CN-122023253-A - Medical image labeling system and method

CN122023253ACN 122023253 ACN122023253 ACN 122023253ACN-122023253-A

Abstract

The invention provides a medical image labeling system and a medical image labeling method, which relate to the technical field of image segmentation, wherein the system comprises a data labeling module for acquiring a clinical original medical image, and labeling pretreatment is carried out on the clinical original medical image to obtain a small-scale high-quality labeling data set; the fine tuning module performs fine tuning on the single-target segmentation model by utilizing the small-scale high-quality labeling data set; the method comprises the steps of obtaining a clinical original medical image by a segmentation module, obtaining a large-scale high-quality data set by using a fine-tuned single-target segmentation model, training a multi-target segmentation model by using the large-scale high-quality data set by a training module, obtaining a clinical medical image to be processed by a medical image labeling module, inputting the clinical medical image to be processed into the trained multi-target segmentation model, and outputting a medical image labeling result by the trained multi-target segmentation model. The method can effectively break through the bottleneck of scarcity of the labeling data, and improve the image labeling efficiency and quality.

Inventors

YANG YUEDONG
TIAN CHONG
ZHAO HUIYING
OU PEIHUA

Assignees

中山大学

Dates

Publication Date: 20260512
Application Date: 20251226

Claims (10)

1. A medical image labeling system, comprising: The data labeling module is used for acquiring clinical original medical images, and labeling and preprocessing the clinical original medical images to obtain a small-scale high-quality labeling data set; The fine tuning module is used for carrying out fine tuning on a preset single-target segmentation model by utilizing the small-scale high-quality labeling data set to obtain a fine-tuned single-target segmentation model; The segmentation module is used for segmenting the clinical original medical image by utilizing the fine-tuned single-target segmentation model to obtain a large-scale high-quality data set; The training module is used for training a preset multi-target segmentation model by utilizing the large-scale high-quality data set to obtain a trained multi-target segmentation model; the medical image labeling module is used for acquiring a clinical medical image to be processed, inputting the clinical medical image to be processed into the trained multi-target segmentation model, and outputting a medical image labeling result by the trained multi-target segmentation model.
2. The medical image labeling system of claim 1, wherein the data labeling module comprises a data loading and visualization sub-module, an interactive segmentation sub-module, and a segmentation result correction sub-module connected in sequence, the labeling preprocessing of the clinical raw medical image comprising: The data loading and visualizing submodule loads the clinical original medical image to obtain a loaded image, and visualizes the loaded image to obtain a visualized image to be segmented; the interactive segmentation submodule segments the image to be segmented to obtain a segmentation mask; and the segmentation result correction submodule corrects the segmentation mask to obtain a new segmentation mask serving as the small-scale high-quality annotation data set.
3. The medical image labeling system of claim 2, wherein the segmenting the image to be segmented comprises: Determining a prompt provided by a user, and encoding the prompt by using a first prompt encoder to obtain a prompt code; inputting the image to be segmented into a first image encoder for feature extraction to obtain comprehensive image features; And inputting the comprehensive image characteristics and the prompt codes into a first mask decoder for fusion to obtain the segmentation mask.
4. A medical image labeling system as in claim 3 wherein said modifying said segmentation mask comprises at least one of: The user adds a positive prompt point or a negative prompt point on the segmentation mask to obtain a new prompt, the first prompt encoder is utilized to encode the new prompt to obtain a new prompt code, and the new prompt code and the comprehensive image characteristics are input into the first mask decoder again to be fused to obtain a new segmentation mask; The user directly adjusts the shape of the segmentation mask by dragging the control point of the edge of the segmentation mask to obtain a new segmentation mask; the user uses the brush tool to directly add or erase pixels in the mask area of the segmentation mask to obtain a new segmentation mask.
5. The medical image labeling system of claim 1, wherein the single-object segmentation model comprises a second hint encoder, a second image encoder, and a second mask decoder, wherein the second hint encoder is configured to input a single-object segmentation hint code, the second image encoder is configured to perform feature extraction on the small-scale high-quality labeling dataset, output an image feature code, and the second mask decoder is configured to fuse the single-object segmentation hint code with the image feature code to obtain a single-object segmentation result.
6. The medical image labeling system of claim 5, wherein the fine-tuning of a pre-set single-target segmentation model with the small-scale high-quality labeling dataset comprises: inputting the small-scale high-quality annotation data set; According to the medical segmentation task corresponding to the labeling data in the small-scale high-quality labeling data set, a plurality of low-rank adapters are deployed in parallel in the second image encoder, and local attention layer parameters are frozen to construct a parameter fine-tuning framework; aiming at the parameter fine tuning framework, a sparse activation strategy of a lightweight routing network is started, and load balancing regularization constraint is applied; And carrying out back propagation training on network parameters of the low-rank adapter in the parameter fine adjustment framework by using the small-scale high-quality annotation data set under the constraint of load balance regularization until training times reach a training threshold value, ending training, and outputting a single-target segmentation model adapting to the medical segmentation task.
7. The medical image labeling system of any of claims 1-6, wherein the multi-target segmentation model comprises a third image encoder and a third image decoder connected in sequence.
8. The medical image labeling system of claim 7, wherein the training of the pre-set multi-objective segmentation model with the large-scale high-quality dataset comprises: Determining a multi-target segmentation task corresponding to the large-scale high-quality data set; injecting a mixed expert-low-rank self-adaptive module into a plurality of full-attention layers responsible for global context modeling in the third image encoder, and setting a multi-scale feature fusion and multi-task collaborative learning strategy; And carrying out counter propagation training on the mixed expert-low rank self-adaptive module in the third image encoder and the network parameters of the third image decoder by utilizing the large-scale high-quality data set under the constraint of the multi-scale feature fusion and multi-task collaborative learning strategy until the training times reach a training threshold value, ending the training, and outputting a multi-target segmentation model adapting to the multi-target segmentation task.
9. A medical image labeling method, characterized by comprising the steps of: s1, acquiring a clinical original medical image, and performing labeling pretreatment on the clinical original medical image to obtain a small-scale high-quality labeling data set; S2, fine-tuning a preset single-target segmentation model by using the small-scale high-quality labeling data set to obtain a fine-tuned single-target segmentation model; S3, segmenting the clinical original medical image by utilizing the fine-tuned single-target segmentation model to obtain a large-scale high-quality data set; s4, training a preset multi-target segmentation model by using the large-scale high-quality data set to obtain a trained multi-target segmentation model; S5, acquiring a clinical medical image to be processed, inputting the clinical medical image to be processed into a trained multi-target segmentation model, and outputting a medical image labeling result by the trained multi-target segmentation model.
10. The computer equipment is characterized by comprising a processor, a memory, a communication interface and a communication bus, wherein the processor, the memory and the communication interface complete communication with each other through the communication bus; The memory is configured to store at least one executable instruction that causes the processor to perform the operations of the medical image labeling method of claim 9.

Description

Medical image labeling system and method Technical Field The invention relates to the technical field of image segmentation, in particular to a medical image labeling system and a medical image labeling method. Background Image segmentation labeling is one of important tasks in medical image analysis, and aims to outline corresponding regions of interest such as organs, lesions, pathological cells and the like, so as to be used for disease diagnosis, operation planning, disease progress monitoring and pathological analysis. In practical clinical application, the medical image segmentation technology still faces a serious challenge, firstly, for special diseases, labeling data is very scarce, available samples are often less than hundred cases, training requirements of a data driving model are difficult to meet, secondly, the labeling process of a single image is highly dependent on a professional doctor, the time can be tens of minutes to hours, and the construction of a large-scale data set is difficult. Although the proposal of U-Net and Vision Transformer significantly promotes the application process of the image segmentation technology in clinic, the bottleneck problem of the data collection link can not be effectively solved, in order to solve the problem, the segmentation universal Model represented by a SEGMENT ANYTHING Model (SAM) and a medical derivative Model MedSAM thereof shows strong generalization and zero-shot capability, the SAM provides a convenient interactive labeling interface and is beneficial to reducing the manual labeling burden, however, the segmentation universal Model has the problems of multi-center heterogeneity, obvious noise interference and the like when labeling medical image data, wherein the multi-center heterogeneity refers to the medical image acquired by different medical institutions by the segmentation universal Model, and has significant difference in data characteristics and labeling standards, the noise interference obviously refers to the fact that image noise can interfere the identification of lesion characteristics by the Model, so that the Model misjudges the noise as a lesion area, or the fact that the Model is unable to identify tiny lesions covered by noise, so that the universal Model is difficult to directly adapt in a real clinical scene, the image labeling efficiency and quality are low, and further the vicious cycle of the data collection and the Model performance is insufficient. Disclosure of Invention In order to solve the problems of scarcity of marking data and low image marking efficiency and quality in the prior art, the invention provides a medical image marking system and a medical image marking method, which can effectively break through the bottleneck of scarcity of marking data and improve the image marking efficiency and quality. In order to achieve the technical effects, the technical scheme of the invention is as follows: a medical image annotation system comprising: The data labeling module is used for acquiring clinical original medical images, and labeling and preprocessing the clinical original medical images to obtain a small-scale high-quality labeling data set; The fine tuning module is used for carrying out fine tuning on a preset single-target segmentation model by utilizing the small-scale high-quality labeling data set to obtain a fine-tuned single-target segmentation model; The segmentation module is used for segmenting the clinical original medical image by utilizing the fine-tuned single-target segmentation model to obtain a large-scale high-quality data set; The training module is used for training a preset multi-target segmentation model by utilizing the large-scale high-quality data set to obtain a trained multi-target segmentation model; the medical image labeling module is used for acquiring a clinical medical image to be processed, inputting the clinical medical image to be processed into the trained multi-target segmentation model, and outputting a medical image labeling result by the trained multi-target segmentation model. Preferably, the data labeling module includes a data loading and visualizing sub-module, an interactive segmentation sub-module and a segmentation result correction sub-module which are sequentially connected, and the labeling preprocessing for the clinical original medical image includes: The data loading and visualizing submodule loads the clinical original medical image to obtain a loaded image, and visualizes the loaded image to obtain a visualized image to be segmented; the interactive segmentation submodule segments the image to be segmented to obtain a segmentation mask; and the segmentation result correction submodule corrects the segmentation mask to obtain a new segmentation mask serving as the small-scale high-quality annotation data set. Preferably, the segmenting the image to be segmented includes: Determining a prompt provided by a user, and encoding the prompt by using a first pro