TWI897721B - An artificial intelligence-assisted identification system using ultrasound imaging to detect fetal brain abnormalities - Google Patents

Abstract

An auxiliary identification system that uses ultrasound imaging and artificial intelligence to detect fetal brain abnormalities can be used when ultrasound images are blurred and local anatomical features are less obvious. The system and method, which utilize artificial intelligence technology to process fetal ultrasound images, can effectively enhance the convolutional neural network's ability to identify key anatomical locations important to fetal development, further reducing the burden on clinical physicians and improving the accuracy of fetal developmental abnormality interpretation.

Description

An artificial intelligence-assisted identification system using ultrasound imaging to detect fetal brain abnormalities

This invention belongs to the field of diagnostic aid measurement technology. Specifically, it is an auxiliary identification system that uses ultrasound imaging and artificial intelligence to detect fetal brain abnormalities. This system can improve the ability to identify key anatomical locations of fetal development, thereby reducing the burden on clinical physicians and improving the accuracy of fetal developmental abnormality interpretation.

Ultrasound imaging is one of the most basic medical imaging diagnostic techniques. Due to its convenience and non-invasive nature, ultrasound is currently the most popular and effective imaging diagnostic tool for fetal structural abnormalities during prenatal development. This is particularly true for fetal brain abnormalities, which occur in 1.4-1.6 out of 1,000 pregnancies and account for 3-6% of fetal stillbirths. Therefore, prenatal fetal ultrasound examinations are crucial for the early diagnosis of fetal structural abnormalities.

However, the quality of ultrasound images is often difficult to control. Ultrasound image resolution is often blurry, making it difficult to locate and identify the fetal position and body posture. Furthermore, issues such as image blurring and noise caused by movement during ultrasound capture often result in unclear ultrasound images, hindering clinical interpretation. Furthermore, fetal ultrasound image interpretation requires a long learning curve for obstetricians and gynecologists, and the ability to expertly interpret subtle fetal brain abnormalities also requires specialized training. Therefore, not all obstetricians and gynecologists specialize in ultrasound operation and interpretation.

In recent years, deep learning technology, using convolutional networks as its primary architecture, has achieved remarkable recognition performance in many medical applications. Patents such as Patent Publication No. I802486, "Method for Interpreting Pediatric Kidney Ultrasound Images Using Artificial Intelligence," Patent Publication No. I810498, "Intelligent Analysis Device for Liver Tumors," and Patent Publication No. I811129, "System and Method for Aiding Target Detection and Recognition in Pediatric Congenital Heart Ultrasound Images," all assist clinical physicians in making judgments in various fields, reducing their workload.

In other words, if artificial intelligence deep learning technology can assist clinicians in identifying the anatomical locations of fetal developmental abnormalities in the fine details of fetal ultrasound images, it will significantly reduce the burden on clinicians in interpreting the findings and lower the risk of missing fetal developmental abnormalities. Detecting fetal developmental abnormalities as early as possible and allowing pregnant women to be referred to medical centers for more detailed examinations are the key issues and solutions that this invention aims to address.

In light of these needs, the inventors of this invention believed that further development was necessary. Leveraging their years of experience in related technologies and product design and manufacturing, they conducted research and development to address these shortcomings, actively seeking solutions. Through continuous research and testing, they developed an auxiliary identification system that uses ultrasound imaging and artificial intelligence to detect fetal brain abnormalities. This system addresses the inconvenience and difficulty in interpreting existing fetal ultrasound images, often caused by the poor resolution of these images.

Therefore, the primary objective of this invention is to provide an auxiliary identification system that uses ultrasound imaging to detect fetal brain abnormalities using artificial intelligence. This system can effectively process fetal ultrasound images even when the ultrasound images are blurred and local anatomical features are less obvious, thereby enhancing the effectiveness of the convolutional neural network in identifying key anatomical locations important to fetal development.

Another primary objective of this invention is to provide an auxiliary identification system that uses ultrasound imaging and artificial intelligence to detect fetal brain abnormalities. This system can be used for automatic identification, measurement, and detection of abnormal development in fetal ultrasound images, effectively reducing the burden on clinical physicians and improving the accuracy of developmental abnormality interpretation, allowing medical personnel to monitor and intervene in real-time.

To this end, the present invention primarily achieves the aforementioned objectives and functions through the following technical means: a hierarchical anatomical region detection module, a key tissue detection module, a key anatomical landmark detection module, a biometrics module, and an abnormal development interpretation module;

The hierarchical anatomical region detection module receives a fetal cerebellum plane ultrasound image as input, and then captures a local region image after hierarchical detection from top to bottom. The process of gradually reducing the local ROI image from top to bottom can be represented by a fetal ultrasound image anatomy tree data structure to analyze the fetal ultrasound image in a highly reliable and stable manner. Separate ROI images layer by layer, and these ROI images contain one or more key tissues. However, the ROI images do not have to be completely aligned with the key tissue images. The definition of the ROI images must take into account that when the convolutional network is detecting the ROI image boundaries, each ROI region can have sufficient and reliable image bioanatomical features provided to the convolutional network so that it can accurately and automatically detect the ROI image;

The key tissue detection module is connected to the hierarchical anatomical region detection module. Its main function is to use object detection methods to find all key tissues within the local ROI image captured by the hierarchical anatomical region detection module.

The key anatomical landmark detection module is connected to the key tissue detection module, and uses the key tissue location information that has been found to find all key anatomical landmarks in the local ROI image using a key point detection method;

The biological measurement module is connected to the key anatomical landmark detection module, which measures the distance between key anatomical landmarks or important biological development characteristics using image segmentation technology;

Furthermore, the abnormal development judgment module is connected to the biometrics module and determines whether a specific developmental abnormality risk condition has occurred based on the biological development characteristic values calculated by the biometrics module in accordance with medical standards or machine learning methods.

In one embodiment of the present invention, the fetal ultrasound image analysis tree T is a tree structure, and the root node of the image analysis tree T represents the entire original fetal ultrasound input image; one or more branch nodes are expanded from the root node, each branch node represents a local ROI image, and this local ROI image is automatically detected and captured from the local ROI image corresponding to the upper-level node through a specific object detection convolutional network. Each branch node can also expand one or more branch nodes, and the upper-level branch node ROI image includes the lower-level branch node ROI image, thereby generating a tree structure of upstream and downstream image analysis relationships. In addition, the end of each branch node has the image analysis tree T One or more terminal branch nodes, each terminal branch node can be further divided into one or more key tissues or key anatomical landmarks, wherein each key tissue is detected from the local ROI image corresponding to the terminal branch node through a specific object detection or image segmentation algorithm, and each key anatomical landmark is detected from the local ROI image corresponding to the terminal branch node through a specific key point detection algorithm.

In one embodiment of the present invention, the hierarchical anatomical region detection module receives a fetal cerebellum plane ultrasound input image, and the first-level node of the fetal ultrasound image analysis tree corresponding to the fetal ultrasound image is a cranial ROI image. I1, and the second-layer node of the cranial ROI image I1 includes a transparent septum ROI image I1,1 and a posterior cervical thickness ROI image I1,2, wherein the transparent septum ROI image I1,1 includes a key tissue transparent septum T1, and the third-layer node of the posterior cervical thickness ROI image I1,2 includes a cerebello-medullary cistern ROI image I1,2,1, and the cerebello-medullary cistern ROI image I1,2,1 includes a key tissue cerebellum T2, as well as three sets of key anatomical landmarks: cerebellum transverse diameter anatomical landmarks P1, P2, cerebello-medullary cistern diameter anatomical landmarks P3, P4, and posterior cervical thickness anatomical landmarks P5, P6.

In one embodiment of the present invention, the hierarchical anatomical region detection module detects the skull, cerebellum, and septum pellucidum in the fetal cerebellum planar ultrasound input image. The local image features that are obvious and difficult to identify independently are the cerebellomedullary cistern, the posterior neck thickness, and six anatomical landmarks closely related to biological developmental characteristics, including the cerebellum transverse diameter anatomical landmarks P1 and P2, the cerebellomedullary cistern diameter anatomical landmarks P3 and P4, and the posterior neck thickness anatomical landmark P5. P6 uses the skull as the most prominent local image feature. First, the skull ROI local image is identified. Images outside the skull are excluded to reduce the complexity of subsequent analysis. After limiting the scope of subsequent analysis to the skull ROI image I1, the next step is to use the cerebellum, a local image with distinct features, combined with the posterior neck skin and occipital bone, with distinct boundary features. These two image features are combined to identify the posterior neck thickness ROI image I1,2 and the cerebellomedullary cistern ROI image I1,2,1, thereby improving the accuracy of identifying the less prominent local images.

In one embodiment of the present invention, the biometrics module calculates the cerebellum transverse diameter by the distance between the aforementioned automatic cerebellum transverse diameter anatomical landmarks P1 and P2, and further verifies the reliability of the measurement data of P1, P2 and the cerebellum transverse diameter based on the detected position of the boundary frame of the key tissue cerebellum T2. The biometrics module also calculates the cerebellum medullary cistern diameter by the distance between the aforementioned automatic cerebellum medullary cistern diameter anatomical landmarks P3 and P4, and further verifies the reliability of the measurement data of P1, P2 and the cerebellum transverse diameter based on the detected position of the boundary frame of the key tissue cerebellum T2. The reliability of the measurement data for P3, P4, and the cisterna magna diameter was verified by the bounding box positions of the cerebellum ROI image I1,2,1 and the key tissue cerebellum T2. The biometrics module also calculated the posterior neck thickness using the distance between the aforementioned posterior neck thickness anatomical landmarks P5 and P6. The reliability of the measurement data for P5, P6, and the posterior neck thickness was further verified based on the bounding box positions of the detected cerebellum cistern ROI image I1,2,1 and the posterior neck thickness ROI image I1,2.

Through the specific implementation of the aforementioned technical means, the present invention can significantly enhance its practicality, increase its added value, and improve its economic benefits.

To help you better understand the structure, features, and other purposes of this invention, several preferred embodiments of this invention are listed below, along with detailed descriptions of the drawings, so that those skilled in the art can implement them.

The accompanying drawings illustrate specific embodiments of the present invention and their components. References to front and back, left and right, top and bottom, upper and lower, and horizontal and vertical are for descriptive purposes only and are not intended to limit the present invention or its components to any particular position or spatial orientation. Dimensions specified in the drawings and description may vary according to the design and requirements of the specific embodiments of the present invention without departing from the scope of the claims.

The present invention is an auxiliary identification system for detecting fetal brain abnormalities by using ultrasound imaging and artificial intelligence. Please refer to FIG1 , which includes a hierarchical anatomical region detection module (10), a key tissue detection module (20), a key anatomical landmark detection module (30), a biometrics module (40) and an abnormal development interpretation module (50), wherein the key tissue detection module (20) is connected to the hierarchical anatomical region detection module (10), the key anatomical landmark detection module (30) is connected to the key tissue detection module (20), and the biometrics module (40) is connected to the abnormal development interpretation module (50). 0) is connected to the key anatomical landmark detection module (30), and the abnormal development judgment module (50) is connected to the biometrics module (40) for inputting a fetal cerebellum plane ultrasound image (P1) into the hierarchical anatomical region detection module (10). After passing through the key tissue detection module (20), the key anatomical landmark detection module (30), the biometrics module (40) and the abnormal development judgment module (50), a fetal ultrasound examination report (100) can be generated to assist clinical physicians in identifying whether the fetal brain growth is abnormal.

Referring to Figure 2, according to a primary embodiment of the present invention, the fetal cerebellum planar ultrasound image (P1) includes, but is not limited to, biparietal diameter, cerebellum diameter, cisterna magna, and nuchal fold thickness. Previous methods using deep learning to measure these important biparietal developmental features primarily involved detecting the smallest bounding boxes (BBs) encompassing these biparietal developmental features, such as BB1 for the cranial bounding box, BB2 for the cerebellum bounding box, BB3 for the cisterna magna bounding box, and BB4 for the nuchal fold thickness.

Please also refer to Figures 3 and 4 . In an embodiment of the present invention, for the input image of the fetal cerebellum planar ultrasound image (P1), the image features of the skull and cerebellum are relatively obvious, and the deep learning object detection technology is used to better identify the skull bounding box BB1 and the cerebellum bounding box BB2. However, the image features of the cisterna magna and the nuchal fold are less obvious and lack a good identification reference, resulting in poor recognition of the cisterna magna bounding box BB3 and the nuchal fold bounding box BB4. As shown in Figure 3, the precision-recall curves for the recognition results of the skull bounding box BB1 and the cerebellum bounding box BB2 are significantly higher than the average of the four bounding box recognition results, with the curves located closer to the upper right. Meanwhile, the precision-recall curves for the recognition results of the cerebellomedullary ventral ... ventral ventral vent ventral vent ventral vent ventral vent ventral vent ventral vent ventral vent ventral vent ventral vent ventral vent ventral vent vent vent vent vent vent vent vent vent vent vent vent vent

Please refer to FIG5 . In an embodiment of the present invention, the hierarchical anatomical region detection module (10) utilizes a method of gradually reducing the image of a region of interest (hereinafter referred to as ROI) from top to bottom, step by step eliminating images irrelevant to the final required biological development characteristics to reduce the complexity of subsequent analysis, and finally retains only the smallest local ROI image containing the biological development characteristics, thereby improving the recognition efficiency of the biological development characteristics. Therefore, the process of gradually reducing the local ROI image from top to bottom can be represented by a data structure of an image parse tree for fetal ultrasound biometry to represent the analysis process of the fetal ultrasound image. A fetal ultrasound image analysis tree T is a tree structure, where the root node of the image analysis tree T represents the entire original fetal ultrasound input image; one or more branch nodes are expanded from the root node, each branch node represents a local ROI image, which is automatically detected and extracted from the local ROI image corresponding to the upper-level node through a specific object detection convolutional network; each branch node can also expand one or more branch nodes, and the upper-level branch node ROI image includes the lower-level branch node ROI image, thus generating a tree structure of upstream and downstream image analysis relationships; the terminal branch nodes are the image analysis tree T Each terminal branch node can be further divided into one or more key tissues or key anatomical landmarks. Each key tissue is detected by a specific object detection or image segmentation algorithm in the local ROI image corresponding to the terminal branch node. Each key anatomical landmark is detected by a specific key point detection algorithm in the local ROI image corresponding to the terminal branch node. The local ROI image is selected based on the principle that it must contain tissues with distinct image features. The purpose is to ensure that the object detection and key point detection algorithms have sufficient convolutional image features to achieve good recognition performance.

Referring also to Figure 6 , in an embodiment of the present invention, a pre-trained cranial ROI detection model is first used to detect a bounding box within a fetal cerebellum planar ultrasound image (P1) within the cranial ROI image I1, excluding other ultrasound images not related to the brain. Within the cranial ROI image I1, the anatomically distinct tissues are the septum pellucidum and cerebellum, while relatively difficult-to-detect tissues include the cerebellomedullary cisterna magna and the posterior cervical thickness. However, the cistern and posterior neck thickness are crucial biometric measurements for diagnosing fetal structural abnormalities. To improve the accuracy of convolutional network detection of these areas, the cerebellum, whose anatomical features are readily detectable, was combined with images of the posterior neck thickness edge to detect the posterior neck thickness ROI region I1,2. Similarly, because the location of the cistern is difficult to identify independently, two new methods were employed to improve its detection: First, through top-down layer-by-layer analysis, the search area for cistern identification was narrowed from the entire ultrasound image to the posterior neck thickness ROI region I1,2, significantly reducing the complexity of cistern identification. Next, the cerebellum, with its distinct anatomical features and easy-to-detect features, is detected within the cerebello-medullary ROI image I1,2,1, using the distinct posterior cranial boundary. By significantly narrowing the search target region to the cerebello-medullary ROI image I1,2,1, keypoint detection techniques are then used within the cerebello-medullary ROI image I1,2,1 to detect the cerebellar transverse diameter anatomical landmarks P1 and P2, as well as the cerebello-medullary diametrical anatomical landmarks P3 and P4. Similarly, the posterior cervical thickness anatomical landmarks P5 and P6 can be found simply within the region already narrowed within the posterior cervical thickness ROI image I1,2. The anatomical tissues of the septum pellucidum T1 and cerebellum T2 can be searched only within the septum pellucidum ROI image I1,1 and the cerebellomedullary cistern ROI image I1,2,1, respectively, which have a reduced search range.

Furthermore, referring to FIG. 7 , in an embodiment of the present invention, for a fetal cerebellum planar ultrasound image (P1), the four branch node ROI images I1, I1,1, I1,2, and I1,2,1 corresponding to the fetal ultrasound image analysis tree in FIG. 6 are shown in the figure. The branch node ROI images are selected based on the principle that the image features are obvious so that the corresponding four object detection models can accurately identify the upper, lower, left, and right frame positions of the ROI image boundary box. For example, I1 can refer to the upper and lower left and right contours of the fetal skull as the basis for I1 bounding box detection; the ROI image I1,1 can refer to the clear septum cavity and the bilateral anterior horns of the ventricles in the fetal brain for the basis for bounding box detection; the right side boundary of the ROI image I1,2 of the nuchal fold can be located based on the skin boundary on the outer side of the skull, but the left side and the upper and lower sides of the nuchal fold do not have sufficient image features to provide for the object detection model for positioning. Therefore, the present invention proposes to use the cerebellum with obvious features to provide the left side and the upper and lower side boundary reference positioning features of the ROI image I1,2. Therefore, the ROI image I1,2 of the nuchal fold is included in the cerebellum region to improve the nuchal fold. Similarly, in this ultrasound image example, although the right boundary of the cerebellomedullary cistern ROI image I1,2,1 has the occipital bone for boundary reference positioning, it also lacks image features for left-lateral and superior-inferior positioning. Therefore, the cerebellum region is also included in the cerebellomedullary cistern ROI image I1,2,1 to provide left-lateral and superior-inferior boundary reference positioning features. After completing the detection of the four branch node ROI images I1, I1,1, I1,2, and I1,2,1 from top to bottom, the search range for the identification of the anatomical tissues of the transparent septum T1 and cerebellum T2 is limited to the local areas of the transparent septum ROI image I1,1 and the cerebellomedullary cistern ROI image I1,2,1, respectively. This can significantly reduce the complexity of the detection problem and improve the efficiency of identifying these two anatomical tissues.

Referring also to Figure 8 , in an embodiment of the present invention, the top-down analysis and development process for a fetal cerebellum ultrasound image (P1) corresponding to the fetal ultrasound image analysis tree in Figure 6 is shown in Figure 8 . The leftmost figure shows the input fetal cerebellum ultrasound image (P1). The bounding box of the skull ROI image partial image I1, as determined by the pretrained skull ROI image detection model, is indicated by the dashed line. After the search range is narrowed to the ROI image partial image I1, the pretrained CSP ROI image detection model and the nuchal fold ROI image detection model are then used to perform detection on the CSP ROI image partial image I1,1 and the nuchal fold ROI image partial image I1,2, respectively. After the search range is narrowed to the I1,2 local image, the pre-trained cisterna magna ROI image detection model is used to perform local image detection of the cisterna magna ROI image I1,2,1. Finally, the anatomical tissue septum transparent T1 was detected from the septum transparent ROI image I1,1 using the pretrained septum transparent detection model; the anatomical tissue cerebellum T2 was detected from the cistern ROI image I1,2,1 using the pretrained cerebellum magna diameter and cerebellar transverse diameter key point detection models; the cerebellar transverse diameter anatomical landmarks P1 and P2, as well as the cerebellum magna diameter anatomical landmarks P3 and P4 were detected from the cistern ROI image I1,2,1; and the posterior cervical thickness anatomical landmarks P5 and P6 were detected from the local image of the posterior cervical thickness ROI image I1,2 using the pretrained posterior cervical thickness key point detection model.

In this embodiment of the present invention, each branch node in the fetal ultrasound image analysis tree represents a local ROI image. This local ROI image is automatically detected and extracted from the local ROI image corresponding to the upper-level node using the corresponding pre-trained object detection convolutional network model. The object detection convolutional network model can use, but is not limited to, the YOLO series object detection model, the R-CNN series object detection model, the SSD series object detection model, the RatinaNet series object detection model, or the EfficientDet object detection model.

In the embodiment of the present invention, each terminal branch node in the fetal ultrasound image analysis tree is the terminal branch node of the image analysis tree T, and each terminal branch node can be further divided into one or more key tissues or key anatomical landmarks. The key tissue detection module (20) automatically detects and captures each key tissue local image based on the corresponding pre-trained object detection convolutional network model or image segmentation algorithm. The object detection convolutional network model may use, but is not limited to, a YOLO series object detection model, an R-CNN series object detection model, an SSD series object detection model, a RatinaNet series object detection model, or an EfficientDet series object detection model; and the image segmentation algorithm may use, but is not limited to, a U-Net series object detection model, a DeepLab series object detection model, or a PSPNet series object detection model. The key anatomical landmark detection module (30) obtains each key anatomical landmark by detecting the local ROI image corresponding to the terminal branch node using the corresponding pre-trained key point detection model. Key point detection models may include, but are not limited to, DeepLabCut, HRNet, AlphaPose, CenterNet, YOLOv8, or Key.Net.

In another embodiment of the present invention, the biometrics module (40) calculates the Euclidean distance between the two paired anatomical landmarks Pi and Pj according to the key anatomical landmark detection module (30) to obtain biometric values. For example, the transverse diameter of the cerebellum is obtained by calculating the Euclidean distance between the anatomical landmarks P1 and P2; the diameter of the cerebellomedullary cisterna magna is obtained by calculating the Euclidean distance between the anatomical landmarks P3 and P4; and the posterior neck thickness is obtained by calculating the Euclidean distance between the anatomical landmarks P5 and P6.

In an embodiment of the present invention, the abnormal development judgment module (50) judges the risk of developmental abnormality based on whether key tissues are detected and the various biometric values calculated by the biometric module (40), and whether the various biometric values exceed the normal range of the biometric statistics. Finally, the fetal ultrasound examination report (100) is generated based on the estimated risk values of the various developmental abnormalities.

As can be seen from the aforementioned system architecture and description, existing fetal ultrasound image analysis technologies primarily classify the entire ultrasound image or directly use object detection technology to detect important anatomical features in fetal ultrasound images. However, anatomical features within ultrasound images often lack specificity due to the volumetric characteristics of the image, resulting in poor performance when directly using object detection for identification. The development of the main technology of the present invention has the following benefits and features, such as:

1. This invention proposes a fetal ultrasound image analysis tree that utilizes a top-down, step-by-step approach to reduce the complexity of subsequent analysis by gradually eliminating images irrelevant to the desired biological developmental characteristics. Ultimately, only the smallest local ROI image containing the biological developmental characteristics is retained, thereby improving the efficiency of identifying biological developmental characteristics.

2. The selection of local ROI images in this invention is based on the principle that they must include tissue with distinct imaging features. This ensures that object detection and keypoint detection algorithms have sufficient convolutional image features to achieve good recognition performance. For example, the cerebellum, whose anatomical features are clearly detectable, is combined with the posterior neck thickness edge image to detect the posterior neck thickness ROI image region.

3. Each terminal branch node in the fetal ultrasound image analysis tree of the present invention is the terminal branch node of the image analysis tree T. Each terminal branch node can be further divided into one or more key tissues or key anatomical landmarks. This allows for the effective detection of key tissues and key anatomical landmarks while minimizing complexity, rather than searching for key tissues and key anatomical landmarks within the entire fetal ultrasound image.

In summary, it can be understood that the present invention is a highly creative creation that not only effectively solves problems faced by practitioners but also significantly enhances efficacy. Furthermore, no identical or similar products have been created or publicly used in the same technical field, and this invention also exhibits enhanced efficacy. Therefore, the present invention meets the requirements of "novelty" and "advancedness" for invention patents, and therefore, an invention patent application is filed in accordance with the law.

10: Hierarchical Anatomical Region Detection Module 20: Key Tissue Detection Module 30: Key Anatomical Landmark Detection Module 40: Biometric Measurement Module 50: Developmental Abnormality Interpretation Module 100: Fetal Ultrasound Examination Report P1: Fetal Cerebellum Planar Ultrasound Image

Figure 1: Schematic diagram of the auxiliary identification system architecture of the present invention. Figure 2: Schematic diagram of the key biological developmental characteristics that need to be measured in fetal cerebellum planar ultrasound imaging according to an embodiment of the present invention, including cranial biparietal diameter, cerebellum biparietal diameter, cisterna magna, and nuchal fold, as well as their corresponding minimum bounding boxes (BBs). Figure 3 is a schematic diagram showing the precision-recall curves of various tissue types in fetal cerebellum planar ultrasound images with varying degrees of difficulty, according to an embodiment of the present invention. The brain and cerebellum have more distinct imaging features, making them easier to identify using the convolutional network, and their precision-recall curves are closer to the upper right corner. The cisterna magna and nuchal fold have less distinct imaging features, making them more difficult to identify using the convolutional network, and their precision-recall curves are farther from the upper right corner. Figure 4: A schematic diagram showing the F1 scores of various tissue types in fetal cerebellar planar ultrasound images with varying degrees of difficulty, according to an embodiment of the present invention. Figure 5: A typical structural diagram of the image parse tree for fetal ultrasound biometry proposed in the present invention. Figure 6: A schematic diagram showing the structural diagram of the fetal cerebellar planar ultrasound input image parse tree, according to an embodiment of the present invention. Figure 7: A schematic diagram showing the ultrasound image, region of interest (ROI), key tissues, and key anatomical landmarks corresponding to the fetal cerebellar planar ultrasound input image parse tree, according to an embodiment of the present invention. Figure 8: A schematic diagram illustrating the top-down analysis process of a planar ultrasound input image of the fetal cerebellum according to an embodiment of the present invention.

10: Hierarchical anatomical region detection module

20: Key Organization Detection Module

30: Key anatomical landmark detection module

40: Biometrics Module

50: Abnormal Development Interpretation Module

100: Fetal Ultrasound Examination Report

P1: Planar ultrasound image of the fetal cerebellum

Claims (5)

An auxiliary identification system for detecting fetal brain abnormalities using ultrasound imaging and artificial intelligence includes a hierarchical anatomical region detection module, a key tissue detection module, a key anatomical landmark detection module, a biometrics module, and an abnormal development interpretation module. The hierarchical anatomical region detection module detects a fetal cerebellum plane ultrasound image from top to bottom, captures a local region image, and selects one or more ROI images with obvious image features. The process of gradually reducing the local ROI image can be represented by a fetal ultrasound image analysis tree data structure to separate the ROI image layer by layer in a highly reliable and stable manner. The ROI image contains one or more key tissues, but the ROI image does not have to be completely aligned with the key tissue image. The definition of the ROI image needs to take into account the fact that there are enough reliable images in each ROI area when the convolution network is detecting the ROI image boundary. The biological anatomical features are provided to the convolutional network so that it can accurately and automatically detect the ROI image. The key tissue detection module is connected to the hierarchical anatomical region detection module. Its main function is to find all the key tissues in the local range of the ROI image captured by the hierarchical anatomical region detection module by using the object detection method. The key anatomical landmark detection module is connected to the key tissue detection module, which cooperates with the key tissue position information found to find the key tissues. A keypoint detection method locates all key anatomical landmarks within the local ROI image. A biometrics module, connected to the key anatomical landmark detection module, measures important biological developmental characteristics by measuring the distance between key anatomical landmarks or using image segmentation techniques. Furthermore, a developmental abnormality assessment module, connected to the biometrics module, uses the biological developmental characteristic values calculated by the biometrics module to determine whether a specific developmental abnormality risk condition has occurred, based on medical standards or machine learning methods. As described in claim 1, an auxiliary recognition system for detecting fetal brain abnormalities using artificial intelligence using ultrasound images, wherein the fetal ultrasound image analysis tree T is a tree structure, and the root node of the image analysis tree T represents the entire original fetal ultrasound input image; one or more branch nodes are expanded from the root node, each branch node represents a local ROI image, and this local ROI image is automatically detected and captured from the local ROI image corresponding to the upper-level node through a specific object detection convolutional network. Each branch node can also expand one or more branch nodes, and the upper-level branch node ROI image includes the lower-level branch node ROI image, thereby generating a tree structure of upstream and downstream image analysis relationships, and each branch node has the image analysis tree T at the end. One or more terminal branch nodes, each terminal branch node can be further divided into one or more key tissues or key anatomical landmarks, wherein each key tissue is detected from the local ROI image corresponding to the terminal branch node through a specific object detection or image segmentation algorithm, and each key anatomical landmark is detected from the local ROI image corresponding to the terminal branch node through a specific key point detection algorithm. As described in claim 2, an auxiliary identification system for detecting fetal brain abnormalities using ultrasound images with artificial intelligence, wherein the hierarchical anatomical region detection module, for a fetal cerebellum plane ultrasound input image, has a first-level node of a fetal ultrasound image analysis tree corresponding to a skull ROI image, and the second-level node of the skull ROI image expanded includes a transparent septum ROI image and a posterior cervical thickness ROI image, wherein the transparent septum ROI image includes a key tissue transparent septum T1, and the third layer node of the expanded posterior cervical thickness ROI image includes a cerebellomedullary cistern ROI image, and the expanded cerebellomedullary cistern ROI image includes a key tissue cerebellum T2, as well as three sets of key anatomical landmarks: cerebellum transverse diameter anatomical landmarks P1, P2, cerebellomedullary cistern diameter anatomical landmarks P3, P4, and posterior cervical thickness anatomical landmarks P5, P6. As described in claim 3, an auxiliary identification system for detecting fetal brain abnormalities using ultrasound imaging and artificial intelligence, wherein the hierarchical anatomical region detection module detects, in a planar ultrasound input image of the fetal cerebellum, the skull, cerebellum, and septum pellucidum with obvious image features, while the image features of the cerebellomedullary cistern, the posterior neck thickness, and six anatomical landmarks closely related to biological developmental characteristics, including the cerebellum transverse diameter anatomical landmarks P1 and P2, the cerebellomedullary cistern diameter anatomical landmarks P3 and P4, and the posterior neck thickness anatomical landmark P5. P6 uses the skull as the most prominent local image feature, first performing skull ROI local image identification. Excluding images outside the skull reduces the complexity of subsequent analysis, limiting the scope of subsequent analysis to the skull ROI image. The next step is to use the cerebellum, a local image with distinct features, combined with the posterior neck skin and occipital bone, with distinct boundary features, to combine these two image features to identify the posterior neck thickness ROI image and the cerebellomedullary cisterna magna ROI image, thereby improving the accuracy of local image identification with less prominent features. As described in claim 4, an auxiliary identification system for detecting fetal brain abnormalities using ultrasound images with artificial intelligence, wherein the biometrics module calculates the cerebellum transverse diameter by the distance between the aforementioned automatic cerebellum transverse diameter anatomical landmarks P1 and P2, and further verifies the reliability of the measurement data of P1, P2 and the cerebellum transverse diameter based on the detected boundary frame position of the key tissue cerebellum T2, and the biometrics module also calculates the cerebellum transverse diameter by the distance between the aforementioned automatic cerebellomedullary cisterna magna anatomical landmarks P3 and P4. The diameter of the cerebellomedullary cistern is measured, and the reliability of the measurement data of P3, P4, and the cerebellomedullary cistern diameter is further verified based on the detected cerebellomedullary cistern ROI image and the bounding box position of the key cerebellar tissue T2. The biometrics module also calculates the posterior neck thickness using the distance between the aforementioned posterior neck thickness anatomical landmarks P5 and P6. The reliability of the measurement data of P5, P6, and the posterior neck thickness is further verified based on the bounding box position of the detected cerebellomedullary cistern ROI image and the posterior neck thickness ROI image.