CN-121147243-B - Ultrasonic artery segmentation method, system and equipment based on positioning knowledge driving

CN121147243BCN 121147243 BCN121147243 BCN 121147243BCN-121147243-B

Abstract

The invention provides an ultrasonic artery segmentation method, system and equipment based on positioning knowledge driving, and relates to the technical field of medical image processing. The method comprises the steps of obtaining an original ultrasonic image, wherein the original ultrasonic image is an image containing an arterial lumen, realizing mask mapping of a rough area of the original ultrasonic image based on a positioning priori learning model to obtain a positioning map containing an arterial highlight area, inputting the original ultrasonic image and the positioning map into a fusion model to obtain a fusion feature map, and segmenting the fusion feature map based on a segmentation network model to obtain a target segmentation image, wherein the fusion model is a model for guiding self-adaptive fusion. The method effectively solves the problem of positioning ambiguity of the arterial lumen in the ultrasonic image, and ensures the correct positioning of the target, so that the subsequent segmentation is more stable and accurate, and the missed segmentation and the wrong segmentation are effectively avoided.

Inventors

FANG ZHEN
YAN MENGXUE
DU LIDONG
WANG PENG
LI ZHENFENG
WU PANG
CHEN XIANXIANG

Assignees

中国科学院空天信息创新研究院

Dates

Publication Date: 20260508
Application Date: 20251118

Claims (8)

1. An ultrasonic artery segmentation method based on positioning knowledge driving, which is characterized by comprising the following steps: acquiring an original ultrasonic image, wherein the original ultrasonic image is an image containing an arterial lumen; Based on a positioning priori learning model, realizing the mask mapping of the rough region of the original ultrasonic image so as to obtain a positioning map containing the arterial highlight region; Inputting the original ultrasonic image and the positioning image into a fusion model to obtain a fusion feature image; based on a segmentation network model, segmenting the fusion feature map to obtain a target segmentation image; the fusion model is a model for guiding self-adaptive fusion; the positioning priori learning model is obtained through pre-training, and the training process of the positioning priori learning model comprises the following steps: acquiring a first training set, wherein the first training set is a coarse-marked training set and comprises a plurality of images to be trained containing arterial lumens; inputting the first training set into a positioning priori learning network, and mapping a rough region mask on each image to be trained through a composite loss function; the training process of the positioning priori learning model further comprises the following steps: Acquiring a second training set, wherein the second training set is a fine-marked training set; selecting a sample image set in the second training set according to a preset proportion; Generating a corresponding pseudo tag data set from the sample image set based on a preset geometric algorithm; and fine tuning the positioning prior learning model based on the pseudo tag data set, and outputting a positioning map containing the arterial highlight region.
2. The positioning knowledge driven ultrasonic artery segmentation method according to claim 1, wherein the fusion model is a multi-layer perceptron structure, and the inputting the original ultrasonic image and the positioning map into the fusion model to obtain a fusion feature map comprises: calculating to obtain an average value of the original ultrasonic image and the positioning map; inputting the average values of the original ultrasonic image and the positioning image into the multi-layer perceptron structure to generate a dynamic weight image containing two channels; and based on the dynamic weight map, channel splicing is carried out to obtain a fusion feature map.
3. The ultrasonic artery segmentation method based on positioning knowledge driving according to claim 1, wherein the segmentation network model is obtained through pre-training, and the training process of the segmentation network model comprises the following steps: Regularizing the segmentation network model through a smooth label boundary.
4. The localization knowledge driven ultrasound artery segmentation method of claim 3, wherein regularizing the segmentation network model with smooth label boundaries comprises: acquiring a third training set, wherein the third training set comprises a plurality of finely marked binary truth masks; Determining the boundary of each binary truth mask by using a morphological gradient; the convolution processing is carried out on the boundary of each binary truth mask by a preset filter; integrating the boundary after convolution processing with a binary truth mask to generate soft label data; regularizing the segmentation network model based on the soft tag data.
5. The method for segmenting the ultrasound artery based on the positioning knowledge driving of claim 4, wherein the segmenting the fusion feature map based on the segmentation network model to obtain the target segmented image comprises: and inputting the fusion feature map into a pre-trained segmentation network model, and using a composite loss function based on soft label data as a standard to complete segmentation of the fusion feature map so as to obtain a predicted target segmentation image.
6. The localization knowledge driven ultrasound arterial segmentation method according to claim 1, wherein the target segmented image is a probability map, the method further comprising: and judging the target segmentation image based on a preset threshold value to obtain a target binary segmentation mask.
7. An ultrasound arterial segmentation system based on localization knowledge driving, comprising: The original image acquisition module is used for acquiring an original ultrasonic image, wherein the original ultrasonic image is an image containing an arterial lumen; The positioning priori learning module is used for realizing the mask mapping of the rough region of the original ultrasonic image based on the positioning priori learning model so as to obtain a positioning map containing the arterial highlight region; The fusion module is used for inputting the original ultrasonic image and the positioning image into a fusion model to obtain a fusion feature image; the segmentation module is used for segmenting the fusion feature map based on a segmentation network model to obtain a target segmentation image; the fusion model is a model for guiding self-adaptive fusion; the positioning priori learning model is obtained through pre-training, and the training process of the positioning priori learning model comprises the following steps: acquiring a first training set, wherein the first training set is a coarse-marked training set and comprises a plurality of images to be trained containing arterial lumens; inputting the first training set into a positioning priori learning network, and mapping a rough region mask on each image to be trained through a composite loss function; the training process of the positioning priori learning model further comprises the following steps: Acquiring a second training set, wherein the second training set is a fine-marked training set; selecting a sample image set in the second training set according to a preset proportion; Generating a corresponding pseudo tag data set from the sample image set based on a preset geometric algorithm; and fine tuning the positioning prior learning model based on the pseudo tag data set, and outputting a positioning map containing the arterial highlight region.
8. An electronic device, comprising: One or more processors; a memory for storing one or more computer programs, Characterized in that the one or more processors execute the one or more computer programs to implement the steps of the method according to any one of claims 1-6.

Description

Ultrasonic artery segmentation method, system and equipment based on positioning knowledge driving Technical Field The invention relates to the technical field of medical image processing, in particular to an ultrasonic artery segmentation method, system and equipment based on positioning knowledge driving. Background Ultrasonic imaging has become a first-line tool for arterial assessment, clinical screening, and health monitoring due to its advantages of non-invasiveness, safety, economy, real-time, etc. The accurate segmentation of arterial lumen in ultrasonic image is the basis for evaluating critical clinical indexes such as vascular stenosis, measuring sectional area, blood flow estimation, and the like, and directly affects the accuracy of clinical decision. However, automatic segmentation of arterial lumens presents a great challenge, which is mainly due to the inherent characteristics of ultrasound images. First, arterial targets are highly similar in characteristics to surrounding collateral tissue (e.g., veins), imaging artifacts, etc., and the image contrast is low, which introduces significant interference and ambiguity for accurate lumen localization, i.e., "localization ambiguity". Secondly, because the ultrasound beam is nearly perpendicular to the vessel sidewall, resulting in lumen contours, particularly the sidewall, often present blurring or signal loss, further exacerbating the difficulty of positioning and accurate tracing. To address these challenges, the prior art has mainly employed implicit strategies to address positioning issues, but with limited effectiveness. For example, based on data enhancement methods, many studies rely on data enhancement means such as random clipping, translation, etc., in an effort to implicitly enhance the model's learning of target location information during the training process. However, this method cannot fundamentally solve the inherent positioning ambiguity of the image content itself. For example, attention mechanism based methods, some of which introduce a position or spatial attention module in the segmented network, attempt to let the model learn autonomously and focus on the target region. However, the generalization ability of such methods is often insufficient, and the performance is significantly degraded when facing the situations of poor image quality or complex background. For another example, a shape prior-based approach, a partial approach introduces a fixed shape prior, such as assuming that the vessel cross-section is elliptical, to constrain the segmentation results. This approach lacks flexibility and is difficult to handle the various irregular vascular morphologies due to pathological changes or imaging angles. In addition, with respect to cascading or coarse-to-fine procedures in the prior art, some methods employ a two-stage procedure in which a region of interest is located with a detection network and then finely segmented within the region by a segmentation network. The final performance of such methods is severely dependent on the accuracy of the first stage positioning, which is prone to error propagation problems. Therefore, the prior art mostly solves the positioning problem in an implicit mode, and can not effectively and directly solve the positioning ambiguity of the arterial lumen in the ultrasonic image, thereby limiting the accuracy and the robustness of the segmentation. Disclosure of Invention The invention provides an ultrasonic artery segmentation method based on positioning knowledge driving, which comprises the steps of obtaining an original ultrasonic image, wherein the original ultrasonic image is an image containing an arterial lumen, realizing rough region mask mapping of the original ultrasonic image based on a positioning priori learning model to obtain a positioning map containing an arterial highlight region, inputting the original ultrasonic image and the positioning map into a fusion model to obtain a fusion feature map, segmenting the fusion feature map based on a segmentation network model to obtain a target segmentation image, and the fusion model is a model guiding self-adaptive fusion. According to the embodiment of the invention, the positioning priori learning model is obtained through pre-training, and the training process of the positioning priori learning model comprises the steps of obtaining a first training set, wherein the first training set is a rough marked training set, the first training set comprises a plurality of images to be trained comprising arterial lumens, inputting the first training set into a positioning priori learning network, and mapping a rough region mask on each image to be trained through a composite loss function. According to the embodiment of the invention, the training process of the positioning priori learning model further comprises the steps of obtaining a second training set which is a fine-marked training set, selecting a sample image set in the second tr