CN-122000031-A - Semi-supervised learning-based labor ultrasonic quantitative measurement method and system

CN122000031ACN 122000031 ACN122000031 ACN 122000031ACN-122000031-A

Abstract

The invention provides a labor ultrasonic quantitative measurement method and system based on semi-supervised learning, wherein the method comprises the steps of constructing an image subset with pixel-level segmentation labels, an image subset with key point labels and an unlabeled image subset based on a perineum ultrasonic image; the method comprises the steps of obtaining a labor ultrasonic quantitative measurement model by combining a medical image segmentation network with a pixel level segmentation marked image subset pre-trained with morphology and a geometric algorithm, performing full-supervision training by using the image subset with a key point mark, constructing a plurality of semi-supervision labor ultrasonic quantitative measurement models based on the labor ultrasonic quantitative measurement model after full-supervision training, performing iterative training by using the image subset with the pixel level segmentation marked image subset and an unlabeled image subset and adopting a sequential cross model strategy, inputting the to-be-detected perineal ultrasonic image into the final semi-supervision labor ultrasonic quantitative measurement model, and outputting a predicted labor ultrasonic quantitative measurement result.

Inventors

DU BO
ZHANG XUEZHI
Tong Lvyang

Assignees

武汉大学

Dates

Publication Date: 20260508
Application Date: 20260127

Claims (10)

1. A labor ultrasonic quantitative measurement method based on semi-supervised learning is characterized by comprising the following steps: Constructing an image subset with pixel-level segmentation labels, an image subset with key point labels and an unlabeled image subset based on the perineum ultrasonic image; Pre-training a medical image segmentation network by using an image subset with pixel level segmentation labels, combining morphology and geometric algorithm of the medical image segmentation network after pre-training to obtain a labor ultrasonic quantitative measurement model, and performing full supervision training on the labor ultrasonic quantitative measurement model by using the image subset with key point labels; Constructing a plurality of semi-supervised labor ultrasonic quantitative measurement models based on the labor ultrasonic quantitative measurement models after the full supervision training, and using an image subset with pixel level segmentation labels and an image subset without labels, adopting a sequential cross model strategy for iterative training, and outputting a final semi-supervised labor ultrasonic quantitative measurement model after the training is finished; And inputting the perineum ultrasonic image to be detected into a final semi-supervised labor ultrasonic quantitative measurement model, and outputting a predicted fetal head and pubic bone combined compact segmentation mask, and marking key points and quantifying clinical indexes of labor.
2. The semi-supervised learning based labor ultrasound quantitative measurement method of claim 1, wherein the process of performing full supervision training on the labor ultrasound quantitative measurement model using the subset of images with the keypoint labels comprises: Inputting each image of the image subset with the key point marks into a pre-trained medical image segmentation network, and outputting an initial fetal head and pubic symphysis segmentation mask of each image of the image subset with the key point marks; Performing hole filling, connected domain denoising and boundary smoothing post-processing on the initial fetal head and pubic symphysis mask, so as to obtain a fetal head and pubic symphysis dense segmentation mask of each image in the image subset with the key point mark; Based on morphology and geometric algorithm, extracting predicted key points of the fetal head and pubic bone joint dense segmentation mask of each image in the image subset with key point marks, and taking the fetal head and pubic bone joint dense segmentation mask of each image in the image subset with key point marks and the predicted key points thereof as training samples in a full supervision stage; and calculating the self-adaptive sample weight of each training sample in the full supervision stage based on the distance error between the predicted key point and the marked key point, training the labor ultrasonic quantitative measurement model by using each training sample weighted by the self-adaptive sample weight, and outputting the trained labor ultrasonic quantitative measurement model.
3. The method for quantitative measurement of labor ultrasound based on semi-supervised learning of claim 2, wherein the steps of extracting key points in the dense segmentation mask of the fetal head and pubic symphysis based on morphology and geometry algorithm comprise: Carrying out connected domain analysis on a pubic symphysis mask region in the fetal head and pubic symphysis compact segmentation mask, extracting a contour point set of the maximum connected region, calculating Euclidean distances between all point pairs on the contour, and selecting the point pair with the maximum Euclidean distance as an upper pubic symphysis endpoint and a lower pubic symphysis endpoint respectively; Extracting the contour of a fetal head mask region in the fetal head and pubic symphysis compact segmentation mask, and searching a fetal head tangent point and a fetal head perpendicular foot point from the contour, wherein the fetal head tangent point is satisfied that a connecting line from the lower end point of the pubis symphysis to the fetal head tangent point is perpendicular to a normal vector of the fetal head contour at the fetal head tangent point.
4. The method for quantitative measurement of labor ultrasound based on semi-supervised learning as set forth in claim 3, wherein the calculation formula of the adaptive sample weights of each training sample in the fully supervised stage is as follows: ; in the formula, Sample weights representing the ith training sample; Is the lower sample weight limit; Predicting Euclidean distance errors of the key points and the marked key points for the ith training sample, Is the median of the global euclidean distance error, Is a half-quarter bit difference.
5. The method for quantitative measurement of labour ultrasonic based on semi-supervised learning as set forth in claim 1, wherein the step of iterative training using a subset of images with pixel-level segmentation labels and a subset of images without labels and using a sequential cross model strategy, the training ending outputting a final semi-supervised labour ultrasonic quantitative measurement model comprises: The method comprises the steps of grouping unlabeled image sets, alternately serving as training image subsets of a semi-supervised iterative training process, taking the image subsets with pixel-level segmentation labels as golden standard data of the semi-supervised iterative training process, starting training, outputting a semi-supervised labor ultrasonic quantitative measurement model of each round of iterative process, stopping training when the semi-supervised iterative training reaches a preset iterative round or a preset training target, and outputting a current semi-supervised labor ultrasonic quantitative measurement model of a final round as a final semi-supervised labor ultrasonic quantitative measurement model after training is finished.
6. The semi-supervised learning based labor ultrasound quantitative measurement method of claim 5, wherein the semi-supervised iterative training process comprises: In the first round of iteration, the fetal head and pubic bone joint compact segmentation mask obtained by processing the first semi-supervised labor ultrasonic quantitative measurement model by the training image subset of the first round of iteration is used as a pseudo tag of the first round of iteration process, unified weight is given, the training image subset of the first round of iteration process, the weighted pseudo tag and gold standard data are combined to train the second semi-supervised labor ultrasonic quantitative measurement model, and the trained second semi-supervised labor ultrasonic quantitative measurement model is used as the current semi-supervised labor ultrasonic quantitative measurement model of the first round of iteration process; The method comprises the steps of processing a current round training image subset by using the initial reference model and the pseudo-label generation model of the second round, respectively obtaining an output fetal head and pubic symphysis compact segmentation mask and key point labels, calculating pseudo-label weights of the second round iteration process by using a cross consistency assessment mechanism, reserving the fetal head and pubic symphysis compact segmentation mask output by the pseudo-label generation model of the second round as a pseudo-label of the second round iteration process, combining the training image subset of the second round iteration process, the weighted pseudo-label and gold standard data to train a third semi-supervised labour ultrasonic quantitative measurement model, and outputting the trained third semi-supervised labour ultrasonic quantitative measurement model as the current semi-supervised labour ultrasonic quantitative measurement model of the second round; And in the subsequent iteration process, taking the current semi-supervised labor ultrasonic quantitative measurement model of the current round and the previous two rounds as an initial reference model of the current round, taking the current semi-supervised labor ultrasonic quantitative measurement model of the current round and the previous round as a pseudo tag generation model of the current round, repeating the steps of the second round of iteration process, and outputting the current semi-supervised labor ultrasonic quantitative measurement model of the current round.
7. The method for quantitative measurement of labor ultrasound based on semi-supervised learning of claim 6, wherein the pseudo tag weights are dynamic confidence weights calculated based on consistency scores of output results of the same unlabeled image processed by an initial reference model and a pseudo tag generation model of a current iteration round number.
8. The utility model provides a labor ultrasonic quantitative measurement system based on semi-supervised learning which characterized in that includes: The training data acquisition module is used for constructing an image subset with pixel-level segmentation labels, an image subset with key point labels and an unlabeled image subset based on the perineum ultrasonic image; The full-supervision training module is used for pre-training the medical image segmentation network by using an image subset with pixel level segmentation labels, combining morphology and geometric algorithm with the pre-trained medical image segmentation network to obtain a labor ultrasonic quantitative measurement model, and performing full-supervision training on the labor ultrasonic quantitative measurement model by using an image subset with key point labels; the semi-supervised training module is used for constructing a plurality of semi-supervised labor ultrasonic quantitative measurement models based on the labor ultrasonic quantitative measurement models after full-supervision training, and using an image subset with pixel-level segmentation labels and an unlabeled image subset to perform iterative training by adopting a sequential cross model strategy, and outputting a final semi-supervised labor ultrasonic quantitative measurement model after training; And the labor ultrasonic quantitative measurement module is used for inputting the to-be-measured perineum ultrasonic image into a final semi-supervised labor ultrasonic quantitative measurement model, and outputting a predicted fetal head and pubic bone combined compact segmentation mask, a key point mark and a labor quantitative clinical index.
9. An electronic device comprising a memory and a processor, the memory storing program instructions for execution by the processor, the processor invoking the program instructions to perform the method of any of claims 1-7.
10. A non-transitory computer readable storage medium, wherein the non-transitory computer readable storage medium stores computer instructions, the computer instructions cause the computer to perform the method of any one of claims 1 to 7.

Description

Semi-supervised learning-based labor ultrasonic quantitative measurement method and system Technical Field The invention belongs to the field of medical image diagnosis, and particularly relates to a labor ultrasonic quantitative measurement method and system based on semi-supervised learning. Background Transperineal ultrasound (TRANSPERINEAL ULTRASOUND, TPU) is the most visual, noninvasive and repeatable imaging mode for evaluating fetal head descent in labor management, provides objective geometric basis for labor progress through real-time imaging of the relative position of fetal head and pubic symphysis, and is a safer and quantifiable method for replacing vaginal examination. Clinical studies show that subtle changes in fetal head position are closely related to delivery outcome, wherein angular progression reflects angular trends in fetal head rotation and descent, and the head-pubic distance quantifies the depth of descent of the fetal head relative to the birth canal, both of which have important reference values in predicting delivery time, judging delivery style, and assessing midwifery intervention. However, conventional AoP and HSD measurement generally rely on an operator to manually mark key points or measurement angles on a static B-mode ultrasonic image, and the operation process is obviously influenced by experience, image definition and probe position, so that the problems of large measurement error, poor repeatability, strong subjectivity and the like exist. In a complex or slow-progressing labour, there may be significant deviations in the measurement results of different doctors for the same case, affecting the clinical judgment. With the increase of the requirements for labor monitoring, an automatic and standardized quantitative measurement method based on transperineal ultrasound becomes an important direction for improving labor management efficiency and safety. Existing automated measurement methods based on deep learning, while making some progress in the ultrasound analysis of labor, still face three key challenges: the method has the key challenges that image noise and structural complexity cause difficult target positioning, and TPU image quality is obviously affected by factors such as speckle noise, sound artifact, probe direction change and the like, so that structural boundary blurring is caused. Pubic Symphysis (PS) usually occurs near tissue areas that are much larger and highlight than the fetal head, making it difficult for the model to distinguish anatomical boundaries and achieve stable localization. In addition, fetal head contour morphology changes with the stage of labor, further increasing the complexity of automatic detection. The key challenges are that high-quality labeling data are scarce, and the requirement of full-supervision learning is difficult to meet, namely that the labeling process of the ultrasound data of the labor process is required to be completed by obstetrical ultrasound doctors with abundant experience, and different operators have differences in the division of key points and structural boundaries. The labeling cost is high, the subjectivity is strong, and the high-quality pixel fraction data is extremely limited. The data scale currently disclosed or clinically available cannot support the training of an end-to-end full-supervision model, so that the model is easy to overfit and lacks generalization capability. The key challenges are that the available data set only contains sparse landmark annotation, the structure information is insufficient, and most of the existing data only provides sparse annotation at the level of key points or line segments and lacks a complete segmentation mask. The manual segmentation of TPU images is labor-intensive, time-consuming and requires expert domain knowledge, and the feasibility of large-scale accurate labeling is greatly limited. Such point-based labels do not adequately capture the geometric topological relationship of the fetal head and pubic symphysis, while segmentation masks can encode richer anatomical and contextual information, which is critical to robust measurement and geometric interpretation. Disclosure of Invention In order to solve the problems of difficult target positioning caused by image noise and structural complexity, difficult target positioning caused by image noise and structural complexity and insufficient structural information caused by sparse landmark annotation contained in an available data set in the existing deep learning-based labor process automatic measurement method, the invention provides a labor process ultrasonic quantitative measurement method and system based on semi-supervised learning, which aim to realize accurate segmentation and automatic extraction of key points of fetal head and pubic bone union and further calculate AoP and HSD two labor process quantitative clinical indexes under the condition of limited labeling samples, and remarkably improve the str