CN-121999310-A - Construction method and device for substrate model library for identifying medical images

CN121999310ACN 121999310 ACN121999310 ACN 121999310ACN-121999310-A

Abstract

The invention discloses a construction method and a construction device for a base model library for identifying medical images, wherein the method comprises the steps of acquiring medical image data, preprocessing and labeling the medical image data to form a data set, training a plurality of deep learning models special for a plurality of anatomical regions based on the data set, enabling each deep learning model to identify and describe morphological characteristics and/or pathological characteristics of the corresponding anatomical region, and constructing a comprehensive base model library by utilizing a multi-scale and/or multi-mode fusion algorithm based on the trained deep learning model. According to the invention, by constructing the base model library for identifying different anatomical regions in the image, an efficient tool is provided, the interconnection and intercommunication of medical data are promoted, and the automatic analysis capability of the medical image can be greatly improved; the method not only has high precision and robustness, but also can dynamically update and adapt to new medical requirements, and has wide clinical application prospect.

Inventors

YANG HUIFANG
PENG XIN
HU LEIHAO

Assignees

北京大学口腔医学院

Dates

Publication Date: 20260508
Application Date: 20241106

Claims (10)

1. A method of constructing a base model library for identifying medical images, the method comprising: acquiring medical image data, preprocessing and marking the medical image data to form a data set; training a plurality of deep learning models specific to a plurality of anatomical regions based on the dataset, each deep learning model being capable of identifying and describing morphological and/or pathological features of the respective anatomical region; Based on the trained deep learning model, a comprehensive base model library is constructed by utilizing a multi-scale and/or multi-modal fusion algorithm.
2. The method according to claim 1, wherein the method further comprises: And optimizing or calibrating the base model library by utilizing an ensemble learning algorithm, so that the base model library can automatically identify and position different anatomical regions in the medical image data.
3. The method of claim 1, wherein the medical image data comprises at least one of a CT image, an MRI image, an ultrasound image, or a PET image, and/or The anatomical region includes at least one of the brain, heart, lung, liver, teeth, digestive system, skeletal system, nervous system, or circulatory system.
4. A method according to any one of claims 1-3, characterized in that the method further comprises: Dynamically updating the base model library, periodically collecting new clinical image data and retraining the model library to accommodate new image data and anatomical study progression, and/or Adapting the existing base model library to a new anatomical region or a new image modality using a transfer learning algorithm, and/or Based on the individuation difference, the adaptability and the accuracy of the base model library are improved by introducing the characteristic features of the patient.
5. A method according to any one of claims 1-3, wherein constructing a comprehensive base model library using a multimodal fusion algorithm based on the trained deep learning model comprises: Extracting text features from a case or a diagnosis report text of a patient by using a large language model, fusing the text features into the deep learning model, and improving understanding and recognition accuracy of each anatomical region, wherein the method specifically comprises at least one of the following steps: Directly adjusting and optimizing the parameters of the base model through the text characteristics by utilizing the natural language understanding capability of the large language model; Collecting feedback and comments by using the large language model through dialogue with a patient or doctor, and iteratively optimizing the deep learning model by using text features extracted by the large language model; Directly generating accurate description features of the anatomical region by using the large language model, automatically detecting abnormality, identifying potential lesion regions, and generating diagnosis suggestions; automatically generating an explanatory report for the image analysis result of the deep learning model by combining the natural language generation capability of the large language model, and carrying out quantitative analysis to the image analysis result to cover qualitative explanation or carrying out risk assessment suggestion; Combining the result of image analysis with a medical knowledge graph by utilizing the text features extracted by the large language model, thereby realizing automatic reasoning, inquiring or recommending solutions; And automatically generating high-quality labels for the medical image data by utilizing the text features extracted by the large language model.
6. A method according to any of claims 1-3, wherein constructing a comprehensive base model library using a multi-scale algorithm based on the trained deep learning model comprises: and constructing a multi-scale feature pyramid, and identifying an anatomical structure from medical image data with different resolutions by utilizing the multi-scale feature pyramid and combining the deep learning model.
7. A method according to any one of claims 1-3, characterized in that the method further comprises: the results of the recognition are presented through three-dimensional visualization operations including rotation, scaling or layering, so that a user can intuitively view the positioning and feature descriptions of each anatomical region, so that the user can view the anatomical region at different viewing angles.
8. A construction apparatus for identifying a base model library of medical images, the apparatus comprising: The acquisition module is suitable for acquiring medical image data, preprocessing and marking the medical image data, and forming a data set; A training module adapted to train a plurality of deep learning models specific to a plurality of anatomical regions based on the dataset, each deep learning model being capable of identifying and describing morphological and/or pathological features of the respective anatomical region; The construction module is suitable for constructing a comprehensive base model library by utilizing a multi-scale and/or multi-mode fusion algorithm based on the trained deep learning model.
9. An electronic device comprising a processor and a memory arranged to store computer executable instructions which, when executed, cause the processor to perform the method of constructing a base model library identifying medical images according to any one of claims 1-7.
10. A computer-readable storage medium storing one or more programs that, when executed by a processor, implement the method of constructing a base model library for identifying medical images according to any one of claims 1-7.

Description

Construction method and device for substrate model library for identifying medical images Technical Field The invention relates to the technical field of medicine, in particular to a method and a device for constructing a base model library for identifying medical images. Background With the development of medical imaging technology, automated image analysis tools are becoming increasingly used in clinical diagnostics. However, due to the complexity and individual differences of the different anatomical regions, accurately locating and characterizing these regions remains challenging. Different imaging techniques (e.g., CT, MRI, PET, etc.) provide different levels of anatomical information, but each of these information has limitations. For example, CT images provide clear anatomical detail but poor soft tissue differentiation, while MRI can display soft tissue structures but has limited spatial resolution. Therefore, how to effectively fuse the image information of the different modes, so as to fully exert respective advantages, and accurately identify the anatomical structure is a great challenge. Furthermore, medical data privacy protection is a non-negligible issue. Data inside hospitals is often severely protected to prevent patient privacy disclosure. The limited data management can influence the training and the use of the model, but by constructing a base database, the privacy of a patient can be effectively protected, and meanwhile, personalized differences of different crowds are provided, so that an effective method is provided for sharing and interconnection of data. The base library construction method of the data can effectively promote sharing of the data and improve accuracy and universality of use of the model in data with different sources. Therefore, developing a base model library that can identify different anatomical regions and automatically extract relevant features is of great importance to improving diagnostic accuracy, and data interconnection. Disclosure of Invention The present invention has been made in view of the above-mentioned problems, and it is an object of the present invention to provide a method and apparatus for constructing a library of basis models for identifying medical images that overcomes or at least partially solves the above-mentioned problems. According to one aspect of the present invention, there is provided a method for constructing a base model library for identifying medical images, the method comprising: acquiring medical image data, preprocessing and marking the medical image data to form a data set; training a plurality of deep learning models specific to a plurality of anatomical regions based on the dataset, each deep learning model being capable of identifying and describing morphological and/or pathological features of the respective anatomical region; Based on the trained deep learning model, a comprehensive base model library is constructed by utilizing a multi-scale and/or multi-modal fusion algorithm. In some embodiments, the method further comprises: And optimizing or calibrating the base model library by utilizing an ensemble learning algorithm, so that the base model library can automatically identify and position different anatomical regions in the medical image data. In some embodiments, the medical image data includes at least one of CT images, MRI images, ultrasound images, or PET images, and/or The anatomical region includes at least one of the brain, heart, lung, liver, teeth, digestive system, skeletal system, nervous system, or circulatory system. In some embodiments, the method further comprises: Dynamically updating the base model library, periodically collecting new clinical image data and retraining the model library to accommodate new image data and anatomical study progression, and/or Adapting the existing base model library to a new anatomical region or a new image modality using a transfer learning algorithm, and/or Based on the individuation difference, the adaptability and the accuracy of the base model library are improved by introducing the characteristic features of the patient. In some embodiments, constructing a comprehensive base model library using a multimodal fusion algorithm based on the trained deep learning model comprises: Extracting text features from a case or a diagnosis report text of a patient by using a large language model, fusing the text features into the deep learning model, and improving understanding and recognition accuracy of each anatomical region, wherein the method specifically comprises at least one of the following steps: Directly adjusting and optimizing the parameters of the base model through the text characteristics by utilizing the natural language understanding capability of the large language model; Collecting feedback and comments by using the large language model through dialogue with a patient or doctor, and iteratively optimizing the deep learning model by using text features ext