CN-121982416-A - OCTA-based hypertension and small vascular lesion recognition model and system

CN121982416ACN 121982416 ACN121982416 ACN 121982416ACN-121982416-A

Abstract

The invention discloses an OCTA-based hypertension and small vascular disease identification model and system, which belong to the technical field of medical image identification and comprise the steps of acquiring OCTA images, synchronously researching association of key retinal parameters with hypertension pathological states and small vascular disease pathological states by combining vascular structure characteristics and neurovascular characteristics of different retinal areas, and establishing a complete framework from OCTA imaging parameter optimization, retinal image acquisition, characteristic extraction, hypertension diagnosis, small vascular disease diagnosis and system microvascular burden index construction. The invention provides a high-efficiency intelligent solution for collaborative screening of hypertension and small vascular diseases, and has important significance for improving the awareness rate and the control rate of hypertension and small vascular diseases in China and reducing the burden of cardiovascular diseases.

Inventors

LIU XIUYUN
Guo Yimai
WANG BOYANG
GENG JIE
He Runnan
MIN HANYI
ZHENG LIUYING
LIU XUANCHANG
JI YONG

Assignees

天津大学

Dates

Publication Date: 20260505
Application Date: 20260204

Claims (10)

1. An OCTA-based hypertension and small vascular disease identification model is characterized by comprising the following steps: s1, data preprocessing and feature selection, namely acquiring unified standard OCTA images of a macular area and a optic disc area, constructing multi-mode input data by combining structured clinical and scale information, screening the data, and ensuring data consistency and comparability; s2, model building and training, namely building a deep neural network based on multi-mode fusion and multi-task joint learning MMTL, wherein the deep neural network sequentially comprises an input and coding layer, a hypertension prediction layer, a multi-mode fusion layer, a multi-task prediction layer, a systematic index layer and a training and optimizing module; S3, detecting and evaluating, namely inputting the preprocessed multi-mode input data into a trained deep neural network, outputting four-system risk probabilities and hypertension stage results including heart system micro-vascular lesions, brain system micro-vascular lesions, kidney system micro-vascular lesions and skin micro-vascular lesions, and calculating a systematic micro-vascular burden index SMBI based on the four-system risk probabilities so as to realize quantitative evaluation of the damage degree of systemic microcirculation.
2. The identification model of hypertension and small vascular lesions based on OCTA according to claim 1, wherein the specific contents in S2 are as follows: S21, designing an input and coding layer, namely building three parallel input branches which respectively correspond to OCTA image features, OCTA quantitative features and clinical and scale features; S22, constructing a hypertension prediction layer, namely taking coding embedded vectors of three input branches as input, adopting a multi-mode fusion structure to realize interactive learning, constructing an ordered classification module, and outputting a hypertension stage result and a hypertension potential representation vector; s23, constructing a multi-mode fusion layer, namely adopting a gating weighted fusion structure, receiving an OCTA image embedded vector, an OCTA quantitative feature embedded vector, a clinical and scale embedded vector and a hypertension potential representation vector, generating a mode weight vector through vector splicing, linear gating and Sigmoid normalization, and calculating to obtain a high-dimensional fusion characterization vector ; S24, designing a multi-task prediction layer, namely setting four parallel prediction heads which respectively correspond to four-system lesion risk predictions including heart, brain, kidney and skin, wherein each prediction head adopts a specified three-layer full-connection structure, a middle layer uses a ReLU activation function, an output layer adopts a Sigmoid function to output probability values, and adopts a training mode of sharing fusion layer parameters and joint optimization; S25, constructing a systematic index layer, namely calculating a unified systematic microvascular burden index SMBI based on a four-system lesion risk prediction result, wherein the unified systematic microvascular burden index SMBI is used for quantifying the damage degree of the systemic microcirculation of an individual; s26, formulating a training and optimizing strategy, namely adopting binary cross entropy with Sigmoid as a basic loss function, and constructing a combined objective function comprising four-system prediction loss and SMBI auxiliary supervision loss.
3. The model for identifying hypertension and small vascular lesions based on OCTA as claimed in claim 2, wherein in S1, the grouping corresponding to the multi-mode input data comprises heart system micro vascular lesions, brain micro vascular lesions, kidney system micro vascular lesions, skin micro vascular lesions and health control groups, wherein the signal intensity threshold screening, contrast enhancement, noise suppression, eye axis length and diopter correction, layer segmentation rechecking and artifact quality control are sequentially carried out on all OCTA images; The characteristics of the multi-mode input data comprise main OCTA characteristics and auxiliary covariates, and the characteristics are as follows: The heart microvascular pathological change group is characterized by main OCTA (OCTA) of central concave choroid thickness SFCT and paracardial superficial capillary plexus blood vessel density SCPVD, and auxiliary covariates of body mass index BMI, age, sex and systolic pressure SBP; the main OCTA characteristic is ganglion cell layer volume GCL and outer plexiform layer volume OPL, and the auxiliary covariates are MoCA score, age, sex and SBP of Montreal cognitive assessment scale; The main OCTA characteristics of the kidney microvascular lesion group are vascular density VD, perfusion density PD of the superficial capillary plexus SVP macular region and thickness pRNFL of retinal nerve fiber layer around the optic disc, and the auxiliary covariates are diabetes course, age, sex and SBP; The skin microvascular lesion group is characterized by the area of the avascular zone of the SCP macula of the superficial capillary plexus and the area of the avascular zone of the DCP macula of the deep capillary plexus, and the auxiliary covariates are diabetes mellitus course, age, sex and SBP.
4. The model of identifying hypertension and small vascular lesions based on OCTA as claimed in claim 3, wherein in S21, the OCTA image characteristic branch adopts EFFICIENTNET-B0 as a main network, is input as a multi-channel image of the macula area and the optic disc area, including shallow capillary plexes, deep capillary plexes and choroidal capillary layers, which are subjected to noise suppression, signal intensity standardization and layer segmentation review, and an SE attention mechanism is embedded between a shallow layer and a deep layer convolution module of the main network to enhance the response capability of the model to critical vascular textures, perfusion modes and boundaries of non-perfusion areas to obtain representative image characteristics, and an image embedding vector is output after multi-layer convolution, pooling and global average pooling; the input of OCTA quantitative characteristic branches is continuous parameters including blood vessel density, perfusion density, non-perfusion area, retinal nerve fiber layer thickness and choroid thickness, and a three-layer fully-connected network is adopted; the clinical and scale characteristic branches are input into structural variables including body mass index, systolic pressure, age, gender, diabetes course and Montreal cognitive scale Score, and after Z-Score standardization and independent heat coding, the structural variables are input into two layers of fully connected networks, and embedded vectors are output.
5. The model for identifying hypertension and small vascular lesions based on OCTA according to claim 4, wherein in S22, the hypertension prediction layer specifically performs interactive learning in a joint feature space by adopting a multi-mode fusion structure, captures a coupling mode between fundus microcirculation change and blood pressure state, inputs the fused features into an ordered classification module, outputs a hypertension stage result, and simultaneously generates a hypertension potential representation vector, wherein the hypertension potential representation vector synthesizes the comprehensive contribution of OCTA image structure, blood vessel density features and clinical parameters to blood pressure change.
6. The OCTA-based model for identifying hypertension and small vascular lesions as claimed in claim 5, wherein in S23, the joint features obtained by vector stitching are expressed as Wherein Embedding vectors into OCTA images, Embedding OCTA quantitative features, Is embedded with a clinical scale, The formula for the high-dimensional fusion is as follows: ; ; Wherein, the The function is activated for Sigmoid, As a matrix of weights, the weight matrix, As a result of the bias term, Is a modal weight vector.
7. The identification model of hypertension and small vascular lesions based on OCTA according to claim 6, wherein in S25, SMBI is calculated according to the formula: ; Wherein the method comprises the steps of For the non-negative constraint weight coefficient, the risk increase of each system is ensured to be increased to be SMBI and accord with physiological interpretation, SMBI is mapped to 0-100 minutes after Sigmoid normalization and divided into three steps of low risk, medium risk and high risk, and the three steps are used for quantitatively representing the whole microvascular burden level of an individual.
8. The identification model of hypertension and small vascular lesions based on OCTA according to claim 7, wherein in S26, the joint objective function is: ; Wherein, the 、、、 Loss values of four system lesion prediction tasks respectively, Is SMBI to assist in supervising the loss values, 、、、、 A weight coefficient for each loss term; The model training optimization process adopts an optimizer AdamW, a cosine annealing learning rate scheduling strategy is introduced to improve convergence stability, dropout and BatchNorm are introduced to each layer of the model to prevent overfitting, and one of E Isotonic regression and Platt scaling calibration is adopted to enhance probability consistency and clinical interpretability of a prediction result.
9. The hypertension and small vascular disease identification system based on OCTA is characterized by comprising a data preprocessing and feature selecting module, a deep neural network module and a result output module; the data preprocessing and feature selecting module is used for acquiring OCTA images with unified standards, constructing multi-mode input data by combining the structured clinical and scale information and screening the data; The deep neural network module is a deep neural network based on multi-modal fusion and multi-task joint learning MMTL, and sequentially comprises an input and coding layer, a hypertension prediction layer, a multi-modal fusion layer, a multi-task prediction layer, a systematic index layer and a training and optimizing module, wherein the training and optimizing module is used for receiving preprocessed multi-modal input data, realizing feature extraction, hypertension staged prediction, multi-modal feature fusion, multi-system small vessel lesion risk prediction, systematic microvascular burden quantification and training and optimizing the deep neural network module by adopting a joint objective function, and improving model prediction precision and generalization capability; The result output module is used for outputting the risk probability of heart system microvascular lesions, brain system microvascular lesions, kidney system microvascular lesions, skin microvascular lesions, the hypertension stage result and the systemic microvascular burden index SMBI, and realizing quantitative evaluation and display of the damage degree of systemic microcirculation.
10. The system for identifying hypertension and small vascular diseases based on OCTA according to claim 9, wherein the deep neural network module realizes hierarchical modeling, and a modeling link comprises multi-mode input data, input and coding layer feature extraction, hypertension prediction layer output stage result and hypertension potential expression vector, multi-mode fusion layer generation fusion feature vector, multi-task prediction layer output four-system disease risk probability, systemic index layer output SMBI, training and optimization module performs training optimization, and complete all-link assessment from OCTA microstructure features to hypertension state to systemic small vascular diseases.

Description

OCTA-based hypertension and small vascular lesion recognition model and system Technical Field The invention relates to the technical field of medical image recognition, in particular to an OCTA-based hypertension and small vascular lesion recognition model and system. Background Cardiovascular health is critical to human quality of life and longevity, and hypertension and high incidence of multisystem small vessel disease have become global significant public health challenges. The early screening rate of the small blood vessel diseases is low, the diagnosis rate in the asymptomatic period is less than 20%, a large number of patients are in an asymptomatic state for a long time due to lack of early screening means, and diagnosis is not confirmed until target organs are damaged, so that the treatment difficulty and the disease burden are obviously increased. The traditional hypertension diagnosis mainly uses cuff type sphygmomanometer to measure peripheral arterial blood pressure, and has obvious defects that firstly, the operation is regulated by an operator to be matched with a patient, the operation is easily influenced by environment (blood vessel contraction caused by low room temperature), physiological state (emotional tension, non-return after exercise) and equipment errors (the cuff size is not matched), the actual blood pressure level is difficult to reflect due to the fact that the average value is needed to be measured for multiple times, secondly, only an instantaneous blood pressure value can be provided, 24-hour blood pressure fluctuation can not be captured, and the identification capability of 'hidden hypertension' and 'white overcoat hypertension' is insufficient, and thirdly, the operation is limited for infants, obese patients and patients with severe arteriosclerosis, and the measurement accuracy is poor. The traditional detection means of the small blood vessel diseases are also limited in that the kidney small blood vessel diseases (DKD, CKD) depend on functional indexes, invasive pathology and combination of conventional images, the early sensitivity of the biological indexes is poor and lags behind the micro blood vessel injury, the kidney puncture biopsy is invasive, the application range is limited, the resolution of conventional imaging (ultrasonic/CT/MRI) is low, the micro blood vessel structure cannot be estimated, the brain small blood vessel diseases (CSVD) are mainly related by image screening, nerve function estimation and dangerous factors, the early sensitivity of the skull MRI is insufficient, microscopic injury cannot be quantized, the nerve function scale inquires about symptom lag, the specificity is low, whether the cognitive disorder is caused by the brain small blood vessel diseases or not is difficult to distinguish, the heart small blood vessel diseases (HFrEF, HFpEF) depend on functional estimation, indirect image and clinical relevance, the coronary artery functional estimation (CFR/IMR) is invasive and low in popularity, the heart functional image (ultrasonic/CMR/SPECT) cannot directly estimate the micro blood vessels, the early sensitivity is poor, only can indirectly infer, the skin small blood vessel diseases (diabetes is sufficient) is not invasive screened, and the diabetes is especially the diabetes is a certain lag after the invasive examination. The traditional diagnosis methods for the two diseases have the problems of strong subjectivity, low early sensitivity, incapability of quantifying microscopic damage and the like, and are difficult to meet the clinical early-stage accurate diagnosis requirement. Disclosure of Invention The invention aims to provide an OCTA (optical coherence tomography) based model and system for identifying hypertension and small vascular lesions, which can avoid subjective deviation and environmental interference of traditional blood pressure measurement and discover target organ damage caused by early hypertension by objectively capturing microscopic structural changes of retina, and simultaneously solve the problem that the traditional detection of small vascular lesions cannot directly quantify early micro vascular damage, so as to realize parallel identification of small vascular lesions. In order to achieve the above purpose, the invention provides an OCTA-based hypertension and small vascular disease identification model, which comprises the following steps: s1, data preprocessing and feature selection, namely acquiring unified standard OCTA images of a macular area and a optic disc area, constructing multi-mode input data by combining structured clinical and scale information, screening the data, and ensuring data consistency and comparability; s2, model building and training, namely building a deep neural network based on multi-mode fusion and multi-task joint learning MMTL, wherein the deep neural network sequentially comprises an input and coding layer, a hypertension prediction layer, a multi-mode fusion layer, a