A method and system for transverse cerebellar mass assessment and biometry

Abstract

The application discloses a transverse cerebellum section quality evaluation and biological measurement method and system, and belongs to the technical field of medical image analysis and processing. The method comprises the following steps: acquiring a transverse cerebellum section sample in a fetal ultrasound image, performing data preprocessing, constructing a target detection and semantic segmentation dataset, training a target detection and semantic segmentation network as a prediction model, and building a corresponding quality evaluation system and biological index post-processing system according to the output result of the prediction model. The prediction model and the evaluation processing system are effectively combined together to realize end-to-end evaluation and measurement of the transverse cerebellum section ultrasound image. The application combines the section quality evaluation and the biological index measurement on the transverse cerebellum section for the first time, establishes a scientific and objective quantitative evaluation system and a fast and accurate biological index measurement method, and facilitates the judgment of the section quality by doctors.

Description

Technical Field

[0001] This invention relates to a method and system for quality assessment and biometry via cerebellar transverse sections, belonging to the field of medical image analysis and processing technology. Background Technology

[0002] With the rapid development of deep learning, there are increasingly more methods for analyzing and processing medical images. Ultrasound image quality assessment is not only a crucial prerequisite for constructing medical image datasets but also plays a vital role in doctors' diagnoses. Traditional medical image quality assessment requires extensive clinical experience and substantial human resources, posing a significant challenge to regions lacking medical resources. Furthermore, medical images acquired without a proper quality assessment process are prone to misinterpretation of patient symptoms, leading to incalculable consequences. In fetal ultrasound images, the transcerebellar cross-section is a crucial basis for doctors to assess the development of the fetal cerebellum. Accurate and reasonable assessment of the sections acquired by the instrument is essential for subsequent diagnosis. Therefore, it is imperative to propose a scientific and objective method for cross-sectional quality assessment.

[0003] Deep neural networks are currently the most popular image processing technology, demonstrating good performance in medical images with indistinct features, high noise levels, and complex structures. Object detection is a crucial research direction within computer vision using deep learning, aiming to automatically identify and locate targets or lesion regions of interest in medical images. This task has wide applications in medical image analysis, particularly in disease screening, diagnosis, and treatment planning. Target regions in medical images often show little difference from surrounding tissues, thus requiring object detection models with very high accuracy to ensure reliable diagnostic results. Medical image annotation typically requires specialized medical personnel and is time-consuming and labor-intensive, resulting in limited labeled data. However, object detection in medical images can significantly assist clinicians in their diagnostic work, especially those with less experience, improving their efficiency and accuracy. This field continues to develop with advancements in deep learning technology. However, target detection alone cannot directly provide a quantitative assessment of the cross-sectional quality. Establishing an objective and scientific evaluation system that links target detection results with the quality of transcerebellar cross-sections in fetal ultrasound images to achieve end-to-end cross-sectional quality assessment is a key focus of research. Furthermore, developing a method for measuring fetal biomarkers based on cross-sectional quality assessment results is crucial for assisting physicians in diagnosis.

[0004] Therefore, there is a need for a method that can assess quality on transcerebellar cross sections of fetal ultrasound images and perform segmentation and biometric measurements based on the assessment results. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method for quality assessment and biometrics via cerebellar transverse sections. By linking section quality assessment with biometric measurements, an evaluation system is established, which enhances the scientific nature and interpretability of quality assessment and biometric measurements. To achieve the above objectives / to solve the above technical problems, the present invention is implemented using the following technical solution:

[0006] First aspect: A method for cerebellar transverse quality assessment and biometrics, said method comprising: The transcerebellar cross-sectional image of the fetus in the ultrasound image to be tested is input into a pre-trained target detection network model, and the output is an ultrasound image with target detection boxes and corresponding confidence scores. The target detection network model is based on the YOLOv11 model structure, which is an improvement on YOLOv5 and YOLOv8. Ultrasound images with target detection boxes and corresponding confidence levels are evaluated through a quality assessment system to identify ultrasound images with acceptable cross-sectional quality. An ultrasound image with acceptable cross-sectional quality is input into a pre-trained segmentation model to obtain segmentation masks for the cerebellum and cisterna magna. The objective function of the segmentation model is a combination of Dice loss and cross-entropy loss. The segmentation masks of the cerebellum and cisterna magna were processed by a post-processing system to obtain two key biometric indicators from the transverse section of the cerebellum.

[0007] Optionally, the training method for the object detection network model includes: Obtain transcerebellar cross-sectional samples from fetal ultrasound images and construct a target detection dataset; Perform data preprocessing on the target detection dataset; Based on the preprocessed object detection dataset, with the goal of minimizing the value of the objective function, a trained object detection network model is obtained according to the final weight parameters of the encoder and decoder.

[0008] Optionally, the step of acquiring transcerebellar cross-sectional samples from fetal ultrasound images and constructing a target detection dataset includes: Based on professional books and expert opinions, the key tissues in the transcerebellar cross section were studied to identify five target tissues required for the section: the strong echo ring of the skull, the cavum septum pellucidum, the thalamus, the cerebellum, and the cisterna magna. Based on the target tissue, target bounding boxes are drawn in the cross-sectional samples, and the original image is matched one-to-one with the image containing the target bounding box to construct an object detection dataset, which is used as the training set.

[0009] Optionally, the target detection network model includes an input layer for data preprocessing and data augmentation, a backbone network for feature extraction, an intermediate layer composed of convolutional neural networks that fuses multi-level information, an output layer for helping the model identify targets at different scales, and an output layer for predicting target bounding boxes, classification, and confidence.

[0010] Optionally, the quality assessment system is constructed based on the confidence results of each part of the target detection results obtained by the target detection network model through cerebellar cross-section processing, and combined with medical prior knowledge to establish a cross-section scoring mechanism.

[0011] Optionally, the section scoring mechanism is as follows: (1) Within each cross section, there is one and only one target detection result for each target organization; (2) The relative positions between different target tissues are fixed. The strong echo ring of the skull needs to include four tissues: cavum septum pellucidum, thalamus, cerebellum and cisterna magna. The four tissues are arranged in sequence along the midline of the brain and maintain a symmetrical relationship with the midline of the brain. The final section score is obtained based on the confidence of the five target tissues, and the cerebellar transverse section is quantitatively evaluated.

[0012] Optionally, the section scoring mechanism uses a 10-point scale for scoring. The initial score for each target organization is two points, which is then multiplied by the confidence level of the target bounding box of the corresponding organization in the target detection result to obtain the final score of the organization. The scores of the five target organizations are then summed to obtain the quantitative result of the quality assessment of the section.

[0013] Optionally, the post-processing system is established based on the segmentation mask of the cerebellum and cisterna magna in the segmentation result obtained by processing the cerebellum cross section using the segmentation model, combined with prior medical knowledge.

[0014] Optionally, the steps to establish a post-processing system are as follows: (1) Based on the segmentation mask of the cerebellum, the algorithm is first used to calculate the convex hull of the region, and the distance between the farthest point pairs on the convex hull is calculated by the rotation caliper method as the measurement value of the transverse diameter of the cerebellum. (2) Based on the cerebellum segmentation mask, fit the corresponding ellipse, and take the direction of the minor axis of the ellipse as a straight line. Then, based on the segmentation mask of the cisterna magna, obtain the distance between the intersection of the extension of the minor axis of the ellipse and the segmentation area of ​​the cisterna magna, and use it as the width of the cisterna magna. Thus, the two key biometric indicators of the cerebellum transverse diameter and the width of the cisterna magna are obtained through the cerebellum transverse section.

[0015] Second aspect: A cerebellar transverse section quality assessment and biometric system, said system comprising: An object detection module is configured to input a transcerebellar cross-sectional image from the ultrasound image of the fetus to be tested into a pre-trained object detection network model, and output an ultrasound image with object detection boxes and corresponding confidence scores, wherein the object detection network model is based on the YOLOv11 model structure, which is an improvement on YOLOv5 and YOLOv8. The quality assessment module is configured to evaluate ultrasound images with target detection boxes and corresponding confidence levels through a quality assessment system to identify ultrasound images with acceptable cross-sectional quality. A semantic segmentation module is configured to input ultrasound images with acceptable cross-sectional quality into a pre-trained segmentation model to obtain segmentation masks for the cerebellum and cisterna magna. The objective function of the segmentation model adopts a combination of Dice loss and cross-entropy loss. A biometric module is configured to process the segmentation masks of the cerebellum and cisterna magna through a post-processing system to obtain two key biometric indicators in the transcerebellar cross section.

[0016] Compared with the prior art, the beneficial effects achieved by the present invention are as follows: This invention is the first to combine transcerebellar transverse section quality assessment with biometric measurement, establishing a scientific and objective quantitative evaluation system and a rapid and accurate biometric measurement method, facilitating physicians' judgment of section quality. It features high precision and high efficiency, effectively addressing the challenges of high noise and complex structures in medical images, providing reliable results for subsequent quality assessment. It is also highly adaptable, easily extendable to quality assessment and biometric measurement tasks in other medical ultrasound images. Compared with the prior art, the target detection model and semantic segmentation model adopted in this invention have the characteristics of high precision and high efficiency, which effectively solves the difficulties of high noise and complex structure in medical images, and provides reliable results for subsequent quality assessment. This invention is highly adaptable and can effectively solve the quality assessment problem of transcerebellar cross-section in fetal ultrasound images. It can also be easily extended to other medical ultrasound image quality assessment tasks. Attached Figure Description

[0017] Figure 1 The diagram shown is a flowchart of the cerebellar cross-section quality assessment and biometric method of the present invention. Figure 2 The diagram shown is a schematic representation of the method flow in a specific embodiment of the present invention. Detailed Implementation

[0018] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.

[0019] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, are used only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, the terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined with "first," "second," etc., may explicitly or implicitly include one or more of that feature. In the description of this invention, unless otherwise stated, "a plurality of" means two or more.

[0020] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art will understand the specific meaning of the above terms in this invention based on the specific circumstances. Example

[0021] See Figures 1-2 This paper presents a method for quality assessment and biometrics of cerebellar transverse sections. This method combines object detection and semantic segmentation techniques from deep learning to establish a scientific and objective quality assessment annotation system. Based on the assessment results, segmentation and post-processing are performed to help doctors evaluate the sections and measure biometric indicators, thereby assisting doctors in subsequent diagnostic work. The steps include the following: Step 1: Obtain transcerebellar cross-sectional samples from fetal ultrasound images, construct a target detection dataset, perform data preprocessing, select a deep neural network model and objective function for target detection, and obtain a trained target detection network model.

[0022] Step 2: Input the transcerebellar cross-sectional image of the fetus in the ultrasound image to be tested into the pre-trained target detection network model, and output the ultrasound image with target detection boxes and corresponding confidence scores. The target detection network model is based on the YOLOv11 model structure, which is an improvement of YOLOv5 and YOLOv8. Step 3: The ultrasound images with target detection boxes and corresponding confidence levels are evaluated through a quality assessment system to identify ultrasound images with acceptable cross-sectional quality. Step 4: Take transcerebellar cross-section samples from fetal ultrasound images, construct a semantic segmentation dataset, perform data preprocessing, select a deep neural network model and objective function for semantic segmentation, and obtain a trained segmentation model. Step 5: Input the ultrasound image with acceptable cross-sectional quality into the pre-trained segmentation model to obtain the segmentation mask of the cerebellum and cisterna magna. The objective function of the segmentation model adopts a combination of Dice loss and cross-entropy loss. Step 6: The segmentation mask of the cerebellum and cisterna magna is processed by the post-processing system to obtain two key biometric indicators of the cerebellum transverse section.

[0023] In the specific implementation of this embodiment, step 1: Select a certain proportion of transcerebellar cross-sectional samples from fetal ultrasound images. First, based on prior medical knowledge, target boxes are drawn for key tissues (strong echo ring of the skull, cavum septum pellucidum, thalamus, cerebellum, and cisterna magna) in this batch of samples. If the target tissue is not fully displayed or is very blurry on the cross-section, it is not necessary to draw the corresponding target box. After processing, the preprocessed training data is obtained.

[0024] The object detection model is constructed based on the YOLOv11 model structure, which is an improvement on YOLOv5 and YOLOv8. Its core structure can be divided into the following parts: an input layer for data preprocessing and data augmentation; a backbone network for feature extraction, which is composed of convolutional neural networks; an intermediate layer that further integrates multi-level information to help the model identify targets at different scales; and an output layer for predicting target bounding boxes, classification, and confidence scores.

[0025] The objective function is a combined loss function designed to optimize three key tasks: bounding box regression, classification, and confidence prediction. Through these optimizations in structure and objective function, YOLOv11 has become a state-of-the-art model in object detection tasks, combining high accuracy and high performance.

[0026] The training methods for the object detection network model described in step 1 include: Obtain transcerebellar cross-sectional samples from fetal ultrasound images and construct a target detection dataset; Perform data preprocessing on the target detection dataset; Based on the preprocessed object detection dataset, with the goal of minimizing the value of the objective function, a trained object detection network model is obtained according to the final weight parameters of the encoder and decoder.

[0027] The data preprocessing workflow for the object detection dataset is as follows: First, based on professional books and expert opinions, the key tissues in the transcerebellar cross-section are studied, and it is determined that the tissues required for the cross-section are mainly the strong echo ring of the skull, the cavum septum pellucidum, the thalamus, the cerebellum, and the cisterna magnum. Then, target bounding boxes are drawn in the cross-section samples according to the target tissues, and the original image is matched one-to-one with the image containing the target bounding box to construct the training set.

[0028] In this embodiment, for step 2, based on the dataset, network structure, and objective function of step 1, a labeled feature extraction network is trained. The training process aims to minimize the value of the objective function. The trained network model is obtained based on the final weight parameters of the discriminator and generator. The trained network model is then used as the final labeled feature extraction model. In this embodiment, step 3 is further explained: Based on the pre-trained model obtained in step 2, the target detection results of the cerebellar transverse section are obtained. First, the target tissue is judged according to the following principles: (1) Within each cross section, there is one and only one target detection result for each target organization; (2) The relative positions of different target tissues are fixed. The strong echo ring of the skull must include the other four target tissues: the cavum septum pellucidum, thalamus, cerebellum, and cisterna magnum. These four tissues must be arranged sequentially along the midline of the brain and maintain a symmetrical relationship with the midline. If the target tissues meet the requirements, the quality of the section is further quantitatively evaluated using a 10-point scoring system. Each target tissue starts with two points, which are then multiplied by the confidence level of the corresponding target box in the target detection results to obtain the final score for that tissue. Finally, the scores of the five target tissues are summed to obtain the quantitative result of the quality evaluation of the section.

[0029] In this embodiment, step 4 is further explained as follows: Transcerebellar cross-sectional samples from fetal ultrasound images are taken to construct a semantic segmentation dataset. Data preprocessing is performed, and key tissues in the transcerebellar cross-section are learned based on professional books and expert opinions. It is determined that the tissues required for biometric measurements are mainly the cerebellum and the cisterna magna. Then, the cross-sectional samples are segmented according to the target tissues to construct a training set. A deep neural network model and objective function are selected for semantic segmentation. The overall network structure continues the classic U-Net framework, including an encoder (downsampling) and a decoder (upsampling), as well as skip connections. The objective function is a combination of Dice loss and cross-entropy loss.

[0030] In this embodiment, regarding step 5: based on the dataset, network structure, and objective function from step 4, a labeled feature extraction network is trained. The training process aims to minimize the value of the objective function, and the trained network model is obtained based on the final weight parameters of the discriminator and generator. The trained network model is then used as the final semantic segmentation and extraction model. In this embodiment, regarding step 6: constructing a post-processing system to obtain biometric indicators using the segmentation results, the main approach is to establish a post-processing system based on the segmentation mask of the cerebellum and cisterna magna in the segmentation results obtained from the cerebellar cross-section processing, using the model described above, and in conjunction with the medical prior knowledge provided in step 3. The steps are as follows: (1) Based on the segmentation mask of the cerebellum, the algorithm is first used to calculate the convex hull of the region, and the distance between the farthest point pair on the convex hull is calculated by the rotation caliper method as the measurement value of the transverse diameter of the cerebellum.

[0031] (2) Based on the segmentation mask of the cerebellum, fit the corresponding ellipse, and draw a straight line along the minor axis of the ellipse. Then, based on the segmentation mask of the cisterna magna, obtain the distance between the intersection point of the extension of the minor axis of the ellipse and the segmented region of the cisterna magna, and use this distance as the width of the cisterna magna. In this way, we obtain two key biometric indicators through the transverse section of the cerebellum, namely the transverse diameter of the cerebellum and the width of the cisterna magna.

[0032] The computer language used in the embodiments of this invention is Python, and the development framework is PyTorch. The basic technologies such as data reading and writing, data operation, and network training are well-known technologies in this field and will not be described in detail here. Example

[0033] This invention also discloses a cerebellar transverse section quality assessment and biometric system, the system comprising: An object detection module is configured to input a transcerebellar cross-sectional image from the ultrasound image of the fetus to be tested into a pre-trained object detection network model, and output an ultrasound image with object detection boxes and corresponding confidence scores, wherein the object detection network model is based on the YOLOv11 model structure, which is an improvement on YOLOv5 and YOLOv8. The quality assessment module is configured to evaluate ultrasound images with target detection boxes and corresponding confidence levels through a quality assessment system to identify ultrasound images with acceptable cross-sectional quality. A semantic segmentation module is configured to input ultrasound images with acceptable cross-sectional quality into a pre-trained segmentation model to obtain segmentation masks for the cerebellum and cisterna magna. The objective function of the segmentation model adopts a combination of Dice loss and cross-entropy loss. A biometric module is configured to process the segmentation masks of the cerebellum and cisterna magna through a post-processing system to obtain two key biometric indicators in the transcerebellar cross section.

[0034] This invention is the first to combine transcerebellar transverse section quality assessment with biometric measurement, establishing a scientific and objective quantitative evaluation system and a rapid and accurate biometric measurement method, facilitating physicians' judgment of section quality. It features high precision and high efficiency, effectively addressing the challenges of high noise and complex structures in medical images, providing reliable results for subsequent quality assessment. It is also highly adaptable, easily extendable to quality assessment and biometric measurement tasks in other medical ultrasound images.

[0035] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for quality assessment and biometrics via cerebellar transverse sections, characterized in that, The method includes: The transcerebellar cross-sectional image of the fetus in the ultrasound image to be tested is input into a pre-trained target detection network model, and the output is an ultrasound image with target detection boxes and corresponding confidence scores. The target detection network model is based on the YOLOv11 model structure, which is an improvement on YOLOv5 and YOLOv8. Ultrasound images with target detection boxes and corresponding confidence levels are evaluated through a quality assessment system to identify ultrasound images with acceptable cross-sectional quality. An ultrasound image with acceptable cross-sectional quality is input into a pre-trained segmentation model to obtain segmentation masks for the cerebellum and cisterna magna. The objective function of the segmentation model is a combination of Dice loss and cross-entropy loss. The segmentation masks of the cerebellum and cisterna magna are processed by a post-processing system to obtain two key biometric indicators from the transverse section of the cerebellum. The quality assessment system is constructed based on the confidence scores of each part of the target detection results obtained by cerebellar cross-section processing using a target detection network model, and combined with prior medical knowledge, to establish a cross-section scoring mechanism, which is as follows: (1) Within each cross section, there is one and only one target detection result for each target organization; (2) The relative positions between different target tissues are fixed. The strong echo ring of the skull needs to include four tissues: cavum septum pellucidum, thalamus, cerebellum and cisterna magna. The four tissues are arranged along the midline of the brain and maintain a symmetrical relationship with the midline of the brain. The final section score is obtained based on the confidence of the five target tissues, and the cerebellar transverse section is quantitatively evaluated. The post-processing system handles the following: (1) Based on the segmentation mask of the cerebellum, the algorithm is first used to calculate the convex hull of the region, and the distance between the farthest point pairs on the convex hull is calculated by the rotation caliper method as the measurement value of the transverse diameter of the cerebellum. (2) Based on the cerebellum segmentation mask, fit the corresponding ellipse, and take the direction of the minor axis of the ellipse as a straight line. Then, based on the segmentation mask of the cisterna magna, obtain the distance between the intersection of the extension of the minor axis of the ellipse and the segmentation area of ​​the cisterna magna, and use it as the width of the cisterna magna. Thus, the two key biometric indicators of the cerebellum transverse diameter and the width of the cisterna magna are obtained through the cerebellum transverse section.

2. The method for cerebellar transverse section quality assessment and biometrics according to claim 1, characterized in that, The training method for the object detection network model includes: Obtain transcerebellar cross-sectional samples from fetal ultrasound images and construct a target detection dataset; Perform data preprocessing on the target detection dataset; Based on the preprocessed object detection dataset, with the goal of minimizing the value of the objective function, a trained object detection network model is obtained according to the final weight parameters of the encoder and decoder.

3. The method for cerebellar transverse section quality assessment and biometrics according to claim 2, characterized in that, The process of acquiring transcerebellar cross-sectional samples from fetal ultrasound images and constructing a target detection dataset includes: Based on professional books and expert opinions, the key tissues in the transcerebellar cross section were studied to identify five target tissues required for the section: the strong echo ring of the skull, the cavum septum pellucidum, the thalamus, the cerebellum, and the cisterna magna. Based on the target tissue, target bounding boxes are drawn in the cross-sectional samples, and the original image is matched one-to-one with the image containing the target bounding box to construct an object detection dataset, which is used as the training set.

4. The method for cerebellar transverse section quality assessment and biometrics according to claim 1, characterized in that, The target detection network model includes an input layer for data preprocessing and data augmentation, a backbone network for feature extraction, an intermediate layer composed of convolutional neural networks that integrates multi-level information to help the model identify targets at different scales, and an output layer for predicting target bounding boxes, classification, and confidence.

5. The method for cerebellar transverse section quality assessment and biometrics according to claim 1, characterized in that, The section scoring mechanism uses a 10-point scale. Each target organization starts with a score of 2 points, which is multiplied by the confidence level of the target bounding box of the corresponding organization in the target detection result to obtain the final score of the organization. The scores of the five target organizations are then summed to obtain the quantitative result of the quality assessment of the section.

6. The method for cerebellar transverse section quality assessment and biometrics according to claim 1, characterized in that, The post-processing system is established based on the segmentation mask of the cerebellum and cisterna magna in the segmentation result obtained by processing the cerebellum transverse section using the segmentation model, combined with medical prior knowledge.

7. A cerebellar transverse section quality assessment and biometric system, characterized in that, The system includes: An object detection module is configured to input a transcerebellar cross-sectional image from the ultrasound image of the fetus to be tested into a pre-trained object detection network model, and output an ultrasound image with object detection boxes and corresponding confidence scores, wherein the object detection network model is based on the YOLOv11 model structure, which is an improvement on YOLOv5 and YOLOv8. A quality assessment module is configured to evaluate ultrasound images with target detection boxes and corresponding confidence levels through a quality assessment system, identifying ultrasound images with acceptable cross-sectional quality. The quality assessment system is constructed based on the confidence levels of various parts of the target detection results obtained through cerebellar cross-section processing using a target detection network model, and incorporates prior medical knowledge to establish a cross-sectional scoring mechanism. The cross-sectional scoring mechanism is as follows: (1) Within each cross section, there is one and only one target detection result for each target organization; (2) The relative positions between different target tissues are fixed. The strong echo ring of the skull needs to include four tissues: cavum septum pellucidum, thalamus, cerebellum and cisterna magna. The four tissues are arranged along the midline of the brain and maintain a symmetrical relationship with the midline of the brain. The final section score is obtained based on the confidence of the five target tissues, and the cerebellar transverse section is quantitatively evaluated. A semantic segmentation module is configured to input ultrasound images with acceptable cross-sectional quality into a pre-trained segmentation model to obtain segmentation masks for the cerebellum and cisterna magna. The objective function of the segmentation model adopts a combination of Dice loss and cross-entropy loss. The biometric module is configured to process the segmentation masks of the cerebellum and cisterna magna through a post-processing system to obtain two key biometric indicators from the cerebellar transverse section. The post-processing system includes: (1) Based on the segmentation mask of the cerebellum, the algorithm is first used to calculate the convex hull of the region, and the distance between the farthest point pairs on the convex hull is calculated by the rotation caliper method as the measurement value of the transverse diameter of the cerebellum. (2) Based on the cerebellum segmentation mask, fit the corresponding ellipse, and take the direction of the minor axis of the ellipse as a straight line. Then, based on the segmentation mask of the cisterna magna, obtain the distance between the intersection of the extension of the minor axis of the ellipse and the segmentation area of ​​the cisterna magna, and use it as the width of the cisterna magna. Thus, the two key biometric indicators of the cerebellum transverse diameter and the width of the cisterna magna are obtained through the cerebellum transverse section.

Images