Search

CN-121999994-A - Medical image AI diagnostic analysis method, system and storage medium

CN121999994ACN 121999994 ACN121999994 ACN 121999994ACN-121999994-A

Abstract

The application relates to the technical field of artificial intelligence, discloses a medical image AI diagnosis and analysis method, a system and a storage medium, and aims to solve the problems of unstable image quality, information fusion failure, poor model interpretation, lack of continuous optimization mechanism and the like caused by non-uniform equipment parameters, difficult multi-mode image registration and semantic gap in the prior art. The method comprises the steps of generating personalized imaging parameters based on physiological parameters and anatomical prior models of patients and controlling scanning, carrying out format normalization and metadata verification on DR, CT, MRI images, and then carrying out rigid and non-rigid registration. According to the application, the image acquisition standardization, the multi-mode information complementation fusion and the continuous evolution of the model are realized through the scheme, the focus detection accuracy is obviously improved, the missed diagnosis and misdiagnosis risk is reduced, the report generation time is shortened to be within 38 seconds, the focus area display is enhanced through eye movement tracking, and the film reading efficiency and the clinical decision confidence are improved.

Inventors

  • WANG GUOWEI

Assignees

  • 杭州市萧山区中医院

Dates

Publication Date
20260508
Application Date
20251229

Claims (10)

  1. 1. A medical image AI diagnostic analysis method, comprising the steps of: Acquiring physiological parameters and examination application information of a patient to be examined, calling an anatomical prior model according to an examination part to generate personalized imaging parameter suggestions, and issuing the personalized imaging parameter suggestions to medical imaging equipment to perform scanning; Receiving original projection data or tomographic images from DR, CT, MRI equipment, performing format normalization and metadata verification, marking the modal type, acquisition time and space coordinate information of each frame of image, performing rigid and non-rigid registration on a multi-modal image sequence, optimizing a transformation matrix by using a mutual information maximization criterion, and establishing a voxel level corresponding relation; The method comprises the steps of extracting a combined characteristic map from a registered image set, generating a comprehensive characterization containing structural and functional information through a cross-modal attention fusion network, inputting the comprehensive characterization into a pre-trained deep convolution neural network, executing focus detection, segmentation and classification tasks, and outputting an abnormal region set with coordinates and semantic labels; Carrying out regional enhancement treatment on the original image according to the diagnosis conclusion, and highlighting the key focus and surrounding anatomical structures thereof; Integrating analysis results to generate a structured diagnosis report, and pushing the structured diagnosis report to a hospital information system for examination and confirmation by doctors; Physician feedback comments are collected and correction samples are extracted for iterative updating of the local model.
  2. 2. The medical image AI diagnostic analysis method of claim 1, wherein the anatomical prior model generates a three-dimensional tissue density distribution prediction graph based on the height, weight and body surface projection image of a patient, calculates optimal tube voltage, tube current, exposure time and layer thickness parameter combinations by combining a dose-image quality mapping relation library corresponding to a historical high-quality image, wherein the cross-modal attention fusion network takes an MRI feature graph as a query source, a CT feature graph as a key and value source, calculates attention weights in two dimensions of a channel and a space, the deep convolutional neural network adopts a nnU-Net framework, automatically adjusts the number of input channels of an encoder according to the number of input modes, and enables synchronous batch normalization, the regional enhancement processing adopts a method of combining guide filtering and Laplacian pyramid decomposition, a base layer keeps consistent overall brightness, a detail layer amplifies high-frequency components, the dynamic range of the enhanced image is controlled within 12-bit, the structural diagnostic report conforms to DICOM standard, adopts XML format packaging, comprises basic information of the patient, a checking item, a visible description, an impression conclusion, AI reliability and a value, a score, a local gradient update factor is set as a local gradient update factor, and a local gradient update factor is set as a local gradient update factor of 0, and a local gradient update is set as a local iteration factor of 0, and a local gradient update factor is set as a local iteration factor of 0, and a local gradient update factor is set as a local iteration factor is mm.
  3. 3. The AI diagnostic analysis method of claim 1, wherein the cross-modal attention fusion network constructs four-dimensional feature tensors with dimensions of a spatial X, Y, Z axis and a modal type axis, and the contribution weights of the modalities are adaptively distributed through a channel attention gating fusion unit.
  4. 4. The AI diagnostic analysis method of claim 1, wherein the depth convolutional neural network is optimized by a Dice loss function when performing a lesion segmentation task, an average Dice coefficient of a target lesion is above 0.89, and a small lesion detection sensitivity is not less than 91%.
  5. 5. The AI diagnostic analysis method of claim 1, wherein the regional enhancement process further supports focus region enhancement based on eye tracking, capturing physician gaze point coordinates in real time, constructing a circular enhancement region centered at that point, and the peripheral region preserving the original gray scale distribution.
  6. 6. The AI diagnostic analysis method of claim 1, wherein the local model iterative update employs a federal learning architecture, periodically exchanging encryption gradient information with other nodes and performing a weight fusion operation.
  7. 7. A medical image AI diagnosis analysis system is characterized by comprising an image data acquisition control module, a multi-mode image fusion analysis module, a deep learning diagnosis reasoning engine, an adaptive image enhancement processor, a diagnosis report generation unit and a data interaction and feedback interface, wherein the image data acquisition control module is used for dynamically adjusting exposure parameters and scanning protocols of imaging equipment according to anatomical features and checking positions of a patient, the multi-mode image fusion analysis module is used for carrying out spatial registration and feature level fusion on original images of DR, CT, MRI different modes to generate a unified focus response map, the deep learning diagnosis reasoning engine is used for carrying out focus detection, classification and benign and malignant assessment based on the fused high-dimensional feature map, the adaptive image enhancement processor is used for carrying out local contrast enhancement and noise suppression on a target area according to diagnosis requirements and display terminal characteristics, the diagnosis report generation unit is used for structuralizing and outputting analysis results into an image-text report conforming to clinical specifications, and the data interaction and feedback interface is used for receiving correction labels of radiologists and reversely updating model parameters.
  8. 8. The AI diagnostic imaging analysis system of claim 7, wherein the image data acquisition control module obtains a preliminary positioning image of a patient's height, weight, body contour and scanned location prior to initiating a scan, generates a three-dimensional tissue density distribution prediction map of the location based on a pre-trained anatomical modeling network, and calculates an optimal tube voltage, tube current, exposure time and layer thickness parameter combination in combination with a dose-image quality mapping library corresponding to historical quality images; the multi-mode image fusion analysis module adopts a dual-path feature extraction structure, a first path extracts bone tissue and calcification focus boundaries by using a high-resolution semantic segmentation network based on U-Net++ aiming at CT and DR images, a second path captures soft tissue abnormal signal areas by using a 3D ResNet-34 network aiming at T1/T2 weighted MRI images, and realizes cross-modal space alignment under non-rigid deformation through deformable convolution, a deep learning diagnosis reasoning engine is internally provided with a three-level reasoning architecture, the first level is a candidate area rapid screener, a light-weight MobileNetV3 model is adopted to carry out sliding window scanning on full-view images, an initial positioning frame of suspicious nodules or lesions is output, the second level is a refined classifier, after each candidate area is cut, the fine grain texture analysis is carried out on Inception-v4 networks, inflammation, benign hyperplasia and malignant tumor categories are distinguished, the third level is a benign and malignant medical record discriminator, age, gender and family history variables in clinical electronic medical record fusion are associated through modeling multisource information of a graph neural network, BI RADS or ng-RADS is finally output, the self-adaptive enhancement processor adopts a local histogram equalization histogram classification algorithm, the diagnosis report generating unit is embedded with a natural language generating template library covering common examination items such as chest CT, abdomen MRI, skeleton DR, etc., can automatically convert the detected focus number, maximum diameter, position description, density/signal characteristics, adjacent structure infringement conditions into standardized text expressions, and appends AI confidence index and recommended follow-up period, the data interaction and feedback interface is configured with an online increment learning unit, the system automatically extracts key feature vectors of the sample and marks correction labels after modifying and confirming AI diagnosis results by doctors, adds a local fine adjustment data set after differential privacy treatment, periodically executes federal learning parameter aggregation, the image data acquisition control module integrates a dynamic dose compensation mechanism, monitors X-ray penetration signal intensity change in real time in a continuous scanning process, and automatically triggers millisecond-level parameter fine adjustment when detecting a tissue density mutation region.
  9. 9. The medical image AI diagnostic analysis system of claim 7, wherein the multi-modality image fusion analysis module constructs a four-dimensional feature tensor after spatial registration is completed and adaptively assigns weights for each modality contribution through a channel attention-gated fusion unit to promote weights of CT in skeletal lesions and MRI in brain glioma margin definitions.
  10. 10. The AI diagnostic imaging analysis system of claim 7, wherein the adaptive image enhancement processor supports eye tracking based focal region enhancement, captures physician gaze point coordinates in real time via an eye tracker coupled to a studio display terminal, builds a circular enhancement region centered at that point, and the peripheral region retains the original gray scale distribution.

Description

Medical image AI diagnostic analysis method, system and storage medium Technical Field The invention belongs to the technical field of artificial intelligence, and particularly relates to a medical image AI diagnostic analysis method, a system and a storage medium. Background Along with the deep fusion of artificial intelligence and medical imaging technology, the utilization of AI to assist doctors in disease screening, diagnosis and efficacy assessment has become an important direction for the development of modern intelligent medical treatment. Medical imaging is used as a core means for information acquisition in clinical diagnosis and treatment, is widely applied to various fields of radiology, oncology, neurology and the like, and the data sources cover various imaging modes, including X-ray digital imaging (DR), computed Tomography (CT) and Magnetic Resonance Imaging (MRI), and the technologies can provide complementary physiological and pathological information from the aspects of structure, function and metabolism, so that the method has important significance for improving diagnosis accuracy. Among them, AI diagnostic analysis based on multi-modal medical images has received attention in recent years. The technology aims at automatically extracting focus features under different imaging modes through a deep learning model and realizing cross-modal information fusion and joint reasoning so as to support more comprehensive and accurate clinical decisions. Especially in the tasks of early lesion detection, lesion boundary segmentation, benign and malignant discrimination and the like, the AI algorithm has the potential of approaching to or even exceeding human expert, and becomes one of key enabling technologies for promoting accurate medical landing. The prior art still has a series of key bottlenecks in the aspect of AI diagnosis and analysis of medical images, namely firstly, the model types of image acquisition equipment are various, imaging parameters are not uniform, so that input data are obviously different in resolution, contrast, noise level and the like, the model generalization capability is seriously affected, secondly, space registration difficulty and semantic gap problems exist among different modality images, the traditional fusion strategy is difficult to effectively align an anatomical structure and a focus area, information redundancy or loss is caused, and thirdly, the traditional diagnosis model mostly adopts an end-to-end training mode, lacks explicit modeling of clinical priori knowledge, causes poor interpretation of an inference process, is difficult to obtain doctor trust, and finally, in actual deployment, a plurality of systems independently operate in single equipment or departments, lacks dynamic tracking and longitudinal analysis capability of full-period image data of patients, and cannot support continuous optimization of personalized diagnosis and treatment schemes. The problems seriously restrict the reliable application and large-scale popularization of the AI technology in complex clinical scenes. Disclosure of Invention The invention aims to make up the defects of the prior art, and provides a medical image AI diagnosis and analysis method, a system and a storage medium, which can effectively solve the problems in the background art. In the application of modes such as DR, CT, MRI and the like of the current medical image, structural technical contradictions such as obvious influence of equipment parameters and individual differences of patients, high missed diagnosis and misdiagnosis rate caused by focus identification depending on doctor experience, difficulty in realizing focus cross-sequence consistency analysis by multi-mode image information isolation processing, limited diagnosis efficiency caused by low automation degree of image post-processing exist. The invention realizes collaborative optimization in three dimensions of image acquisition control, intelligent diagnosis decision and image enhancement processing by constructing an end-to-end intelligent analysis architecture, and remarkably improves the image quality stability, focus detection accuracy and clinical diagnosis efficiency. The medical image AI diagnosis analysis system comprises an image data acquisition control module, a multi-mode image fusion analysis module, a deep learning diagnosis reasoning engine, an adaptive image enhancement processor, a diagnosis report generation unit, a data interaction and feedback interface, a diagnosis report generation unit and a model parameter updating unit, wherein the image data acquisition control module is used for dynamically adjusting exposure parameters and scanning protocols of imaging equipment according to anatomical features and examination parts of patients; Preferably, the image data acquisition control module acquires the height, weight, body shape profile and preliminary positioning image of the scanned part of the patient before starting