A method for predicting NAION and distinguishing acute stage based on deformable convolution and multi-site OCTA

Abstract

This invention relates to a method for NAION prediction and acute phase differentiation based on deformable convolution and multi-site OCTA, belonging to the field of ophthalmic disease diagnosis and medical image processing technology. This invention acquires OCTA images of ocular samples and extracts data from key sites. Using the OCTA deformable convolution feature extraction module as the core feature extraction backbone network, its optimal performance is verified through single-site experiments, and a multi-site joint diagnostic framework is constructed. Radiomics features are introduced, and deep fusion of deformable convolution features and radiomics features is achieved through a bidirectional cyclic feature interaction module. Finally, a hybrid expert module is added to the three-site joint model to decouple NAION prediction from the acute phase differentiation task of NAION / ON, improving diagnostic accuracy. This invention effectively solves the problems of difficult accurate differentiation between the acute phases of NAION and ON and insufficient early prediction of NAION, providing reliable technical support for ophthalmic clinical diagnosis.

Description

Technical Field

[0001] This invention relates to a method for predicting and differentiating NAION in the acute phase based on deformable convolution and multi-site OCTA, belonging to the interdisciplinary field of ophthalmic disease diagnostic technology and medical image processing. Specifically, it is applicable to the early prediction of non-arteritis anterior ischemic optic neuropathy (NAION) and the accurate differentiation of NAION and ON in the acute phase. Background Technology

[0002] Non-arteritis anterior ischemic optic neuropathy (NAION) and optic neuritis (ON) are both common acute optic nerve diseases in clinical practice. Their acute symptoms are similar (both present with sudden visual loss and visual field defects, such as...). Figure 1 (As shown in the image), but the pathogenesis and treatment differ significantly: NAION is mostly caused by optic nerve head ischemia, requiring priority to improve local blood supply; ON is mostly related to autoimmunity, requiring immunosuppressive therapy. If the diagnosis is confused, it can easily lead to treatment delays, or even worsen optic nerve damage.

[0003] Current clinical diagnosis mainly relies on fundus examination, visual field examination, and optical coherence tomography (OCT), but there are limitations: fundus examination is difficult to quantify changes in optic nerve head blood flow; OCT can only reflect retinal structure and cannot capture abnormalities in deep tissues such as the choroidal layer and the radial capillary layer around the optic disc; and NAION patients have a potential risk of developing the disease in the contralateral healthy eye, making early prediction difficult with current technology.

[0004] Optical coherence tomography (OCTA) can non-invasively acquire vascular texture in multiple parts of the eye, providing a new dimension for disease diagnosis. OCTA analysis of acute NAION and acute ON reveals retinal vessel rupture and reduced blood flow leading to vascular reduction, providing a basis for disease differentiation. NAION is a sudden loss of vision, making it difficult to capture the fundus condition of the affected eye before onset. However, the contralateral healthy eye of the affected eye is more likely to be involved in NAION, thus we used the contralateral healthy eye of a NAION patient to simulate the fundus condition before the onset of NAION.

[0005] Existing OCTA-based diagnostic methods mostly employ single-site approaches or traditional convolutional neural networks, which have two main shortcomings: First, the fixed receptive field of traditional convolutional networks cannot adapt to the morphological differences of different parts of the eye, resulting in limited feature extraction accuracy. Second, they do not fully integrate multi-site information and radiomics features, and fail to decouple the "NAION prediction" and "NAION / ON differentiation" tasks, leading to low multi-classification accuracy. Therefore, there is an urgent need for a method that can combine multi-site OCTA, optimize feature extraction, and decouple diagnostic tasks to improve the accuracy of NAION prediction and acute-phase differentiation. Summary of the Invention

[0006] In view of the shortcomings and deficiencies of existing technologies, such as difficulty in distinguishing the acute phase of NAION from ON and low early prediction accuracy of NAION, this invention proposes a method for NAION prediction and acute phase differentiation based on deformable convolution and multi-site OCTA. This method effectively integrates multi-dimensional ocular imaging information and quantitative features, significantly improves diagnostic accuracy, and provides technical support for precise clinical diagnosis and treatment.

[0007] The technical solution of this invention is: a method for NAION prediction and acute phase differentiation based on deformable convolution and multi-site OCTA, comprising:

[0008] OCTA images of four types of eye samples were collected: normal healthy eyes, contralateral healthy eyes of NAION patients, eyes in the acute phase of ON, and eyes in the acute phase of NAION. Data from three key areas were extracted: the choroidal layer, the radial capillary layer around the optic disc, and the superficial retina.

[0009] Using the OCTA deformable convolutional feature extraction module ODCFI Block as the core feature extraction backbone network, after verifying its optimal performance through single-site experiments, a multi-site joint diagnostic framework was constructed.

[0010] Extract radiomics features and achieve deep fusion of deformable convolutional features and radiomics features through the bidirectional cyclic feature interaction module BCFI Block;

[0011] Finally, a hybrid expert module is added to the three-part joint model to decouple NAION prediction from the NAION / ON acute phase differentiation task, and the final classification result is output through sub-task collaborative optimization.

[0012] Furthermore, the method includes the following steps:

[0013] Step 1. Collect OCTA image data of the eye and classify the sample types. The samples include four categories: normal healthy eye, healthy eye on the opposite side of NAION patient, eye in the acute phase of ON, and eye in the acute phase of NAION. For each type of sample, OCTA images of three parts are extracted: choroidal layer, radial capillary layer around the optic disc, and superficial retina.

[0014] Step 2. Preprocess the single-part OCTA images acquired in Step 1 and divide them into training and test sets;

[0015] Step 3. Use masking techniques to extract radiomics features from the OCTA images in Step 1;

[0016] Step 4. Construct the deformable convolutional feature extraction module ODCFIBlock for OCTA based on the DCN verified in Step 2, and build a multi-part joint diagnostic framework: After each OCTA of different parts passes through the ODCFI Block, features are fused through a bidirectional cyclic feature interaction module to achieve unified classification on a single OCTA of a single part; based on single-part OCTA, the number of OCTA parts is continuously increased to achieve joint unified classification of two parts and three parts; during the feature extraction process, the features extracted by each part through the ODCFI Block need to interact with the features of other parts through a bidirectional cyclic feature interaction module to achieve information exchange between OCTAs on different parts, so as to provide more accurate features for subsequent classification;

[0017] Step 5. Add a hybrid expert module to decouple the four-class classification task into the NAION prediction subtask and the NAION / ON acute phase molecular task. The final classification result is output through the collaborative optimization of the subtasks.

[0018] Furthermore, in Step 2, the specific preprocessing includes:

[0019] Step 2.1. ROI delineation: Based on the anatomical structure of the eye OCTA, professional radiologists manually delineate the choroidal layer, the radial capillary layer around the optic disc, and the vascular distribution area of ​​the superficial retina to ensure accurate positioning of the target area.

[0020] Step 2.2. Contrast Enhancement: The CLAHE algorithm is used to enhance the details of blood vessel texture. The algorithm block size is set to 8×8 and the contrast limit is set to 2.0 to improve the distinction between blood vessels and background.

[0021] Step 2.3. Zero-mean normalization: using the formula:

[0022]

[0023] To eliminate differences in image grayscale distribution, the formula is as follows: This represents the pixel value at (x, y) in the original image. The average gray level of the image. The standard deviation of grayscale Minimum value to avoid To prevent calculation errors from occurring, ensure data distribution consistency.

[0024] Furthermore, in Step 4, the processing flow of the OCTA deformable convolutional feature extraction module is as follows:

[0025] The input features are first processed by 3×3 deformable convolution and batch normalization (BN), and then the first residual is added to the original input to retain the initial features.

[0026] The summed features are then subjected to a 3×3 deformable convolution and a ReLU activation function, and then added to the first residual output for a second residual summation to obtain the final feature output.

[0027] The present invention also provides a NAION prediction and acute phase differentiation system based on deformable convolution and multi-site OCTA, the system comprising: a module for performing the NAION prediction and acute phase differentiation method based on deformable convolution and multi-site OCTA.

[0028] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the NAION prediction and acute phase differentiation method based on deformable convolution and multi-site OCTA.

[0029] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the NAION prediction and acute phase differentiation method based on deformable convolution and multi-site OCTA.

[0030] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the NAION prediction and acute phase differentiation method based on deformable convolution and multi-site OCTA.

[0031] The beneficial effects of this invention are:

[0032] 1. More accurate feature extraction: The ODCFI Block is constructed using deformable convolution, and its adaptive receptive field can...

[0033] It adapts to the morphological differences of different parts of the eye, and its single-part classification AUC is better than traditional networks such as ResNet18 and VIT, with a maximum improvement of 0.068. It can efficiently capture fundus vascular abnormalities and lay a high-quality feature foundation for diagnosis.

[0034] 2. Full utilization of information from multiple parts: Information interaction between parts is achieved through a multi-part joint framework and FI Block. The AUC of the three-part joint model reaches 0.955, which is 0.074 higher than that of a single part. This makes up for the shortcomings of the one-sided information of a single part and solves the problem of difficulty in capturing deep tissue anomalies.

[0035] 3. High efficiency in cross-modal feature fusion: BCFI Block achieves deep fusion of deformable convolutional features and 91 radiomics features, improving AUC by up to 0.014 compared to direct stitching, transforming visual information into quantitative data, and supplementing the diagnosis with objective dimensions.

[0036] 4. Task decoupling improves diagnostic performance: The hybrid expert module breaks down the four-class classification task, specifically addressing the difficulties in clinical identification. The final model achieves a four-class classification accuracy of 0.912 and an AUC of 0.972, especially improving the eye classification accuracy in the acute phase of ON, providing support for early intervention and precision treatment of NAION. Attached Figure Description

[0037] Figure 1 This is a comparative image of the four types of OCTA of the present invention in three locations (choroidal layer, perioptic disc radial capillary layer, and superficial retina).

[0038] Figure 2 The ROC curves for the single-site model comparison experiment of this invention (choroidal layer, radial capillary layer around the optic disc, and superficial retina).

[0039] Figure 3 This is a schematic diagram of the standard convolution and deformable convolution (DCN) principles of this invention;

[0040] Figure 4 This is a structural diagram of the model framework of the present invention;

[0041] Figure 5 This is a schematic diagram of the ODCFI module and FI module of the present invention;

[0042] Figure 6 The confusion matrix and ROC curve of the model of this invention under the unified classification and hybrid expert modes respectively;

[0043] Figure 7 This is a flowchart of the method of the present invention. Detailed Implementation

[0044] Example 1: As Figures 1-7 As shown, a method for NAION prediction and acute phase differentiation based on deformable convolution and multi-site OCTA includes the following specific steps:

[0045] Step 1: First, four types of ocular OCTA samples were collected, with clear collection criteria for each type: Normal healthy eyes must have no history of eye disease and exclude underlying conditions such as hypertension and diabetes; the contralateral healthy eye of NAION patients must be the asymptomatic eye of patients in the acute phase of NAION, with no abnormalities on fundus examination; the acute phase eye of ON (optic neuritis) must be a clinically diagnosed eye with an acute phase of optic neuritis ≤14 days; the acute phase eye of NAION must be a clinically diagnosed eye with an acute phase of non-arteritis anterior ischemic optic neuropathy ≤14 days. OCTA images of three regions were extracted for each type of sample: the choroid, the peridiscal radial capillary layer (RPC), and the superficial retina. The image resolution was uniformly set to 400×400 pixels and stored in PNG format. The number of samples collected for each type is shown in Table 1.

[0046] Table 1. Sample data for various types of OCTA

[0047] As Step 1 in this invention, its core is to ensure the standardization and consistency of the samples, providing a reliable data foundation for subsequent model training and validation. The OCTA comparison charts of the four types of samples at three locations are shown below. Figure 1 As shown, the differences in images between different samples can be observed intuitively.

[0048] Step 2: The collected single-part OCTA images were preprocessed and the dataset was divided. The optimal feature extraction network was verified through model comparison experiments. Preprocessing included ROI delineation, contrast enhancement, and zero-mean normalization. The dataset was partitioned using a hierarchical random partitioning strategy, splitting the four classes into training (80%) and test (20%) sets in an 8:2 ratio. Subsequently, eight models—ResNet18, VGGNet, VIT, SwinTransformer, DenseNet, ShuffleNet, ConvNeXt_Tiny, and Deformable Convolutional Network (DCN)—were used to train the four-class classification on the three parts of the OCTA images. Accuracy (ACC) and area under the curve (AUC) were used as evaluation metrics to verify model performance. The experimental results are shown in Table 2, and the ROC curves for each part are shown in Table 2. Figure 2 As shown.

[0049] Table 2. Classification results of commonly used deep learning networks and deformable convolutional networks on OCTA in various parts.

[0050] As Step 2 in this invention, the specific preprocessing steps are as follows:

[0051] Step 2.1. ROI delineation: Based on the anatomical structure of the eye OCTA, professional radiologists manually delineate the choroidal layer, the radial capillary layer around the optic disc, and the vascular distribution area of ​​the superficial retina to ensure accurate positioning of the target area.

[0052] Step 2.2. Contrast Enhancement: The CLAHE algorithm is used to enhance the details of blood vessel texture, with the algorithm block size set to 8×8 and the contrast limit set to 2.0, to improve the distinction between blood vessels and the background;

[0053] Step 2.3. Zero-mean normalization: using the formula:

[0054]

[0055] To eliminate differences in image grayscale distribution, the formula is as follows: This represents the pixel value at (x, y) in the original image. The average gray level of the image. The standard deviation of grayscale Minimum value to avoid To prevent calculation errors from occurring, ensure data distribution consistency.

[0056] The experimental environment was set as follows: Intel(R) Core(TM) i7-10700F CPU @ 2.90GHz, 64.0 GB of memory, and NVIDIA GeForce RTX3060 graphics card (8GB VRAM). The optimizer used was SGD, and the learning rate decay strategy was to update the learning rate every 30 epochs with a multiplication factor of 0.1. Table 2 shows that DCN achieved the best ACC and AUC in all three regions (e.g., the ACC of the superficial retinal DCN was 0.806, and the AUC was 0.883). Its adaptive receptive field can better adapt to the morphological differences of different parts of the eye. The principle diagrams of standard convolution and deformable convolution are shown below. Figure 3 As shown, DCN was therefore chosen as the core feature extraction network.

[0057] Step 3: Next, radiomics features are extracted and preprocessed. Professional ophthalmologists use masking techniques to extract radiomics features from all parts of the OCTA images of the four types of samples. These features encompass 91 items across 7 categories, including diagnostic-related features, first-order statistical features, and gray-level co-occurrence matrix features. The number of features in each category is shown in Table 3.

[0058] Table 3. Types and quantities of radiomics features extracted by OCTA from different body parts.

[0059] After standardizing the extracted features, they are used to construct a fusion scheme of "DCN image features + radiomics features". The fusion effect of feature stitching (Concat) and bidirectional cyclic feature interaction module (BCFI Block) is compared through ablation experiments to provide the optimal solution for subsequent feature fusion.

[0060] Step 4: Subsequently, construct a multi-site joint diagnostic framework: using the DCN validated in Step 2 as the core, form the OCTA deformable convolutional feature extraction module (ODCFI Block), the module structure of which is as follows: Figure 5 As shown, the process is as follows: the input features are first processed by 3×3 deformable convolution and batch normalization (BN), and then the first residual is added to the original input to retain the initial features; the added features are then processed by 3×3 deformable convolution and ReLU activation function, and the second residual is added to the first residual output to obtain the final feature output. In the experiment of combining single-site OCTA with radiomics features, three classification schemes were adopted based on DCN: "OCTA features only", "OCTA features + radiomics features (Concat)" and "OCTA features + radiomics features (BCFI Block)". The experimental results are shown in Table 4.

[0061] Table 4 Ablation experiments based on deformable convolutional networks in single-site OCTA combined with radiomics features

[0062] The results show that the BCFI Block fusion effect is superior (e.g., the ACC of the choroid layer after using the BCFI Block is 0.863 and the AUC is 0.910). This module is based on a bidirectional LSTM design with a bidirectional double-loop fusion mechanism. First, deformable convolutional features are propagated forward and radiomics features are propagated backward. Then, deformable convolutional features are propagated backward and radiomics features are propagated forward. The outputs of the two loops are superimposed to fully fuse complementary information. Based on this, multi-site OCTA fusion experiments are extended. By default, radiomics features of each site are added and BCFI Block fusion is used. At the same time, a feature interaction module (FIBlock, structure as shown) is introduced in the feature extraction process. Figure 5 As shown in the figure, cross-attention enables information interaction between different parts. Four models were constructed: "choroidal layer + perioptic disc radial capillary layer", "choroidal layer + superficial retina", "perioptic disc radial capillary layer + superficial retina", and "three-part combination". The experimental results are shown in Table 5.

[0063] Table 5 Ablation experiments based on deformable convolutional networks using combined radiomics features of OCTA in multiple locations.

[0064] It is evident that the three-part joint model exhibits the best performance (ACC 0.899, AUC 0.955), effectively integrating information from multiple parts to compensate for the limitations of a single part. The overall model framework is as follows: Figure 4 As shown.

[0065] Step 5: Finally, a hybrid expert module is introduced to decouple the tasks: Based on the three-part joint model, the four-class classification task is split into a "NAION prediction subtask" (distinguishing between normal healthy eyes and healthy eyes contralateral to NAION) and a "NAION / ON acute phase zone molecular task" (distinguishing between eyes in the acute phase of NAION and eyes in the acute phase of ON). The module handles the subtasks separately by setting two classification heads, avoiding classification interference. Experiments compare the performance of the three-part joint unified classification model and the hybrid expert model; the results are shown in Table 6.

[0066] Table 6. Classification results after decoupling from hybrid expert task under unified classification based on combined multi-site OCTA and radiomics features.

[0067] The classification performance of the two models in each sample category is shown in Table 7. The model confusion matrix and ROC curve are shown in Table 7. Figure 6 As shown:

[0068] Table 7 Performance of Unified Classification and Hybrid Expert Task Decoupling on Each Category

[0069] According to the above implementation process, it can be combined with Figure 7 The working principle of this invention is summarized as follows: 1. Sample acquisition in Step 1 and preprocessing in Step 2 are performed on four types of ocular OCTA images to obtain a standardized dataset; 2. Based on the DCN validated in Step 2, an ODCFI Block is constructed. Combined with the radiomics features extracted in Step 3 and the BCFI Block, a multi-site joint framework is built in Step 4 to achieve feature extraction and fusion; 3. In Step 5, a hybrid expert module is introduced to decouple the tasks, complete model training and classification prediction, and output NAION prediction and acute phase differentiation results. Experimental results show that the hybrid expert model has the best performance, with an ACC of 0.912 and an AUC of 0.972 for the four categories. The classification accuracy of each sample category is improved (e.g., the ACC for acute-phase ON eyes increases from 0.842 to 0.895 and the AUC increases from 0.932 to 0.965), effectively solving the problems of difficulty in differentiating NAION from acute-phase ON and low early-phase prediction accuracy of NAION. Although the sample size for the acute phases of NAION and ON was relatively small (72 and 84 cases respectively), resulting in slightly lower performance of this subtask compared to the NAION prediction subtask, it still provides reliable technical support for early intervention and precision treatment of NAION in clinical practice.

[0070] The specific embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of the present invention.

Claims

1. A method for NAION prediction and acute stage differentiation based on deformable convolution combined with multi-site OCTA, characterized in that, include: OCTA images of four types of eye samples were collected: normal healthy eyes, contralateral healthy eyes of NAION patients, eyes in the acute phase of ON, and eyes in the acute phase of NAION. Data from three key areas were extracted: the choroidal layer, the radial capillary layer around the optic disc, and the superficial retina. Using the OCTA deformable convolutional feature extraction module ODCFI Block as the core feature extraction backbone network, after verifying its optimal performance through single-site experiments, a multi-site joint diagnostic framework was constructed. Extract radiomics features and achieve deep fusion of deformable convolutional features and radiomics features through the bidirectional cyclic feature interaction module BCFI Block; Finally, a hybrid expert module is added to the three-part joint model to decouple NAION prediction from the NAION / ON acute phase differentiation task, and the final classification result is output through sub-task collaborative optimization.

2. The NAION prediction and acute stage differentiation method based on deformable convolution combined with multi-site OCTA according to claim 1, characterized in that, The method includes the following steps: Step 1. Collect OCTA image data of the eye and classify the sample types. The samples include four categories: normal healthy eye, healthy eye on the opposite side of NAION patient, eye in the acute phase of ON, and eye in the acute phase of NAION. For each type of sample, OCTA images of three parts are extracted: choroidal layer, radial capillary layer around the optic disc, and superficial retina. Step 2. Preprocess the single-part OCTA images acquired in Step 1 and divide them into training and test sets; Step 3. Use masking techniques to extract radiomics features from the OCTA images in Step 1; Step 4. Construct the OCTA deformable convolutional feature extraction module ODCFI Block based on the DCN verified in Step 2, and build a multi-site joint diagnostic framework: After each OCTA of different sites passes through the ODCFI Block, features are fused through a bidirectional cyclic feature interaction module to achieve unified classification on a single site OCTA; based on single site OCTA, the number of OCTA sites is continuously increased to achieve joint unified classification of two and three sites; during the feature extraction process, the features extracted by each site through the ODCFI Block need to interact with the features of other sites through a bidirectional cyclic feature interaction module to achieve information exchange between OCTAs on different sites, so as to provide more accurate features for subsequent classification; Step 5. Add a hybrid expert module to decouple the four-class classification task into the NAION prediction subtask and the NAION / ON acute phase molecular task. The final classification result is output through the collaborative optimization of the subtasks.

3. The NAION prediction and acute stage differentiation method based on deformable convolution combined with multi-site OCTA according to claim 1, characterized in that, In Step 2, the specific preprocessing includes: Step 2.

1. ROI delineation: Based on the anatomical structure of the eye OCTA, professional radiologists manually delineate the choroidal layer, the radial capillary layer around the optic disc, and the vascular distribution area of ​​the superficial retina to ensure accurate positioning of the target area. Step 2.

2. Contrast Enhancement: The CLAHE algorithm is used to enhance the details of blood vessel texture, with the algorithm block size set to 8×8 and the contrast limit set to 2.0, to improve the distinction between blood vessels and the background; Step 2.

3. Zero-mean normalization: using the formula: ； Eliminate image gray level distribution difference, where is the pixel value at the original image (x, y), is the image gray level mean value, is the gray level standard deviation, is the minimum value to avoid calculation error when, ensure data distribution consistency.

4. The NAION prediction and acute stage differentiation method based on deformable convolution combined with multi-site OCTA according to claim 1, characterized in that, In Step 4, the processing flow of the OCTA deformable convolutional feature extraction module is as follows: The input features are first processed by 3×3 deformable convolution and batch normalization (BN), and then the first residual is added to the original input to retain the initial features. The summed features are then subjected to a 3×3 deformable convolution and a ReLU activation function, and then added to the first residual output for a second residual summation to obtain the final feature output.

5. A system for NAION prediction and acute stage differentiation based on deformable convolution combined with multi-site OCTA, characterized in that, The system includes a module for performing the NAION prediction and acute phase differentiation method based on deformable convolution and multi-site OCTA as described in any one of claims 1 to 4.

6. An electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the NAION prediction and acute phase differentiation method based on deformable convolution and multi-site OCTA as described in any one of claims 1 to 4.

7. A non-transitory computer-readable storage medium having stored thereon a computer program, characterized in that, When the computer program is executed by the processor, it implements the NAION prediction and acute phase differentiation method based on deformable convolution and multi-site OCTA as described in any one of claims 1 to 4.

8. A computer program product comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the NAION prediction and acute phase differentiation method based on deformable convolution and multi-site OCTA as described in any one of claims 1 to 4.

Images