An AI technology-based postoperative vision intelligent prediction and evaluation system for cataract patients

Abstract

This invention discloses an AI-based intelligent prediction and assessment system for postoperative vision in cataract patients, comprising an information acquisition module, a macular OCT image acquisition module, an image analysis module, a model training module, a vision prediction module, and a treatment suggestion module. The information acquisition module collects basic information about the individual user; the macular OCT image acquisition module acquires macular OCT images of the patient; the image analysis module automatically extracts and analyzes the external morphological features of the macula in the images acquired by the macular OCT image acquisition module; the model training module trains the vision prediction module; the vision prediction module loads the optimal model parameters obtained from the model training module and uses them to predict the postoperative vision of cataract patients; the treatment suggestion module provides targeted treatment suggestions based on the predicted postoperative vision obtained from the vision prediction module. This invention can accurately and quickly predict the postoperative vision of cataract patients based on their preoperative macular OCT images.

Description

Technical Field

[0001] This invention belongs to the field of medical image processing, and in particular relates to an AI-based intelligent prediction and assessment system for postoperative vision of cataract patients. Background Technology

[0002] Cataracts are the leading cause of blindness worldwide. With increasing life expectancy and a growing aging population, the incidence of cataracts is rising annually. Surgery is the only effective treatment for cataracts. Most patients recover good vision after cataract surgery. However, some patients, due to complex fundus diseases, do not experience any improvement in vision after surgery.

[0003] Computer vision technology is frequently used in the field of rapid intelligent image processing, such as image classification, object detection, and object retrieval. Computer vision simulates human visual mechanisms and has advantages such as high detection speed and low cost. In recent years, with the application of deep learning in computer vision, especially in the field of medical imaging, breakthroughs have been achieved. The traditional image interpretation mode, which relied on doctors manually interpreting images, has been broken. Data-driven deep learning technology allows computers to assist in the discovery of lesions and improve diagnostic accuracy by combining imaging and medical image processing techniques with computer analysis and computation. Based on this, computer-aided diagnostic systems have been built, and in many diseases, the level of diagnosis has reached or even surpassed that of human doctors.

[0004] For example, Chinese patent document CN112016634A discloses a medical image recognition method, including: inputting a medical image into a disease grading network, obtaining a category activation map output by the disease grading network, along with its disease category and disease confidence, wherein the category activation map can characterize the relevant region in the medical image indicating the corresponding disease category, and the classification of the disease category is related to one or more lesions; inputting the medical image into a pathological sign recognition network, obtaining one or more lesion probability maps output by the pathological sign recognition network, wherein each pixel of each lesion probability map indicates the probability that a corresponding sub-region in the medical image includes a lesion, and when the corresponding disease confidence is greater than a preset confidence level, the similarity between the category activation map and each lesion probability map in the related one or more lesion probability maps is greater than a set threshold.

[0005] The macula is the most sensitive part of the retina and a crucial factor in determining postoperative visual acuity in cataract patients. Optical coherence tomography (OCT) is a non-invasive retinal imaging method that helps ophthalmologists qualitatively and quantitatively assess changes in the microstructure of the macular region. However, most clinicians currently rely on empirical analysis of OCT images to roughly assess macular function, and there is currently no standardized AI assessment system based on a large sample size to accurately predict postoperative visual acuity in cataract patients.

[0006] Predicting visual acuity before cataract surgery can help doctors make informed surgical decisions for patients whose postoperative vision may not improve, reducing their financial burden and avoiding unnecessary waste of medical resources. It also helps patients adjust their expectations appropriately, preventing disputes. Therefore, there is an urgent need for a system that can efficiently and accurately predict postoperative visual acuity in cataract patients. Summary of the Invention

[0007] This invention provides an AI-based intelligent prediction and assessment system for postoperative visual acuity in cataract patients. It can accurately and quickly predict postoperative visual acuity in cataract patients based on preoperative macular OCT images, providing important information for clinicians to make reasonable surgical decisions for cataract patients and promoting the rational and effective use of medical resources.

[0008] A postoperative vision intelligent prediction and assessment system for cataract patients based on AI technology includes a computer memory, a computer processor, and a computer program stored in the computer memory and executable on the computer processor. The computer memory contains an information acquisition module, a macular OCT image acquisition module, an image analysis module, a model training module, a vision prediction module, and a treatment suggestion module.

[0009] The information collection module is used to collect basic information of individual users, including name, gender, age, eye type, and preoperative visual acuity.

[0010] The aforementioned macular OCT image acquisition module is used to acquire macular OCT images of the patient;

[0011] The image analysis module is used to automatically extract and analyze the external morphological features of the macula in images acquired by the macula OCT image acquisition module. Specifically, it removes image noise through fast denoising, filters discrete white dots through dilatation and erosion, performs edge detection using the Canny operator, performs macula ILM layering, performs angle calibration using the boundary lines of the layered ILM, performs RPE layering using the U-Net segmentation network after calibration, and extracts external morphological features using the extracted ILM and RPE layers.

[0012] The model training module is used to train the vision prediction module. It takes the external morphological features of the macula extracted by the image analysis module, the image, and the preoperative vision as input, and uses the postoperative vision as a label to train the deep learning model.

[0013] The vision prediction module loads the optimal model parameters obtained from the model training module and uses them to predict the postoperative vision of cataract patients.

[0014] The aforementioned treatment suggestion module is used to provide different targeted treatment suggestions based on the predicted postoperative visual acuity obtained from the visual acuity prediction module.

[0015] Furthermore, when the macular OCT image acquisition module acquires images of the patient's macula, it can clearly and intuitively understand the cellular structure and lesion distribution of each layer of the retina in the macular region of the fundus, and clearly display the morphology, size, and boundaries of various lesions in the macular region, which has the advantages of being simple and fast.

[0016] Furthermore, in the image analysis module, the U-Net segmentation network includes 8 convolutional layers and pooling layers, of which 4 layers are used for downsampling and the remaining 4 layers are used for upsampling.

[0017] Furthermore, the extracted external morphological features of the macula in the image analysis module include: maximum temporal retinal thickness, maximum temporal slope, foveal thickness, maximum nasal retinal thickness, maximum nasal slope, pit depth, and foveal diameter.

[0018] Furthermore, the vision prediction module consists of two parts based on a deep learning algorithm; the first part is a convolutional neural network, which is used to extract global features from OCT images; the second part is an attention fusion network, which is used to fuse the embeddings of each modality, including macular external morphological features, image features and preoperative vision features, and performs modality fusion through a Transformer layer to predict postoperative vision.

[0019] Furthermore, the first part of the convolutional neural network includes four stages, each stage consisting of three residual blocks, and each residual block consisting of two 1x1 convolutions, one 3x3 convolution, one pooling layer, and one activation layer.

[0020] Furthermore, the attention fusion network in the second part includes 20 Transformer layers and 1 Head layer. Each Transformer layer consists of two merged layers, one attention layer, and one feedforward network layer; the Head layer consists of two fully connected layers.

[0021] Furthermore, when the model training module trains the vision prediction module, it uses SGD to optimize the model in order to minimize the root mean square error (RMSE) to calculate the loss function; the maximum number of training iterations is set to 1200, and the initial learning rate is set to 0.01.

[0022] The original dataset was randomly divided into three parts in a 6:2:2 ratio: training set, validation set, and test set. After each training round, the model was validated using the validation set. After all training rounds were completed, the final trained model was tested using the test set to evaluate the performance of the classification model.

[0023] Furthermore, the treatment recommendation module determines whether cataract surgery is necessary by comparing the postoperative visual acuity of the cataract patient obtained by the visual acuity prediction module with the preoperative visual acuity collected by the information collection module. The judgment criterion is: surgery is recommended when the postoperative visual acuity improves by ≥0.2 logMAR.

[0024] Compared with the prior art, the present invention has the following beneficial effects:

[0025] 1. This invention extracts and analyzes the features of the patient's preoperative macular OCT image, and then uses a deep learning algorithm to train the model, which can accurately and quickly predict the postoperative vision of cataract patients and provide different targeted treatment suggestions.

[0026] 2. This invention can quickly and easily predict the degree of postoperative visual recovery of cataract patients based on preoperative macular OCT images. It uses a deep learning artificial intelligence system for model training and constructs a treatment suggestion module, providing important information for clinicians to make reasonable surgical decisions for cataract patients and promoting the rational and effective use of medical resources. Attached Figure Description

[0027] Figure 1 This is an overall flowchart of an AI-based intelligent prediction and assessment system for postoperative vision in cataract patients according to the present invention.

[0028] Figure 2 This is a schematic diagram of the U-Net network structure in the image analysis module of this invention;

[0029] Figure 3 This is a schematic diagram illustrating the extraction of external morphological features of the macula.

[0030] Figure 4 This is a schematic diagram illustrating how the model training module trains the vision prediction module in this invention.

[0031] Figure 5 This is a schematic diagram of the diagnostic and treatment suggestion module in this invention. Detailed Implementation

[0032] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be noted that the embodiments described below are intended to facilitate the understanding of the present invention and do not constitute any limitation thereof.

[0033] like Figure 1 As shown, an AI-based intelligent vision prediction system for cataract patients after surgery includes an information acquisition module, a macular OCT image acquisition module, an image analysis module, a model training module, a vision prediction module, and a treatment suggestion module.

[0034] The information collection module is used to collect basic information about individual users, including: name, gender, age, eye type, and preoperative visual acuity. The macular OCT image acquisition module is used to acquire macular OCT images of patients, which are obtained through OCT equipment. This equipment can be from brands such as Heidelberg, Zeiss, and Colin, which are commonly used in clinical settings.

[0035] The image analysis module extracts the morphology of the fovea centralis. First, it uses the `fastNlMeansDenoising` function in OpenCV for denoising. Then, it uses dilation and erosion operations for boundary enhancement and discrete white point removal. Next, it uses the Canny algorithm for edge detection to obtain the ILM boundary. After obtaining the detected ILM boundary, it calibrates to make the ILM boundaries parallel, and then uses a trained U-Net network for RPE boundary segmentation. Figure 2 As shown, the U-Net network first downsamples the image for feature encoding, then introduces jump links during the upsampling process to copy the downsampled features before upsampling again, concatenating the features across multiple dimensions. After upsampling, it predicts the foreground and background for each pixel. Finally, through joint analysis of the ILM and RPE layers, it calculates macular features, including five specific indicators: maximum nasal retinal thickness, maximum temporal retinal thickness, foveal depth, diameter, and thickness.

[0036] like Figure 3 As shown, the morphological features of the fovea are automatically extracted by the algorithm, and five marker points A, B, C, D, and E are automatically identified. Points A and E represent the points where the slope of the ILM layer is 0 on the temporal and nasal sides, respectively; point C is the point where the slope of the fovea pit is 0; and points B and D represent the points where the slope of the ILM layer is maximum on the temporal and nasal sides, respectively. Thus, five basic features of the fovea are extracted: the maximum thickness of the nasal retina, the maximum thickness of the temporal retina, the depth, diameter, and thickness of the fovea.

[0037] like Figure 4As shown, the model training module, based on the foveal-specific morphological values ​​extracted by the image analysis module, uses a convolutional neural network to extract global image features at multiple scales, focusing more on the lesion area. The extracted image features from the two views are mapped to 256 dimensions, and combined with a preoperative visual acuity feature (also mapped to 256 dimensions) and a macular external morphological feature to form a 4*256 dimensional vector, which is then fused through an attention fusion network. The attention network consists of multiple Transformers, each learning multimodal features through a self-attention mechanism. This self-attention mechanism captures the dependencies and correlations between different global modalities, resulting in higher model accuracy. Deep fusion of multimodal features is achieved through multiple Transformers, using preoperative visual acuity and external morphological features to guide image feature learning. The learned features are aggregated and condensed by a classification head, and the final result is predicted by the Head layer based on the classification head. The visual acuity prediction module loads the model parameters obtained from the model training module into the model, enabling prediction of postoperative visual acuity after inputting an OCT image.

[0038] like Figure 5 As shown, the treatment recommendation module determines whether cataract surgery is necessary by comparing the postoperative visual acuity of cataract patients obtained from the visual acuity prediction module with the preoperative visual acuity collected by the information collection module. The judgment criterion is: surgery is recommended when the postoperative visual acuity improves by ≥0.2 logMAR.

[0039] The embodiments described above provide a detailed explanation of the technical solutions and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the present invention. Any modifications, additions, and equivalent substitutions made within the scope of the principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An AI technology-based intelligent postoperative vision prediction and evaluation system for cataract patients, comprising a computer memory, a computer processor, and a computer program stored in the computer memory and executable on the computer processor, characterized in that: The computer memory contains an information acquisition module, a macular OCT image acquisition module, an image analysis module, a model training module, a vision prediction module, and a treatment suggestion module; ​ The information collection module is used to collect basic information of individual users, including name, gender, age, eye type, and preoperative visual acuity. The aforementioned macular OCT image acquisition module is used to acquire macular OCT images of the patient; The image analysis module is used to automatically extract and analyze the external morphological features of the macula in the images acquired by the macula OCT image acquisition module. Specifically, it removes image noise through fast denoising, filters discrete white dots through dilatation and erosion, performs edge detection through the Canny operator, performs macula ILM layering, performs angle calibration through the boundary lines of the layered ILM, performs RPE layering using the U-Net segmentation network after calibration, and extracts external morphological features through the extracted ILM and RPE layers. The extracted external morphological features of the macula include: maximum temporal retinal thickness, maximum temporal slope, foveal thickness, maximum nasal retinal thickness, maximum nasal slope, pit depth, and foveal diameter; the U-Net segmentation network consists of 8 convolutional layers and pooling layers, of which 4 layers are used for downsampling and the remaining 4 layers are used for upsampling; The model training module is used to train the vision prediction module. It takes the external morphological features of the macula extracted by the image analysis module, the image, and the preoperative vision as input, and the postoperative vision as the label for deep learning model training. The vision prediction module consists of two parts based on deep learning algorithms. The first part is a convolutional neural network, which is used to extract global features from OCT images. The second part is the attention fusion network, which fuses the embeddings of each modality, including macular external morphological features, image features, and preoperative visual acuity features. Modality fusion is performed through Transformer layers, and postoperative visual acuity is predicted. The convolutional neural network in the first part includes four stages, each stage consisting of three residual blocks. Each residual block consists of two 1x1 convolutions, one 3x3 convolution, one pooling layer, and one activation layer. The attention fusion network in the second part includes 20 Transformer layers and one Head layer. Each Transformer layer consists of two merged layers, one attention layer, and one feedforward network layer. The Head layer consists of two fully connected layers. The vision prediction module loads the optimal model parameters obtained from the model training module and uses them to predict the postoperative vision of cataract patients. The aforementioned treatment suggestion module is used to provide different targeted treatment suggestions based on the predicted postoperative visual acuity obtained from the visual acuity prediction module. 2.The AI technology-based postoperative vision intelligent prediction and evaluation system for cataract patients according to claim 1, wherein When the model training module trains the vision prediction module, it uses SGD to optimize the model in order to minimize the root mean square error (RMSE) to calculate the loss function; the maximum number of training iterations is set to 1200, and the initial learning rate is set to 0.

01.

3. The AI-based intelligent prediction and assessment system for postoperative vision of cataract patients according to claim 1, characterized in that, The aforementioned treatment recommendation module, based on the postoperative visual acuity of cataract patients obtained by the visual acuity prediction module, compares it with the preoperative visual acuity collected by the information collection module to determine whether cataract surgery is necessary. The judgment criteria are as follows: Surgery is recommended when postoperative visual acuity improves by ≥0.2 logMAR.

Images