CN-121983300-A - Pet pathology whole-piece image analysis and auxiliary report system based on visual language AI model

CN121983300ACN 121983300 ACN121983300 ACN 121983300ACN-121983300-A

Abstract

The invention provides a pet pathology auxiliary diagnosis system based on visual language artificial intelligence. The computer vision AI pathological analysis module analyzes the whole pathological section graph, detects a tumor area and identifies key pathological features, the intelligent pathological report generation module automatically generates a pathological diagnosis report based on a large language model by combining an information retrieval enhancement generation (RAG) technology, and the cloud digital section management system provides a remote virtual reading function, supports a pathological expert to access and manage pathological sections anytime and anywhere, integrates AI auxiliary diagnosis and report generation capability, and realizes full-flow information tracking and management. The auxiliary diagnosis system can reduce manual film reading time, optimize pathological resource allocation and improve overall working efficiency.

Inventors

CHANG HONG
LI BIN
CUI ZHANFENG
CAI MINGSHAN

Assignees

杭州泰旸生物科技有限公司

Dates

Publication Date: 20260505
Application Date: 20250902

Claims (11)

1. A pathology report generation method based on a large language model, comprising: Obtaining pathological analysis data of a patient according to an original full-slice image of the patient; Acquiring basic information of a patient; inputting basic information and pathology analysis data of a patient into a trained large language model, and automatically generating a pathology report of the patient, wherein the professional pathology report is searched by a RAG technology and used as a knowledge base of the large language model.
2. The method of claim 1, further comprising obtaining diagnostic information of a pathologist; The basic information, the pathological analysis data and the diagnosis information of a pathologist are input into the trained large language model to automatically generate a pathological report of the patient.
3. The method of claim 1 or 2, further comprising recording the editing content of a pathology expert in the pathology report in response to the editing operation of the pathology report by the pathology expert.
4. A method according to any one of claims 1-3, wherein said obtaining pathological analysis data of the patient from the original whole-slice image of the patient comprises: respectively partitioning the original full-slice image to obtain each patch image block of the original full-slice image; Respectively extracting features from each patch image block of the original full-slice image to obtain feature vectors of each patch image block; Inputting the feature vectors of each patch image block into a trained multi-instance learning neural network model, and outputting attention weights of each patch image block, wherein the attention weights represent the contribution of the feature vectors of each patch image block to a classification result; According to the attention weight of each patch image block, carrying out weighted calculation on the feature vector of each patch image block to obtain the overall feature vector of the original full-slice image, wherein the overall feature vector integrates the features of each patch image block; Inputting the total feature vector of the original full-slice image into a trained linear classifier, outputting to obtain a preliminary classification result of whether a focus and a focus type exist in the original full-slice image, and taking the preliminary classification result as pathology analysis data; the feature vector of each patch image block and the attention weight of each patch image block output by the trained multi-instance learning neural network model are used as graph nodes, the graph nodes are input into the trained graph neural network model, and an attention weight thermodynamic diagram is output and is used for representing the final classification result of the focus area in the original slice image; and determining at least one of focal region occupation area, focal region boundary, focal radius, focal infiltration depth and nuclear mitosis density as pathological analysis data according to the detected focal region.
5. The method of claim 4, wherein prior to pathology classification, a multi-instance learning neural network model and a linear classifier are obtained by: dividing the full-section image of the training sample into patch image blocks of the full-section image of the training sample, wherein the full-section image of the training sample is provided with a section-level label which is used for representing the focus type of the full-section image of the training sample; Respectively extracting features from each patch image block of the training sample full-slice image to obtain feature vectors of each patch image block; inputting the feature vector of each patch image block into a multi-instance learning neural network model to be trained, and outputting to obtain the attention weight of each patch image block; According to the attention weight of each patch image block, weighting calculation is carried out on the feature vector of each patch image block, and the overall feature vector of the training sample full-slice image is obtained; Inputting the total feature vector of the full-section image of the training sample into a linear classifier to be trained, and outputting a preliminary classification result of whether a focus exists in the full-section image of the training sample; And carrying out loss calculation according to the preliminary classification result and the slice-level label of the full-slice image of the training sample, and completing the training process through the gradient of the back propagation loss function.
6. The method of claim 4, wherein prior to pathology classification, a trained graph neural network model is obtained by: the feature vector of each patch image block of the training sample full-slice image and the attention weight of each patch image block output by the multi-instance learning neural network model when training is finished are used as graph nodes, input into the graph neural network model to be trained, and output to obtain the final classification result of the focus area in the training sample full-slice image; And carrying out loss calculation according to the final classification result and the slice-level label of the full-slice image of the training sample, and completing the training process through the gradient of the counter propagation loss function.
7. The method of claim 6, wherein the at least a portion of the training sample full-slice image further has a region-level label for characterizing a lesion area in a patch image block of the training sample full-slice image and is input as a graph node into a graph neural network model to be trained to participate in training.
8. The method of claim 7, wherein the graph neural network model includes a 3-layer graph convolutional layer and a 1-layer attention pooling layer.
9. The method of claim 7, wherein the up to 10% training sample full slice image has a region-level label.
10. A computer device, characterized in that, Comprising a processor, a memory storing machine-readable instructions executable by the processor for executing the machine-readable instructions stored in the memory, which when executed by the processor, perform the steps of the method according to any one of claims 1 to 9.
11. A computer readable storage medium, characterized in that the computer readable storage medium has stored thereon a computer program which, when run by a computer device, performs the steps of the method according to any of claims 1 to 9.

Description

Pet pathology whole-piece image analysis and auxiliary report system based on visual language AI model Technical Field The application belongs to the field of artificial intelligent medical diagnosis, and particularly relates to a pet pathology full-film image analysis and auxiliary reporting system based on a visual language AI model. Background With the rise of health consciousness of pets and the development of veterinary technology, the average life of dogs and cats is gradually prolonged. However, cancer has become one of the major causes of death in dogs and cats. However, the current veterinary pathological diagnosis system has serious unbalance of supply and demand, especially in the pathological diagnosis link, the number of the professional veterinary pathologists is far insufficient, so that the period of pathological detection is long and the efficiency is low, and early detection and accurate treatment of the pet diseases are affected. In the existing veterinary pathological diagnosis process, samples are sent from a veterinary clinician to a pathological laboratory for film making, then the film is read by a pathological expert for diagnosis, and finally the clinical doctor is returned, so that the process is complex and lengthy, usually requires several days to one week, and the requirement of clinical quick decision making is difficult to meet. In addition, the accuracy of pathological diagnosis is highly dependent on the personal experience and training level of pathological specialists, and subjective judgment of different specialists may have great difference, so that diagnosis consistency is affected. Finally, due to the complexity of the contents of the clinical pathology report, animal basic data, medical history, sample description, diagnosis, pathology diagnosis opinion, follow-up treatment advice, references, etc. must be included, and it is time for the report writing process to occupy up to 80%, and for the pathologist to process complex cases and studies. Artificial intelligence (ARTIFICIAL INTELLIGENCE, AI) and machine learning (MACHINE LEARNING, ML) have shown significant advantages in the field of medical image analysis, but applications in the field of veterinary pathology remain in the preliminary stage. The traditional computer vision method faces a plurality of challenges when processing pathological images, and the main reason is that the resolution of the pathological images is extremely high, the data volume of a single image is huge, and the traditional deep learning method cannot be directly used for high-efficiency processing. In addition, the labeling data resources of veterinary pathology are limited, and the application effect of the existing AI method is further limited. At present, no mature intelligent auxiliary system exists in the field of veterinary pathological diagnosis, and the following main defects exist in the prior art: First, the pathology expert resource is short, the diagnosis process is long The number of veterinary pathologists is limited, and the increasing demand for pathological detection of pets is difficult to meet; from the sample sending of veterinary clinicians to the generation of pathological diagnosis reports, the traditional flow involves a plurality of links, so that the diagnosis period is long, and the treatment timeliness is affected; (II) subjective judgment of diagnosis dependent expert, accuracy and consistency are difficult to guarantee The existing pathological diagnosis mainly depends on personal experience of pathological specialists, and is easily influenced by factors such as training background, working state and the like; Different pathologists may have different interpretations for the same case, lack of standardization and consistency, and increase the risk of misdiagnosis and missed diagnosis. (III) the pathologist has heavy workload and high repeated labor occupation ratio The pathologist needs to conduct repeated work such as a large number of film reading, report writing, case comparison and the like, and fatigue and human error are extremely easy to cause; the traditional diagnosis mode can not effectively reduce repeated labor of experts, so that the working efficiency is low, and the overall diagnosis capability is improved. Fourth, the existing computer vision and machine learning methods are difficult to be directly applied to pathological image analysis The resolution of the pathological image is extremely high, the data volume of a single whole pathological section image is huge, the traditional deep learning model is difficult to directly process, and the computational resource requirement is extremely high; the existing computer vision method and vision language model are difficult to accurately identify the pathological region in the pathological image, and lack effective modeling capability for pathological image features; The veterinary pathology data set is deficient, the labeling cost is high, and th