CN-121837277-B - Retinal image abnormality intelligent recognition system based on ophthalmology auxiliary diagnosis

CN121837277BCN 121837277 BCN121837277 BCN 121837277BCN-121837277-B

Abstract

The invention discloses an ophthalmic auxiliary diagnosis-based retina image anomaly intelligent recognition system, which belongs to the technical field of medical image processing analysis, and is characterized in that a plurality of groups of images of the same detected eye are acquired and form an analysis sequence, an anatomic structure feature image and a tissue texture feature image of each frame of images in the sequence are extracted in parallel, conditional normal texture construction and intra-group comparison, inter-multi-frame texture consistency verification and anatomic structure deviation quantitative analysis are performed on the anatomic structure feature image and the tissue texture feature image, a texture anomaly intensity image, a texture space-time variation image, a texture region stability classification image and a structure deviation degree image are generated, and the multiple types of results are subjected to self-adaptive weighted fusion and hierarchical judgment.

Inventors

WU TONG
WANG RUI
WU XIAOCHANG
ZHAO MENGJIA

Assignees

中国人民解放军总医院海南医院

Dates

Publication Date: 20260508
Application Date: 20260313

Claims (8)

1. The retina image abnormity intelligent recognition system based on the ophthalmologic auxiliary diagnosis is characterized by comprising an image acquisition and processing module, a control module and a control module, wherein the image acquisition and processing module is used for controlling imaging equipment to acquire a plurality of groups of retina images when the same eye to be inspected executes at least two different preset eyeball actions, and carrying out standardized preprocessing on the plurality of groups of retina images so as to form an analysis sequence; The dual-path characteristic decoupling extraction module is used for receiving an analysis sequence, synchronously inputting a pre-trained anatomical structure characteristic extraction branch and a pre-trained tissue texture characteristic extraction branch in parallel, correspondingly outputting an anatomical structure characteristic image and a tissue texture characteristic image of each frame of image in the analysis sequence, and forming an anatomical structure characteristic image sequence and a tissue texture characteristic image sequence; The multiple comparison analysis module is used for receiving the anatomical structure feature map sequence and the original tissue texture feature map sequence, executing conditional normal texture construction and intra-group comparison analysis, multi-frame texture consistency verification analysis and anatomical structure deviation quantization analysis, and generating an intra-frame texture abnormal strength map, an inter-frame texture uncontrolled variation map, a texture region stability classification map and a structure deviation map according to multiple comparison analysis results; The information fusion checking module is used for receiving the multi-class image data, carrying out self-adaptive weighted fusion to obtain a comprehensive abnormal intensity image, and carrying out grading judgment and positioning on the focus based on collaborative analysis of the comprehensive abnormal intensity image, the texture region stability classification image and the structure deviation image.
2. The system for intelligently identifying retinal image abnormalities based on ophthalmic assisted diagnosis as recited in claim 1, characterized in that an anatomical feature extraction branch is constructed based on a pre-trained retinal image segmentation network for extracting anatomical features including vascular morphology, optic disc geometry, fovea location, and a tissue texture feature extraction branch is trained by self-supervised learning for extracting texture features decoupled from the anatomical structure and representing tissue texture, pigment, reflection, hemorrhage, and pigmentation.
3. The retinal image abnormality intelligent identification system based on the ophthalmic auxiliary diagnosis according to claim 2, wherein the specific process of the conditional normal texture construction and the intra-group comparison analysis comprises the following steps: Inputting an anatomical structure feature map of a t frame image in the anatomical structure feature map sequence as a condition, generating a network through a pre-training condition, reconstructing a predicted normal tissue texture feature map corresponding to the current anatomical structure state, and calculating pixel-by-pixel differences between the tissue texture feature map of the frame image and the predicted normal tissue texture feature map to obtain an original difference map; The original difference image is mapped into an intra texture anomaly intensity image with the pixel value range of [0,1] through a convolution layer with the scale of 1×1 and a Sigmoid activation function.
4. The retinal image abnormality intelligent identification system based on the ophthalmic auxiliary diagnosis according to claim 3, wherein the specific process of the inter-multiframe texture consistency verification analysis includes: performing cross-frame cooperative comparison on the tissue texture feature map sequence, and calculating the texture feature similarity of any two frames of images in the sequence at corresponding spatial positions to obtain a plurality of inter-frame similarity matrixes; Obtaining the median, mean or other aggregation statistics of a plurality of interframe similarity matrixes, and generating interframe texture uncontrolled variation graphs representing the overall variation intensity of textures at all positions; Based on a plurality of inter-frame similarity matrixes, classifying and marking each spatial position as a stable texture area and a transient texture area according to a preset threshold and a continuous frame number judging rule, and generating a texture area stability classification diagram.
5. The intelligent recognition system of retinal image anomalies based on the ophthalmic assisted diagnosis according to claim 4, wherein the process of acquiring the transient texture region and the stable texture region comprises: In the similarity matrix among the frames, when the similarity of the texture features of the N frames (N is more than or equal to 2) on the corresponding space positions is larger than a first preset threshold value, the corresponding space positions are marked as stable texture regions, and when the similarity of the texture features of the N frames on the corresponding space positions is smaller than a second preset threshold value, the corresponding space positions are marked as transient texture regions, wherein the first preset threshold value is larger than the second preset threshold value.
6. The intelligent recognition system for retinal image anomalies based on the ophthalmic assisted diagnosis according to claim 5, wherein the specific process of the quantitative analysis of the deviation of the anatomical structure includes: And (3) comparing the anatomical structure feature map of each frame of image with a pre-stored standard normal anatomical structure feature template pixel by pixel in a multi-channel manner, carrying out linear weighted fusion on the quantized deviation values of all channels, and mapping the quantized deviation values to a structure deviation degree map with the pixel values in a [0,1] interval through a nonlinear function.
7. The system for intelligently identifying retinal image anomalies based on the ophthalmic auxiliary diagnosis according to claim 6, wherein the specific process of the information fusion verification module for adaptively weighting and fusing comprises the following steps: s1, constructing a characteristic tensor, namely splicing an intra-frame texture abnormal intensity image, an inter-frame texture uncontrolled variation image, a texture region stability classification image and a structure deviation degree image along a channel dimension to form a multi-source evidence characteristic tensor with four-dimensional evidence vectors; S2, self-adaptive weight generation, namely inputting the multi-source evidence characteristic tensor into a self-adaptive weight generation network based on a spatial attention mechanism, and dynamically generating self-adaptive weight tensors corresponding to each pixel position one by one; S3, carrying out weighted fusion and initial confidence calculation, namely carrying out pixel-level weighted summation on four types of evidence by using an adaptive weight tensor to generate a comprehensive abnormal intensity map; S4, grading and positioning focus based on stability, namely carrying out joint judgment on the comprehensive abnormal intensity image and the texture region stability classification image and the structural deviation image according to a preset threshold rule, and marking and positioning focuses of different grades.
8. The retinal image abnormality intelligent identification system based on the ophthalmic auxiliary diagnosis according to claim 7, wherein the specific process of "stability-based lesion classification and localization" in S4 includes: S4.1, judging a high-confidence focus, namely judging a pixel marked as a stable texture region in a texture region stability classification diagram as the high-confidence focus if the corresponding value of the pixel in the comprehensive abnormal intensity diagram is larger than a preset first judging threshold value; S4.2, judging a focus to be rechecked, namely judging a pixel marked as a transient texture region in a texture region stability classification diagram as the focus to be rechecked if the corresponding value of the pixel in the comprehensive abnormal intensity diagram is larger than a preset second judging threshold value and the corresponding value of the pixel in a structure deviation diagram is larger than a preset third judging threshold value, wherein the second judging threshold value is smaller than the first judging threshold value; And S4.3, marking suspected artifacts, namely marking the pixels marked as transient texture areas in the texture area stability classification diagram as the suspected artifacts if the corresponding values in the comprehensive abnormal intensity diagram are larger than a preset second judgment threshold value, but the corresponding values in the structure deviation diagram are smaller than a preset third judgment threshold value.

Description

Retinal image abnormality intelligent recognition system based on ophthalmology auxiliary diagnosis Technical Field The invention relates to the technical field of medical image processing analysis, in particular to an ophthalmic auxiliary diagnosis-based retinal image abnormality intelligent recognition system. Background With the deep application of artificial intelligence technology in the field of medical image analysis, automatic analysis of retina images (such as fundus color photographs and optical coherence tomography images) based on deep learning has become a research hotspot for ophthalmologic auxiliary diagnosis, and the aim is to automatically identify pathological changes in images, such as diabetic retinopathy, glaucoma, age-related macular degeneration and the like, so as to improve screening efficiency and assist clinical decisions. The conventional technical scheme mainly follows the following two types of paradigms: the supervised end-to-end disease classification model is that the method relies on a large-scale and finely marked disease image data set to train a deep convolutional neural network to directly output specific disease labels or severity grades; The core limitation of the paradigm is that firstly, the performance is severely limited by the scale and quality of the labeling data, the cost of acquiring medical expert labeling is high and the period is long, and secondly, the model is essentially in learning the distribution of existing disease labels, the unknown anomaly types which do not occur or are rare in the training data lack identification capability, and the generalization is limited. Unsupervised or self-supervised anomaly detection model to reduce reliance on anomaly labeling data, such methods attempt to define a "normal" pattern by learning a large number of feature distributions of normal retinal images, and then determine samples that deviate from that pattern as anomalies. This paradigm, while reducing reliance on abnormal labeling, has the disadvantage that the generic image reconstruction model is difficult to integrate into the prior knowledge of the anatomy specific to the retina (e.g., vessel tree morphology, anatomical locations of optic disc and macula), resulting in insensitivity to structural aberrations (e.g., enlarged optic disc pits), or false positives to normal anatomical variations due to significant physiological individual differences in retinal anatomy, and such a retrieval scheme can misjudge normal structural variations as abnormal textures, resulting in low specificity. In addition, the two paradigms are analyzed for a single Zhang Jingtai image, and dynamic information and consistency constraints contained in a multi-view image sequence acquired by eyeballs in different gazing directions cannot be utilized, so that noise and artifact interference in a single frame image are easy to occur, and the specificity is required to be improved. In summary, the core contradiction faced by the current retinal image anomaly identification technology lies in how to construct an intelligent identification system capable of simultaneously sensing structural distortion and fine texture changes in a sharp manner and providing an anatomic interpretable result on the premise of reducing the dependence on large-scale anomaly labeling data. Disclosure of Invention The invention aims to solve the existing problems, and provides an ophthalmic auxiliary diagnosis-based retinal image abnormality intelligent recognition system compared with the prior art. The invention aims at realizing the technical scheme that the retina image abnormity intelligent identification system based on the ophthalmologic auxiliary diagnosis comprises an image acquisition and processing module, a control module and a control module, wherein the image acquisition and processing module is used for controlling imaging equipment to acquire a plurality of groups of retina images when the same eye to be inspected executes at least two different preset eyeball actions, and carrying out standardized pretreatment on the plurality of groups of retina images so as to form an analysis sequence; The dual-path characteristic decoupling extraction module is used for receiving an analysis sequence, synchronously inputting an anatomical structure characteristic extraction branch and a tissue texture characteristic extraction branch which are pre-trained and optimized in parallel, correspondingly outputting an anatomical structure characteristic image and a tissue texture characteristic image of each frame of image in the analysis sequence, and forming an anatomical structure characteristic image sequence and a tissue texture characteristic image sequence; The multiple comparison analysis module is used for receiving the anatomical structure feature map sequence and the original tissue texture feature map sequence, executing conditional normal texture construction and intra-group comparison analysis, multi