CN-122023443-A - Lesion region segmentation method and system of medical image

CN122023443ACN 122023443 ACN122023443 ACN 122023443ACN-122023443-A

Abstract

The invention discloses a lesion region segmentation method and a lesion region segmentation system of a medical image, which relate to the technical field of image processing and comprise the following steps of constructing a knowledge graph based on multi-mode disease data of different patients; the method comprises the steps of taking a lesion area of a medical image to be segmented as a state, taking a classification result corresponding to the lesion area as an action, taking the consistency degree of the classification result and a disease label given in a knowledge graph as a reward, inputting the state of a current iteration into a segmentation network to obtain a segmentation result, inputting the segmentation result into the classification network to obtain the current action, obtaining the current reward based on the consistency degree of the current action and the disease label given in the knowledge graph, adjusting the state through the current reward to obtain the state of the next iteration, outputting a target lesion area after the iteration is completed, and inputting the target lesion area into the segmentation network for segmentation. The invention establishes a microscopic-to-macroscopic cross-scale information mapping relation through reinforcement learning.

Inventors

CHEN YUANYUAN
XU YUNGANG
YANG MENGYUAN
YANG ZEYUAN

Assignees

西安交通大学

Dates

Publication Date: 20260512
Application Date: 20260131

Claims (7)

1. A lesion region segmentation method of a medical image, comprising the steps of: Constructing a knowledge graph based on multi-modal disease data of different patients; Collecting a medical image to be segmented, taking a lesion area of the medical image to be segmented as a state, taking a classification result corresponding to the lesion area as an action, and taking the consistency degree of the classification result and a disease label given in a knowledge graph as a reward; Positioning a lesion area of a medical image to be segmented through reinforcement learning to obtain a target lesion area, wherein the state of the current iteration is input into a segmentation network to obtain a segmentation result, the segmentation result is input into a classification network to obtain a current action; And after the iteration is completed, outputting the target lesion area, and inputting the target lesion area into a segmentation network for segmentation.
2. The method for segmenting a lesion area in a medical image according to claim 1, wherein the multi-modality data includes images, pathology images, genes and clinical cases, and the knowledge graph is constructed based on the multi-modality disease data of different patients, comprising the steps of: Extracting features of the multi-mode data to obtain spatial feature vectors of images, multi-scale vectors of pathological images, structural embedded representation of gene data and semantic vector representation of clinical cases; Aligning and feature fusion are carried out on the space feature vector, the multi-scale vector, the structure embedded representation and the semantic vector representation to obtain multi-mode features; Retrieving the multi-mode data to obtain relevant knowledge of lesions; Performing biomedical entity identification from the related knowledge, identifying the entities, and extracting the relationship between the entities; and combining the multi-modal characteristics, the entities and the relations thereof to form a core knowledge graph.
3. A method for segmenting a lesion region in a medical image according to claim 2, wherein said feature extraction of multi-modal data comprises the steps of: extracting features of the image data through a three-dimensional convolutional neural network; extracting features of the pathological image through HRNet; extracting the characteristics of the gene data through a graph neural network; and extracting the characteristics of the clinical cases through a medical pre-training language model.
4. The method for segmenting a lesion area in a medical image according to claim 1, wherein the degree of consistency is evaluated by using the similarity of the classification result and the semantic cosine of the disease label given by the knowledge graph.
5. The method of claim 1, wherein the adjustment of the state includes a left shift, a right shift, an up shift, a down shift, a zoom in and a zoom out of the lesion area.
6. The method for segmenting a lesion area in a medical image according to claim 1, wherein the segmentation network is a U-net network, and the classification model is a 6-layer fully connected convolutional neural network.
7. A lesion field segmentation system for a medical image, comprising: the construction module is used for constructing a knowledge graph based on multi-mode disease data of different patients; the definition module is used for collecting the medical image to be segmented, taking a lesion area of the medical image to be segmented as a state, taking a classification result corresponding to the lesion area as an action, and taking the consistency degree of the classification result and the disease label given in the knowledge graph as a reward; The positioning module is used for positioning a lesion area of the medical image to be segmented through reinforcement learning to obtain a target lesion area, wherein the state of the current iteration is input into a segmentation network to obtain a segmentation result, the segmentation result is input into a classification network to obtain a current action; And the segmentation module is used for outputting the target lesion area after the iteration is completed, and inputting the target lesion area into the segmentation network for segmentation.

Description

Lesion region segmentation method and system of medical image Technical Field The present invention relates to the field of image processing technologies, and in particular, to a method and a system for segmenting a lesion region in a medical image. Background Primary hepatocellular carcinoma (HCC) is one of the malignant tumors with high morbidity and mortality in China, and the prognosis of patients is poor, with the total survival rate of only about 12.1% in 5 years. The realization of accurate diagnosis and personalized treatment of liver cancer is a key for improving survival of patients. In the diagnosis and treatment process of liver cancer, the method plays a vital role in accurate segmentation of lesion areas in medical images (such as CT and MRI) and pathological images. The segmentation result is the basis of quantitative analysis of the size, shape and position of the tumor and the relation with surrounding blood vessels/tissues, directly influences the tumor stage and the selection of treatment schemes, and can provide basis for the accurate sketching of the surgical excision range or the radiotherapy target zone, so as to strive to clear the focus to the maximum extent and protect the normal liver tissues. Meanwhile, by comparing the changes of the segmentation results before and after treatment, the treatment effect (such as tumor volume reduction rate) can be objectively evaluated and recurrence can be monitored. In addition, the focus area obtained by segmentation is a core anchor point for integrating imaging features with trans-scale and multi-modal information such as genomics, pathology, clinical data and the like for comprehensive analysis. Therefore, the development of efficient, robust and accurate medical image lesion segmentation technology, especially for complex liver cancer lesions, has urgent clinical requirements and important application value. Currently, the mainstream medical image segmentation method mainly depends on a deep learning technology, but has significant limitations and challenges including 1, depending on a single image mode, and insufficient information utilization, wherein the existing method is mostly based on single type image data (such as single CT or single MRI) for segmentation. Failure to effectively integrate multimodal information (e.g., MRI, PET-CT, ultrasound, gene expression profiling, clinical pathology data, etc. of a patient, of different sequences) limits the overall and thorough understanding and segmentation accuracy of the model to the lesion, especially in cases of foci boundary blurring, high heterogeneity, or low contrast with background tissue. 2. Lack of trans-scale information integration capability-existing segmentation processes lack efficient utilization of broader biological background and clinical context knowledge, resulting in segmentation results that may lack interpretability at a biological meaning level and that are difficult to accommodate for complex individual differences. 3. Model feedback and self-adaptive optimization mechanisms lack of fixed parameters after traditional segmentation model training is finished, and an effective mechanism is lacking, so that potential of the model in dynamic clinical decision support is limited. In summary, the existing liver cancer medical image segmentation technology has obvious defects in the aspects of fully utilizing multi-mode information, realizing cross-scale knowledge fusion and constructing a closed-loop self-adaptive optimization mechanism. There is a need to develop a novel intelligent segmentation method to break through the bottleneck and provide a more reliable technical support for accurate diagnosis and treatment of liver cancer. Disclosure of Invention Based on the defects in the prior art, the invention provides a lesion area segmentation method and a lesion area segmentation system for medical images, which solve the existing problems. The invention adopts the following technical scheme: In a first aspect, the present invention provides a lesion region segmentation method of a medical image, comprising the steps of: Constructing a knowledge graph based on multi-modal disease data of different patients; Collecting a medical image to be segmented, taking a lesion area of the medical image to be segmented as a state, taking a classification result corresponding to the lesion area as an action, and taking the consistency degree of the classification result and a disease label given in a knowledge graph as a reward; Positioning a lesion area of a medical image to be segmented through reinforcement learning to obtain a target lesion area, wherein the state of the current iteration is input into a segmentation network to obtain a segmentation result, the segmentation result is input into a classification network to obtain a current action; And after the iteration is completed, outputting the target lesion area, and inputting the target lesion area into a segmentation netw