achieving automation and efficiency of the entire process, which can lay a solid foundation for subsequent intelligent diagnosis and structural analysis.

[0143] The comparison of the output heat map of the model VM-Unet (A) trained with synthetic data in this application and the mainstream anomaly detection models (Skip-TS, RD4AD and IKD) is shown in Figure 2. Figure 6 As shown in the figure, it can be seen that although VM-Unet (A) was not trained using any manually annotated real GMH-IVH data, the heat map generated by its final layer is able to achieve fine-grained and accurate segmentation of the lesion area. The three unsupervised anomaly detection models can only roughly locate the abnormal area, among which IKD performs relatively well, while Skip-TS has the weakest positioning ability. This comparison result further highlights that the method of this application can still achieve high-level automatic lesion detection and segmentation without manual segmentation.

[0144] This application uses segmentation model enhancement to visualize the slice-level false positive segmentation results. Figure 7 As shown, the first row is the original input brain slice; the second row is the reference standard of the bleeding area annotated by the expert; the third row is the heat map output by VM-Unet (A) (i.e., the model trained with synthetic data proposed in this application), which can indicate the high probability area of bleeding, but there are false positives in non-lesion tissues (indicated by red arrows); the fourth row is the point prompt generated according to the maximum coordinates of the heat map (also marked with red arrows); the fifth row is the result of segmentation by the adaptive SAM model. Under the guidance of the point prompt, the model only achieves accurate segmentation of the real bleeding area, effectively eliminating the false detection of the non-lesion area, thereby improving the accuracy and specificity of the segmentation.

[0145] This application uses post-processing based on the segmentation model to enhance the visualization of slice-level false negative segmentation effects. Figure 8As shown, the first row is the original brain slice input; the second row is the bleeding area standard annotated by experts; the third row is the heat map output by VM-Unet (A) (i.e., the model trained with synthetic data proposed in this application). Although the signals of some bleeding areas are weak, the highest probability points are still located in the real bleeding areas, while other areas have false negative omissions; the fourth row is the point prompt (marked by a red arrow) generated based on the maximum coordinates of the heat map. Although the signal of this point is not strong in the lesion area, it still accurately points to the bleeding area; the fifth row is the segmentation result processed by the adaptive SAM model, which can accurately capture and segment the real bleeding area under the guidance of the point prompt, thereby effectively correcting the original false negative segmentation error of the model.

[0146] In general, compared with the prior art, this application can achieve the following beneficial effects:

[0147] 1. Deep Learning without Labeled Data: This application addresses the core challenge of the extreme scarcity of labeled fetal GMH-IVH data. By synthesizing pseudo-abnormalities guided by prior medical knowledge, deep learning model training is achieved without the need for labeled data from real cases. This lowers the barrier to data acquisition and increases the feasibility of model development and application. This approach has consistently maintained stable performance on multi-center validation data, demonstrating the rationality and practical value of its technical approach.

[0148] 2. Compared to existing clinical routine methods based on subjective judgment by radiologists, this application achieves an AUROC of 0.950 and an AUPR of 0.921 for case-level GMH-IVH diagnosis, far exceeding conventional clinical diagnostic standards. This significantly enhances the ability to detect early-stage, small, and occult lesions, effectively reducing missed diagnoses and facilitating early intervention.

[0149] 3. Dual Optimization of Segmentation Accuracy and Efficiency: This application achieved a 0.599 accuracy rate for slice-level GMH-IVH segmentation and an 81.0% accuracy rate for point hinting. Furthermore, the system reduced the processing time for a single fetal MRI examination from an average of 30-45 minutes for expert manual analysis to less than 2 minutes, improving segmentation efficiency and laying the foundation for large-scale clinical application.

[0150] 4. Versatility and Scalability of Pseudo-Anomaly Synthesis Technology: This application pioneers a pseudo-anomaly synthesis technology guided by medical prior knowledge. This technology is not only applicable to fetal GMH-IVH but can also be extended to the detection and segmentation of other rare medical lesions. Experiments have demonstrated that this technology excels in the authenticity and diversity of synthesized images. The generated pseudo-lesions are highly similar to real lesions in texture, location, and visual appearance, providing a new approach to addressing the widespread data scarcity problem in the medical imaging field.

[0151] 5. Multi-model collaborative architecture optimizes performance: The collaborative working mechanism of VM-UNet and SAMUS based on SAM realizes a solution from lesion detection to precise segmentation, especially with obvious advantages in accurately depicting lesion boundaries and identifying tiny lesions.

[0152] 6. Enhanced value of clinical decision support: This application can provide objective auxiliary judgment basis for prenatal diagnosis, help clinical physicians assess the severity and location distribution of GMH-IVH, and provide important reference for accurate risk assessment, pregnancy management and postnatal treatment plan formulation.

[0153] 7. Optimization of the timing of prenatal intervention: This application provides a valuable time window for clinical intervention by accurately identifying fetal GMH-IVH early.

[0154] According to the unsupervised fetal brain hemorrhage detection and lesion segmentation method proposed in the embodiment of the present application, medical prior knowledge can be used to synthesize diverse, high-quality pseudo fetal lesions and intraventricular hemorrhage images from magnetic resonance imaging data of the normal fetal brain, and training data can be generated based on multiple pseudo fetal lesions and intraventricular hemorrhage images. The training data is used to train the segmentation model, and no real case labeled data is required for training the segmentation model, which reduces manual participation and data acquisition thresholds, thereby improving the feasibility of model development and application. The trained segmentation model can achieve segmentation of target fetal brain lesions and intraventricular hemorrhage, which has a higher accuracy rate than manual clinical subjective judgment, providing more reliable technical support for subsequent prenatal diagnosis and risk assessment.

[0155] Next, the unsupervised fetal cerebral hemorrhage detection and lesion segmentation device proposed in accordance with an embodiment of the present application will be described with reference to the accompanying drawings.

[0156] Figure 9 3 is a block diagram of an unsupervised fetal cerebral hemorrhage detection and lesion segmentation device according to an embodiment of the present application.

[0157] like Figure 9 As shown, the unsupervised fetal cerebral hemorrhage detection and lesion segmentation device 10 includes: an acquisition module 100, a synthesis module 200 and a training module 300.

[0158] Among them, the acquisition module 100 is used to obtain magnetic resonance imaging data of the normal fetal brain; the synthesis module 200 is used to use medical prior knowledge to synthesize multiple pseudo fetal lesions and intraventricular hemorrhage images from the magnetic resonance imaging data of the normal fetal brain, and generate training data based on the multiple pseudo fetal lesions and intraventricular hemorrhage images; the training module 300 is used to use the training data to train the segmentation model, input the magnetic resonance imaging data of the target fetal brain into the trained segmentation model, and the segmentation model outputs the segmentation results of the target fetal brain lesions and intraventricular hemorrhage.

[0159] In an embodiment of the present application, the synthesis module 200 is further used to: obtain tissue segmentation results from magnetic resonance imaging data of a normal fetal brain; utilize medical prior knowledge to probabilistically select the bleeding location in the tissue segmentation result during each synthesis, generate a noise image of the same dimension as the input image, and generate a random shape mask based on the noise image; combine the random shape mask with the bleeding area to obtain a bleeding area mask, and generate an image of a pseudo fetal lesion and intraventricular hemorrhage based on the bleeding area mask during each synthesis.

[0160] In an embodiment of the present application, the synthesis module 200 is further used to extract slices along the axial, sagittal and coronal planes of the brain volume from the magnetic resonance imaging data of the normal fetal brain, and generate tissue segmentation results based on the slices along the axial, sagittal and coronal planes of the brain volume.

[0161] In an embodiment of the present application, the synthesis module 200 is further used to: perform Gaussian blur processing on the noise image; scale the image after Gaussian blur processing to obtain an image with different light and dark areas; perform threshold operations, opening operations and closing operations on the image with different light and dark areas to obtain a mask image, and obtain a random shape mask by element-by-element multiplication of the mask image and the brain mask.

[0162] In an embodiment of the present application, the synthesis module 200 is further used to: scale the pixel values within the hemorrhage area mask area during each synthesis; perform Gaussian blur softening processing on the hemorrhage area mask area to obtain images of pseudo fetal lesions and intraventricular hemorrhage.

[0163] In an embodiment of the present application, the segmentation model includes an image segmentation layer, stacked downsampling and encoding blocks, an upsampling path consisting of an upsampling layer and a convolution block, and a downsampling path consisting of a convolution block and a maximum pooling layer. The jump connection of the UNet variant structure adopts an addition operation.

[0164] In an embodiment of the present application, the training module 300 is further used to: optimize the training process of the segmentation model using a segmentation optimization model, wherein the segmentation optimization model includes an adjustment module, a convolutional neural network branch, a cross-branch attention mechanism and an optimizer, the adjustment module adjusts the image encoder input size according to the medical image resolution, and when the convolutional neural network branch and the cross-branch attention mechanism are trained, they receive images and point prompts randomly selected based on the segmentation mask as input, and the optimizer optimizes the training setting parameters of the segmentation model.

[0165] It should be noted that the above explanation of the embodiment of the unsupervised fetal cerebral hemorrhage detection and lesion segmentation method is also applicable to the unsupervised fetal cerebral hemorrhage detection and lesion segmentation device of this embodiment, and will not be repeated here.

[0166] According to the unsupervised fetal brain hemorrhage detection and lesion segmentation device proposed in the embodiment of the present application, medical prior knowledge can be used to synthesize diverse, high-quality pseudo fetal lesions and intraventricular hemorrhage images from the magnetic resonance imaging data of the normal fetal brain, and training data can be generated based on multiple pseudo fetal lesions and intraventricular hemorrhage images. The training data is used to train the segmentation model, and no real case labeled data is required for training the segmentation model, which reduces manual participation and data acquisition thresholds, thereby improving the feasibility of model development and application. The trained segmentation model can achieve segmentation of target fetal brain lesions and intraventricular hemorrhage, which has a higher accuracy rate than manual clinical subjective judgment, providing more reliable technical support for subsequent prenatal diagnosis and risk assessment.

[0167] Figure 10 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present application. The electronic device may include:

[0168] A memory 1001 , a processor 1002 , and a computer program stored in the memory 1001 and executable on the processor 1002 .

[0169] When the processor 1002 executes the program, the unsupervised fetal cerebral hemorrhage detection and lesion segmentation method provided in the above embodiment is implemented.

[0170] Furthermore, the electronic device further includes:

[0171] The communication interface 1003 is used for communication between the memory 1001 and the processor 1002 .

[0172] The memory 1001 is used to store computer programs that can be run on the processor 1002 .

[0173] The memory 1001 may include a high-speed RAM (Random Access Memory) memory, and may also include a non-volatile memory, such as at least one disk memory.

[0174] If the memory 1001, the processor 1002, and the communication interface 1003 are implemented independently, the communication interface 1003, the memory 1001, and the processor 1002 can be connected to each other via a bus and communicate with each other. The bus can be an ISA (Industry Standard Architecture) bus, a PCI (Peripheral Component Interconnect) bus, or an EISA (Extended Industry Standard Architecture) bus. The bus can be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Figure 10Only one thick line is used in the diagram, but this does not mean that there is only one bus or one type of bus.

[0175] Optionally, in a specific implementation, if the memory 1001, the processor 1002 and the communication interface 1003 are integrated on a chip, the memory 1001, the processor 1002 and the communication interface 1003 can communicate with each other through an internal interface.

[0176] The processor 1002 may be a CPU (Central Processing Unit), or an ASIC (Application Specific Integrated Circuit), or one or more integrated circuits configured to implement the embodiments of the present application.

[0177] An embodiment of the present application also provides a computer-readable storage medium having a computer program or instruction stored thereon. When the computer program or instruction is executed by a processor, the above-mentioned unsupervised fetal cerebral hemorrhage detection and lesion segmentation method is implemented.

[0178] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present application. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or N embodiments or examples in a suitable manner. In addition, those skilled in the art can combine and combine different embodiments or examples described in this specification and features of different embodiments or examples without contradiction.

[0179] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be understood to indicate or imply relative importance or implicitly specify the number of technical features indicated. Thus, a feature specified as "first" or "second" may explicitly or implicitly include at least one such feature. In the description of this application, "N" means at least two, for example, two, three, etc., unless otherwise specifically defined.

[0180] Any process or method description in a flowchart or otherwise described herein may be understood to represent a module, fragment or portion of code comprising one or N executable instructions for implementing a custom logical function or process step, and the scope of the preferred embodiments of the present application includes alternative implementations in which functions may be performed in a different order than shown or discussed, including performing functions in a substantially simultaneous manner or in a reverse order depending on the functions involved, which should be understood by those skilled in the art to which the embodiments of the present application pertain.

[0181] It should be understood that various parts of the present application can be implemented using hardware, software, firmware, or a combination thereof. In the above-described embodiments, the steps or methods can be implemented using software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented using hardware, as in another embodiment, any one of the following technologies known in the art or a combination thereof can be used to implement the method: a discrete logic circuit having a logic gate circuit for implementing a logic function on a data signal, an application-specific integrated circuit having a suitable combination of logic gate circuits, a programmable gate array, a field programmable gate array, etc.

[0182] A person skilled in the art may understand that all or part of the steps carried out in the method for implementing the above-mentioned embodiment may be completed by instructing the relevant hardware through a program, and the above-mentioned program may be stored in a computer-readable storage medium, which, when executed, includes one of the steps of the method embodiment or a combination thereof.

[0183] Although the embodiments of the present application have been shown and described above, it can be understood that the above embodiments are exemplary and cannot be understood as limitations on the present application. Ordinary technicians in this field can change, modify, replace and modify the above embodiments within the scope of the present application.

Claims

1. An unsupervised fetal cerebral hemorrhage detection and lesion segmentation method, characterized in that: The following steps are involved: Obtaining magnetic resonance imaging data of the normal fetal brain; synthesizing a plurality of pseudo images of fetal lesions and intraventricular hemorrhage from the magnetic resonance imaging data of the normal fetal brain using prior medical knowledge, and generating training data based on the plurality of pseudo images of fetal lesions and intraventricular hemorrhage; The training data is used to train a segmentation model, and the magnetic resonance imaging data of the target fetal brain is input into the trained segmentation model, and the segmentation model outputs a segmentation result of the target fetal brain lesion and intraventricular hemorrhage.

2. The unsupervised fetal cerebral hemorrhage detection and lesion segmentation method according to claim 1, characterized in that: The method of synthesizing a plurality of pseudo fetal lesions and intraventricular hemorrhage images from the magnetic resonance imaging data of the normal fetal brain using medical prior knowledge includes: Acquire a tissue segmentation result from the magnetic resonance imaging data of the normal fetal brain; Using medical prior knowledge, the bleeding location in the tissue segmentation result is selected by probability during each synthesis, a noise image with the same dimension as the input image is generated, and a random shape mask is generated based on the noise image; The random shape mask is combined with the hemorrhage region to obtain a hemorrhage region mask, and an image of a pseudo fetal lesion and intraventricular hemorrhage is obtained according to the hemorrhage region mask during each synthesis.

3. The unsupervised fetal cerebral hemorrhage detection and lesion segmentation method according to claim 2, characterized in that: The obtaining of a tissue segmentation result from the magnetic resonance imaging data of the normal fetal brain includes: Slices are extracted from the magnetic resonance imaging data of the normal fetal brain along the axial, sagittal and coronal planes of the brain volume, and the tissue segmentation result is generated according to the slices of the brain volume along the axial, sagittal and coronal planes.

4. The unsupervised fetal cerebral hemorrhage detection and lesion segmentation method according to claim 2, characterized in that: Generating a random shape mask according to the noise image includes: Performing Gaussian blur processing on the noise image; The image after Gaussian blur processing is scaled to obtain an image with different light and dark areas; A threshold operation, an opening operation, and a closing operation are performed on the image with different light and dark areas to obtain a mask image, and a random shape mask is obtained by element-by-element multiplication of the mask image and the brain mask.

5. The unsupervised fetal cerebral hemorrhage detection and lesion segmentation method according to claim 2, characterized in that: The masking of the hemorrhage area into pseudo fetal lesions and intraventricular hemorrhage images according to each synthesis comprises: Scaling the pixel values within the bleeding area mask during each synthesis; Gaussian blurring and softening of the boundary of the hemorrhage area mask region are performed to obtain images of pseudo fetal lesions and intraventricular hemorrhage.

6. The unsupervised fetal cerebral hemorrhage detection and lesion segmentation method according to claim 1, characterized in that: The segmentation model includes an image segmentation layer, stacked downsampling and encoding blocks, an upsampling path consisting of an upsampling layer and a convolution block, and a downsampling path consisting of a convolution block and a maximum pooling layer. The jump connection of the UNet variant structure adopts an addition operation.

7. The unsupervised fetal cerebral hemorrhage detection and lesion segmentation method according to claim 1, characterized in that: The process of training the segmentation model using the training data further includes: A segmentation optimization model is used to optimize the training process of the segmentation model, wherein the segmentation optimization model includes an adjustment module, a convolutional neural network branch, a cross-branch attention mechanism and an optimizer, the adjustment module adjusts the image encoder input size according to the medical image resolution, the convolutional neural network branch and the cross-branch attention mechanism receive images and point prompts randomly selected based on segmentation masks as input during training, and the optimizer optimizes the training setting parameters of the segmentation model.

8. An unsupervised fetal cerebral hemorrhage detection and lesion segmentation device, characterized in that: include: an acquisition module, for acquiring magnetic resonance imaging data of a normal fetal brain; a synthesis module, configured to synthesize a plurality of pseudo images of fetal lesions and intraventricular hemorrhage from the magnetic resonance imaging data of the normal fetal brain using prior medical knowledge, and generate training data based on the plurality of pseudo images of fetal lesions and intraventricular hemorrhage; A training module is used to train a segmentation model using the training data, input the magnetic resonance imaging data of the target fetal brain into the trained segmentation model, and the segmentation model outputs the segmentation results of the target fetal brain lesions and intraventricular hemorrhage.

9. An electronic device, characterized in that: include: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the unsupervised fetal cerebral hemorrhage detection and lesion segmentation method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program or instruction stored thereon, characterized in that: When the computer program or instructions are executed, the unsupervised fetal cerebral hemorrhage detection and lesion segmentation method according to any one of claims 1 to 7 is implemented.

Images