CN-121999268-A - Medical image focus automatic identification system based on deep learning

CN121999268ACN 121999268 ACN121999268 ACN 121999268ACN-121999268-A

Abstract

The invention discloses a medical image focus automatic identification system based on deep learning, which relates to the technical field of medical image intelligent analysis, and aims to support access of CT, MRI and X-ray equipment, adapt to various image formats and calibrate to keep physical meaning consistent, treat image artifacts and unify the size of a region of interest by adopting a denoising-normalizing-cutting three-stage process, extract local and global features by adopting a multi-branch CNN and transducer mixed model based on large-scale labeling dataset training, output focus category, size, position and malignancy risk level, support structural report, DICOM labeling and visual display, periodically update the model by adopting a layered storage architecture, monitor system operation indexes and trigger alarm when the indexes are abnormal. The invention improves the consistency of image processing and the accuracy of focus identification, optimizes risk judgment by combining clinical information, ensures data privacy and system quality, and has high diagnosis efficiency and reliability.

Inventors

GUO PENG
HUANG YAQI
LU JIALE
YIN XIAOLU
LI XIAOQIN

Assignees

湖南工程学院

Dates

Publication Date: 20260508
Application Date: 20251210

Claims (10)

1. The medical image focus automatic identification system based on deep learning is characterized by comprising: The image acquisition adaptation module supports the access of three medical image devices, namely CT, MRI and X-ray, receives real-time image data through a hardware interface, and simultaneously performs device calibration on images output by different devices; The image preprocessing module adopts a denoising-normalization-clipping three-stage processing flow, wherein a denoising link integrates Gaussian filtering, median filtering and non-local mean filtering, metal artifact removal is additionally carried out on a CT image, motion artifact correction is carried out on an MRI image based on an iterative reconstruction algorithm, and a compensation range is-pi to pi based on phase compensation; the deep learning model module adopts a CNN+transducer mixed architecture to extract image features, captures global feature association and dynamically distributes weight of the fusion module, wherein model training is based on a medical image dataset, comprises tumor, inflammation and nodular lesions, and adopts AdamW optimizers for training iteration times of 100-200 rounds; The focus positioning and classifying module is used for mapping to focus probability distribution through a full-connection layer based on 2048-dimensional feature vectors output by the deep learning model, adopting non-maximum suppression to screen candidate focus areas, generating focus thermodynamic diagrams, wherein the confidence level corresponding to the color depth is reduced, and the confidence level of red, yellow and green is reduced; The result output and interaction module comprises a structured report, a DICOM labeling file, a visual interface display and interaction function, wherein the structured report comprises a focus list, description and suggestion, and is used for automatically inserting a thermodynamic diagram and an image screenshot; And the quality control module monitors the system operation index in real time, triggers an alarm mechanism when the index exceeds a threshold value, slightly alarms and only prompts an interface, seriously alarms and triggers a short message to inform an administrator, and meanwhile, automatically switches to a standby model, namely a historical optimal model, and generates a quality analysis report.
2. The automatic recognition system of medical imaging lesions based on deep learning of claim 1, further comprising: The image enhancement module is used for performing self-adaptive enhancement processing on low-quality images, including X-ray low-dose images and MRI low-signal-to-noise ratio images, separating an image illumination component and a reflection component by adopting a Retinex algorithm, improving detail contrast of focus areas by combining histogram equalization, and performing super-resolution reconstruction on micro focus in a CT image.
3. The automatic recognition system of medical imaging lesions based on deep learning of claim 1, further comprising: The multi-mode image fusion module is used for receiving the mode images of the same patient, wherein the mode images comprise CT+MRI and X-ray+CT, different mode images are aligned to the same coordinate system through a spatial registration algorithm, the spatial registration algorithm is optimized based on mutual information, a fusion link adopts an attention weighted fusion strategy, the weight of each mode image in a focus area is calculated, CT image weight density characteristics and MRIT2 weight image weight soft tissue characteristics are weighted, a fusion image is generated through pixel-level weighted summation, a fusion formula is F (X, y) =w1×CT (X, y) +w2×MRI (X, y), w1+w2=1, w1 and w2 are dynamically calculated by the attention module, and the fusion image retains the advantage characteristics of each mode.
4. The automatic recognition system of medical imaging lesions based on deep learning of claim 1, further comprising: the focus region confidence evaluation module is used for calculating the overall confidence based on the pixel probability distribution in the focus region of interest, wherein the calculation formula is as follows Wherein C is the confidence coefficient of a focus area, S is the area of the focus interested area, f (x, y) is the probability value of the focus belonging to the pixel point (x, y) in the interested area, dxdy is the area infinitesimal of the pixel point, when C is more than or equal to 0.8, the focus with high confidence coefficient is judged to be directly included in a result list, when C is more than or equal to 0.5 and less than or equal to 0.8, the focus with medium confidence coefficient is judged to be marked in a report to be combined with clinical confirmation, and when C is less than or equal to 0.5, the focus with low confidence coefficient is judged to be automatically marked as the suspicious area to be manually checked by an administrator.
5. The automatic recognition system of medical imaging lesions based on deep learning of claim 1, further comprising: The clinical information association module is used for interfacing with the hospital electronic medical record system through an HL7FHIR interface to obtain clinical data of patients, including ages, sexes, medical histories and laboratory examination results such as tumor markers CEA, AFP values, blood sugar and blood fat indexes, establishing a mapping rule of the clinical data and focus characteristics in an association analysis link, and automatically inserting related clinical data in a structural report.
6. The automatic recognition system of medical imaging lesions based on deep learning of claim 1, further comprising: the model iteration optimization module adopts a gradient descent method to minimize a model loss function to realize dynamic parameter updating, and the optimization is as follows And (3) wherein θt is a model parameter in the t-th iteration, and comprises CNN branch weight, a transducer branch attention matrix and fusion module weight, θt+1 is a model parameter after the t+1th iteration, and η is learning rate self-adaptive adjustment.
7. The automatic recognition system of medical imaging lesions based on deep learning of claim 1, further comprising: The false positive focus filtering module is used for filtering the false positive region based on clinical rules and image features, extracting gray level co-occurrence matrix features of the false positive region, including contrast and relativity, by adopting texture analysis, comparing the gray level co-occurrence matrix features with a true focus feature library, reducing the false positive rate of a system after filtering by more than or equal to 1.5%, generating a false positive analysis log containing filtering basis and rule matching results, and supporting an administrator to optimize the rule library.
8. The automatic recognition system of medical imaging lesions based on deep learning of claim 1, further comprising: the focus growth trend prediction module predicts focus volume change in future time period based on focus volume data of patient examination, and the prediction formula is that V (T) is the predicted focal volume on the T day, V 0 is the focal volume of the last examination, T is the predicted time length, tau is the integral variable, R (tau) is the focal volume change rate at tau moment, the fitting model is obtained based on historical volume data fitting, the fitting model adopts linear regression, namely the volume uniform velocity change or exponential regression, namely the volume acceleration change, the fitting goodness R2 is more than or equal to 0.9, the predicted result is output as a volume change curve and trend judgment, the time axis of the volume change curve is T, and the Y axis is volume.
9. The automatic recognition system of medical imaging lesions based on deep learning of claim 1, further comprising: The remote consultation docking module supports docking with each center remote consultation platform, adopts real-time audio and video communication, simultaneously transmits patient image data, focus identification results and clinical information, supports breakpoint continuous transmission for DICOM format of the patient image data, enables each expert to mark focus real-time synchronous marking content in the consultation process, enables a main expert in a shared operation interface to display different images and results through a control system, records consultation video content, automatically adds a time stamp and consultation staff information to be stored in the data storage module, automatically generates consultation reports after consultation is finished, and comprises expert opinion summarization, final diagnosis conclusion and treatment suggestion, and is confirmed by a participant expert electronic signature.
10. The automatic recognition system of medical imaging lesions based on deep learning of claim 1, further comprising: The privacy protection module is used for realizing collaborative training of all central data by adopting a federal learning architecture to avoid transmission of original data across mechanisms, each medical institution is used as federal node to support 10-100 nodes to participate simultaneously, a local training model uses own labeling data to not upload the original data, only uploading model parameters is encrypted by adopting a homomorphic encryption algorithm Paillier, a federal server aggregates all node parameters by adopting weighted average, the weight is the node data volume duty ratio, the higher the data volume is, global model parameters are generated and then sent to all nodes, all nodes update the local model by using the global parameters, iteration is carried out until the global model accuracy rate converges to achieve continuous 3 rounds of lifting <0.5%, meanwhile, desensitization processing is carried out on patient data, patient ID is replaced by a random character string, and face/name identification information is removed from images.

Description

Medical image focus automatic identification system based on deep learning Technical Field The invention relates to the technical field of intelligent analysis of medical images, in particular to an automatic medical image focus recognition system based on deep learning. Background In the current medical image focus identification process, the image acquisition link has obvious problems of equipment compatibility and data consistency. In clinic, CT, MRI, X-ray and other image equipment are complicated in brand and model, the difference of image formats output by different equipment is large, some old equipment only supports a non-standard format, so that data access is difficult, even if the equipment supports a universal format, the output image physical parameters of the equipment also lack uniform calibration, for example, the HU value range of the CT equipment can deviate due to different manufacturer settings, the signal intensity of the MRI has no fixed reference standard, and the accuracy of the subsequent focus feature extraction is directly affected. Meanwhile, the processing capability of the low-quality image is insufficient, the signal-to-noise ratio of the X-ray low-dose image is low, MRI motion artifacts are obvious, CT metal artifacts are seriously interfered, the existing preprocessing method mostly adopts a single filtering algorithm, various noises and artifacts are difficult to eliminate in a targeted manner, the detail of a tiny focus is covered, and the missed diagnosis risk is increased. The deep learning model has the advantages of limited performance and insufficient clinical suitability, and further restricts the focus recognition effect. Most of the existing systems adopt a single CNN architecture, can only capture local image features, have insufficient global association analysis on focuses and surrounding tissues, and particularly have low accuracy rate on focus identification across anatomical regions, have limited data set scale on which model training depends, have extremely low rare focus sample ratio, and cause weak recognition capability of the model on the focuses, and have high false positive rate. In addition, the focus classification and risk level determination only depend on image features, and the risk level misdetermination may be caused by not incorporating critical factors such as smoking history and age into the malignant risk assessment of the pulmonary nodule, which may affect clinical decision. Meanwhile, the model updating mechanism is rigidified, the newly added data needs to be trained in full quantity again, time and effort are consumed, and the model updating mechanism is difficult to adapt to the continuously accumulated case data and the newly discovered focus type in clinic. Contradiction of privacy protection and multi-center collaboration, lack of quality control, and insufficient support for remote consultation are also important obstacles to system landing. Medical data privacy requirements are strict, multi-center data are difficult to directly share, so that the model cannot cooperatively train by utilizing multi-mechanism data, performance improvement is limited, and partial systems lack of perfect quality control mechanisms, and when the success rate of image acquisition is reduced and the accuracy rate of the model fluctuates, standby schemes cannot be timely alarmed and switched, so that misdiagnosis can be caused. In addition, the high-quality medical resources of the primary hospital are deficient, the existing system is difficult to dock a remote consultation platform, the focus identification capability of the superior hospital cannot be extended to the primary hospital, focus missed diagnosis and misdiagnosis rate of the primary hospital are high, and patients are difficult to obtain timely and accurate diagnosis. Disclosure of Invention The invention provides a medical image focus automatic identification system based on deep learning, which aims to solve the problems in the prior art. In order to achieve the purpose, the invention adopts the following technical scheme that the medical image focus automatic identification system based on deep learning comprises: The image acquisition adaptation module supports the access of three medical image devices, namely CT, MRI and X-ray, receives real-time image data through HDMI2.1 and USB3.2 interfaces, and performs device calibration on images output by different devices; The image preprocessing module adopts a denoising-normalization-clipping three-stage processing flow, a denoising link integrates Gaussian filtering, median filtering and non-local mean filtering, performs metal artifact removal additionally on a CT image, performs motion artifact correction on an MRI image based on an iterative reconstruction algorithm, and compensates a range-pi to pi based on phase compensation; The deep learning model module adopts a CNN+transducer mixed architecture to extract image features, captures g