MTM postoperative vision prediction method, system and equipment under small sample

Abstract

The invention discloses an MTM postoperative vision prediction method, system and equipment under a small sample. According to the method, OCT image features are enriched by constructing a radiomics feature fusion image strategy, an efficient local attention mechanism module is introduced according to intertissue filament features and traction features existing in an MTM patient eye OCT image, and the network is helped to better master image key features by using attention features constructed according to the height dimension and the width dimension of the efficient local attention mechanism module. Radiomics features are introduced on the basis of the OfficientNetb0 network to enrich the features of the image, meanwhile, an ELA attention mechanism is introduced to transform a network trunk, and due to the fact that the ELA attention mechanism constructs attention from the height dimension and the width dimension, the ELA attention mechanism has the advantages that the number of parameters is small, and long-strip-shaped features are effectively concerned; therefore, the network is effectively helped to realize MTM postoperative prediction feature acquisition under a small sample. Experiments show that compared with a current mainstream prediction model, the MTM postoperative vision prediction performance under a small sample is effectively improved.

Description

Technical Field

[0001] This invention relates to the field of medical classification, and specifically to a method for predicting postoperative visual acuity after MTM in a small sample size. Background Technology

[0002] Myopic traction macular disease (MTM), a highly destructive retinal complication in patients with high myopia, is fundamentally caused by abnormal traction between the vitreous interface and the retinal layers, leading to anatomical disruption of the macular region. When patients meet the diagnostic criteria for high myopia (axial length ≥26mm or equivalent spherical power ≤-6.00D), the excessively elongated axial length triggers a cascade of changes, including retinal thinning, progressive choroidal atrophy, and vitreous liquefaction. These changes collectively form the anatomical basis for the development of MTM. Under continuous mechanical traction, the macular region can experience multi-layered pathological changes: from abnormal adhesion between the internal limiting membrane and the posterior vitreous membrane to tractional retinal schisis, which may eventually evolve into blinding complications such as macular holes, choroidal neovascularization, and even retinal detachment. These progressive damages irreversibly destroy the patient's central visual function.

[0003] The main way for physicians to observe the structure of the macular region is optical coherence tomography (OCT). Features such as retinal layering, neuroepithelial separation and choroidal thickness changes in OCT images provide important basis for the early diagnosis and treatment of MTM. At present, the main treatment for MTM is surgical intervention, including vitrectomy, internal limiting membrane peeling

[18] . These surgeries can alleviate macular traction symptoms to a certain extent, improve the anatomical structure of the macular region, and promote the recovery of patients' vision. However, since MTM lesions involve multiple layers of retinal tissue and complex interactions between the vitreous and retina, the postoperative visual recovery of patients is easily affected by a variety of factors, such as the preoperative macular structure, the degree of retinal lesions, the degree of vitreous liquefaction, choroidal thickness, surgical method and individual differences of patients. Even patients with similar imaging manifestations will still have significant differences in their postoperative visual recovery. Therefore, accurately predicting the postoperative visual recovery of MTM patients is of great importance to improving the success rate of surgery and the prognosis assessment of patients. However, it is difficult to accurately predict the postoperative visual recovery of MTM patients based solely on the doctor's experience or a single image indicator. In this forecasting task, it is necessary to establish a multi-factor comprehensive forecasting model. Summary of the Invention

[0004] To address the shortcomings of the application of deep learning in postoperative visual acuity prediction after microscopic transcranial Doppler (MTM) surgery and the slow accumulation of datasets for this task, this invention proposes a method, system, and device for predicting postoperative visual acuity after MTM surgery using a small sample size.

[0005] This invention first proposes a strategy to enrich OCT image features by constructing images fused with radiomics features. Secondly, it introduces an Efficient Local Attention (ELA) module to address the interstitial filamentous features and traction characteristics present in OCT images of MTM patients' eyes. This module utilizes ELA's ability to construct attention along both height and width dimensions to help the network better grasp key image features. To solve the challenge of accurate postoperative visual acuity prediction in MTM with small sample sizes, the technical solution adopted in this invention includes the following steps: S1. Collect eye data: Optical coherence tomography (OCT) images of the eyes of different patients were acquired. Based on postoperative visual acuity, the images were divided into two categories: visual acuity below a set threshold indicated poor visual recovery, while visual acuity above the set threshold indicated good visual recovery. The images were then categorized into two classes using a set of thresholds, and the resulting datasets were further divided into a test set and a training set.

[0006] The threshold is set based on the logarithm of the minimum discernible angle (LogMAR). A LogMAR greater than or equal to 0.3 is considered to indicate good visual recovery, while a LogMAR less than or equal to 0.3 is considered to indicate poor recovery.

[0007] S2. Enrich image features by constructing a radiomics feature fusion image strategy, as detailed below: S2-1. Manually annotate the key regions of each OCT image in the dataset to obtain a unique annotated region for each image.

[0008] S2-2. Extract radiomics features from the labeled regions.

[0009] S2-3. The N radiomics features extracted from each image are filtered using mutual information methods, retaining the top N1 radiomics features with the strongest relationship to the target variable. Four physician-collected patient information items are then introduced, resulting in N1+4 features. These four physician-collected patient information items include: patient gender, age, axial length, and preoperative visual acuity.

[0010] S2-4. The N1+4 features are further filtered using the Recursive Feature Elimination (RFE) method to obtain the N2 high-order features that contribute significantly to the prediction task.

[0011] S2-5. Perform logarithmic normalization on the N2 high-order features, then divide them into two groups on average. Then, perform dimensionality transformation on the N2 / 2 features in each group to obtain two feature matrices of (N2 / 2)^(1 / 2)*(N2 / 2)^(1 / 2). Then, perform unpooling on the feature matrices to obtain a feature matrix of scale 224*224. Perform dot product operation on the feature matrix with the second and third channels of the patient's eye OCT image to achieve the fusion requirement.

[0012] S3. Data processing of OCT images of MTM patients' eyes based on Efficient Local Attention-Efficient Network (ELA-EfficientNet); S3-1. In the front end of the ELA-EfficientNet network, a convolutional layer is used to expand the input image (i.e., the eye OCT image) from 3 channels to 32 channels, and a convolution with a stride of 2 is used to reduce the feature map size. After the convolutional layer, the BatchNorm and Swish activation functions are passed in sequence. The backbone of the S3-2.ELA-EfficientNet network consists of multiple layers of efficient local attention-moving inverse residual convolutional layers (ELA-MBConv). The ELA-MBConv module gradually expands and compresses the number of channels, while enhancing the expressive power of features through the attention mechanism. The ELA-MBConv module combines depthwise separable convolution, attention mechanism and residual connection.

[0013] S3-3. At the end of the network, convolution is first used to expand the number of channels to 1280, then global average pooling is applied to compress the feature map to 1*1, and finally the compressed feature map is flattened and mapped through a fully connected layer to obtain the corresponding class probability. The class with the highest probability is selected as the final prediction result.

[0014] S3. Construct a loss function using the training set, and train the ELA-EfficientNet network using the binary cross-entropy loss function. After N training rounds, obtain the best-performing ELA-EfficientNet network.

[0015] S4. Input the data from the test set described in step S1 into the trained ELA-EfficientNet network to perform the prediction task.

[0016] A small-sample MTM postoperative prediction system includes a data acquisition module, a feature processing and image fusion module, an ELA-EfficientNet network module, a training module, and a testing module; Data acquisition module: Collects eye data and divides it into training and test sets; Feature processing and image fusion module: It filters key macular radiomic features and some physician-extracted information from patients' eye OCT images, retains 32 types of high-contribution features, and performs feature fusion operations on small sample eye OCT images; The ELA-EfficientNet network module: At the front end of the ELA-EfficientNet network, convolutional layers are used to expand the input image from 3 channels to 32 channels, and convolution is used to reduce the feature map size. After the convolutional layers, the BatchNorm and Swish activation functions are applied sequentially. The backbone of the ELA-EfficientNet network consists of multiple layers of efficient local attention-moving inverse residual convolutional layers (ELA-MBConv). The ELA-MBConv module is used to gradually expand and compress the number of channels, while the attention mechanism enhances the expressive power of the features. At the tail of the ELA-EfficientNet network, convolution is first used to expand the number of channels to 1280, then global average pooling is applied to compress the feature map to 1*1, and finally the compressed feature map is flattened and mapped through a fully connected layer to obtain the corresponding class probability. The class with the highest probability is selected as the final prediction result.

[0017] The ELA-EfficientNet network module is based on the EfficientNet_b0 model and performs novel feature fusion on eye OCT images. The EfficientNet_b0 model utilizes depthwise separable convolution and ELA attention mechanism to accurately grasp the features of eye OCT images with a limited number of parameters and to process data from small sample eye OCT images. Training module: Construct a loss function using the training set, and train the ELA-EfficientNet network using the loss function. After N training rounds, obtain the best-performing ELA-EfficientNet network. Test module: Input the test set into the trained ELA-EfficientNet network to perform postoperative visual acuity prediction after MTM.

[0018] A device for predicting postoperative visual acuity after MTM in small sample sizes, the key feature of which is that the device is equipped with a prediction system for postoperative MTM in small sample sizes.

[0019] The beneficial effects of the present invention are as follows: This invention proposes a postoperative prediction algorithm for MTM (Malignant Transmission of the Eye) based on radiomics feature fusion images and ELA-EfficientNet. Existing models for MTM postoperative visual acuity prediction are limited by overfitting issues with large parameter sets on small datasets and struggle to accurately capture key features related to postoperative recovery in MTM patients. Radiomics features can characterize patient lesion information to some extent and can therefore be combined with images for deep learning methods. EfficientNet is known for its parameter efficiency and exhibits excellent generalization in small sample tasks, but its Squeeze-and-Excitation (SE) attention mechanism is incompatible with the feature layout of OCT images of MTM patients' eyes. This invention introduces radiomics features to enrich the image features on the EfficientNet_b0 network and simultaneously introduces the ELA attention mechanism to modify the network backbone. Since the ELA attention mechanism constructs attention from both height and width dimensions, it has the advantages of fewer parameters and effective focus on long, narrow features, which effectively helps the network acquire predictive features for MTM postoperative recovery even with small sample sizes.

[0020] This invention proposes a postoperative prediction algorithm for MTM based on radiomics feature fusion images and ELA-EfficientNet, which can effectively solve the problem of postoperative visual acuity prediction for MTM with small sample sizes. Attached Figure Description

[0021] Figure 1 The structure of the ELA-EfficientNet algorithm; Figure 2 This is a structural diagram of the feature processing and image fusion module; Figure 3 Here is a structural diagram of the ELA-MBConv module; Figure 4 Here is a diagram of the ELA module structure; Figure 5 Annotated OCT images of the eyes of MTM patients; Figure 6 Comparison of OCT images of the eyes of MTM patients; Detailed Implementation

[0022] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0023] This embodiment uses a postoperative prediction method for MTM based on radiomics feature fusion images and ELA-EfficientNet, specifically including the following steps: S1. Constructing the dataset S1-1: The dataset W2HOSPITAL-177 required for this invention is constructed using OCT images of the eyes of different MTM patients. The dataset contains 177 OCT images of the eyes, of which 141 are used as the training set and 36 are used as the test set.

[0024] S1-2: Manually annotate the above dataset, focusing on the patient's main lesion area, specifically the macular region. Examples of W2HOSPITAL-177 ocular OCT images and annotated images are shown below. Figure 5 As shown.

[0025] S-3: Based on postoperative visual acuity testing, patients with LogMAR less than 0.3 are classified into Class 0, indicating poor postoperative recovery, while those with LogMAR greater than or equal to 0.3 are classified into Class 1, indicating good postoperative recovery. Images for Class 0 and Class 1 are shown below. Figure 6 As shown.

[0026] S2. The eye OCT image is fused based on the feature processing and image fusion module.

[0027] S3. Feature extraction of eye OCT images based on ELA-EfficientNet network to achieve prediction.

[0028] The basic framework of the method proposed in this invention is as follows: Figure 1 As shown.

[0029] Figure 1 As the basic framework of this invention, the input image first undergoes four steps in the feature processing and image fusion module: feature extraction, feature filtering, feature preprocessing, and feature-image fusion. After passing through this module, the original image obtains a fused image containing some high-value radiomics features, thereby improving image specificity and helping to improve the model's multi-scale feature expression capability. The fused image is then fed into the ELA-EfficientNet network. It first goes through a 3*3 convolutional layer to expand from 3 channels to 32 channels, and then uses a convolution with a stride of 2 to reduce the size. After the convolutional layer, it goes through the BatchNorm and Swish activation functions. The initial processed feature matrix is ​​used to extract key features through a seven-layer ELA-MBConv module. During the process of extracting features layer by layer, the size of the feature matrix is ​​compressed and the number of channels is expanded so that specific category probabilities can be output at the end of the network.

[0030] At the end of the network, a 1*1 convolution is first used to expand the number of channels of the feature matrix to 1280. Then, global average pooling is applied to compress the feature map to 1*1. Finally, the feature matrix is ​​flattened and mapped to the number of classes through a fully connected layer to output the final prediction result.

[0031] Feature processing and image fusion module like Figure 2 As shown, it mainly includes four steps: feature extraction, feature selection, feature preprocessing, and feature fusion.

[0032] In the feature extraction part, the semantic segmentation model is used to segment the lesions of the patient, and the radiomics features of the segmented regions are extracted using the Python toolkit Pyradiomics. This yields seven major categories of features, including the original two-dimensional shape features, the original first-order features, the original gray-level co-occurrence matrix, the original gray-level long-running matrix, the original gray-level region matrix, the original gray-level dependency matrix, and the original non-uniformity matrix. Since each major category of features will have multiple sub-features, such as the original two-dimensional shape features including a series of features such as extensibility, principal axis length, and maximum diameter, a total of 496 feature categories are obtained for each image.

[0033] In the feature selection process, directly applying the Recursive Feature Elimination (RFE) method to obtain the optimal features from the 496 feature classes would require a large amount of computation. Therefore, the mutual information method was first used for selection. Mutual information is used to measure the correlation between different features and labels, and features with a strong relationship with the target variable are selected. Finally, 102 features were initially selected and then four physician-extracted features (patient gender, age, axial length, and preoperative visual acuity) were introduced to obtain 106 feature classes. At this point, RFE was performed on the 106 feature classes based on the SVM algorithm, and 32 features that contributed the most to the model's predictive ability were retained, including patient axial length, preoperative visual acuity, original first-order feature (mean absolute deviation), and original gray-level co-occurrence matrix (autocorrelation).

[0034] In the feature preprocessing section, since the 32 feature values ​​after feature screening have a large range, the general normalization method is prone to small data values ​​being too close to 0, thus losing their feature value. Therefore, the logarithmic transformation method is used to normalize the 32 feature classes. In order to meet the requirements of feature and image fusion operation, the features are matrix transformed to obtain a 4*4*2 three-dimensional feature matrix, and the matrix is ​​unpooled to obtain a 224*224*2 three-dimensional feature matrix with the same scale as the original input image.

[0035] In the feature fusion section, to preserve some original features from the three channels of the original input image, only the second and third channels of the original image are used to perform a matrix multiplication operation with the preprocessed three-dimensional feature matrix to obtain the final fused image. ELA-MBConv module like Figure 3As shown, the network comprises an expansion layer, depthwise separable convolutions, an ELA module, a projection layer, and residual connections. The expansion layer uses 1x1 convolutions to increase the number of channels from c to t*c, where t is the expansion factor (t=6 in this invention). This layer increases the representational power of features. The depthwise separable convolutions extract spatial features by independently performing 3x3 or 6x6 depthwise convolutions on each input channel, thus reducing computational cost. The ELA module calculates attention weights in both the height and width directions using depthwise separable convolutions. This enhances the model's focus on local features and allows the network to better integrate information from elongated or localized lesion areas. This improvement provides finer-grained feature enhancement and improves the model's discriminative ability. The projection layer uses 1x1 convolutions to reduce the number of channels back to the original number c. Finally, if the stride is 1 and the number of input channels equals the number of output channels, residual connections are used. This means the input is added to the output after convolution and attention processing, improving network stability and facilitating gradient propagation.

[0036] ELA module like Figure 4 As shown, the input x has dimensions (b, c, h, w), where b is the training batch size, c is the number of channels, and h and w are the height and width of the feature map, respectively. First, the input feature map is average-pooled along the height dimension, generating a tensor of size (b, c, h, 1). Then, through convolution and population normalization, a weight between [0, 1] is generated using the sigmoid activation function to weight the feature map along the height dimension. Similarly, average-pooling is performed on the features along the width dimension, and the same operation generates attention weights along the width dimension. Finally, the input feature map is element-wise multiplied with the attention maps along both the height and width dimensions to generate the weighted feature map. This method, which calculates attention separately for both height and width, enables the network to capture important regional information in both the horizontal and vertical directions. This effectively increases the network's attention to local information, especially elongated regions, while maintaining the integrity of global features. This is highly suitable for the stretching and traction characteristics of the macular region in postoperative prediction tasks of MTM (Malignant Transformer), and also aligns with the feature layout of lesions such as internal and external splits and internal and external holes that affect the postoperative recovery of MTM patients. Furthermore, because it uses mean pooling and lightweight convolutions, rather than computationally intensive Transformers or self-attention modules, this method avoids overfitting with limited data, making it more suitable for small-sample medical image analysis tasks. S3. Train the entire model in 50 rounds, and finally retain the model with the best performance on the validation set.

[0037] S4. Input the test set images into the trained model to extract and make predictions.

[0038] The experimental server used in this invention employed an NVIDIA TITAN RTX 4090 GPU, and the deep learning framework used was PyTorch. The input images were 224*224 resolution 3-channel ocular ultrasound images. The training batch size was 8, and the training run consisted of 50 epochs. The Adam optimizer was used to optimize the loss function, with a learning rate of 3*10⁻⁶. -4 The weight decay factor is 10. -5 .

[0039] Example Due to the relatively limited dataset size in this chapter, a transfer learning strategy was employed to reduce the risk of overfitting and improve the model's generalization ability. To ensure the fairness and comparability of the experimental results, several mainstream convolutional neural network models were also constructed as comparison methods while training the improved model, including ResNet18, ResNet34, MobileNetV3, EfficientNetV2-S, and EfficientNet_b0. All models were initialized using parameters pre-trained on the ImageNet dataset to achieve more stable convergence performance and mitigate the risk of overfitting due to the small sample size. For the current binary classification task, the output layer of each model was uniformly replaced with a fully connected layer structure adapted for binary classification, and training was conducted on the W2HospitalMTM-177 dataset. By uniformly adopting the transfer learning strategy, the training fairness of different network structures under the same initial conditions was effectively guaranteed, making the model performance comparison more objective and reliable.

[0040] The following is an experimental data analysis of the algorithm proposed in this invention based on the partitioned test set data. The prediction performance of this invention with ResNet18, ResNet34, MobileNetV3, EfficientNetV2-S and EfficientNet_b0 algorithms on the W2HOSPITAL-5000 dataset is compared for original images and fused images.

[0041] To verify the superiority of the model of this invention, four classification performance indicators, namely AUC, sensitivity, specificity, and F1 score, are used to evaluate the classification performance of the algorithm model in this chapter. AUC is the area under the ROC operating characteristic curve, which measures the classifier's ability to distinguish between positive and negative samples. The AUC value ranges from [0, 1], with a higher value indicating better classification performance.

[0042] In medical classification tasks, an AUC of 0.5 indicates that the classifier performs the same as random sampling, while an AUC of 1.0 indicates that the classifier perfectly distinguishes all positive and negative samples. A higher AUC means that the classifier can effectively distinguish between healthy and sick individuals at different thresholds, and is not affected by a single threshold.

[0043] Sensitivity represents the proportion of all real unhealthy cases that are correctly identified as unhealthy by the model, i.e., the ability to detect real cases. The calculation formula is as follows:

[0044] In the formula, TP refers to cases that are actually unhealthy but are correctly classified as unhealthy; FN refers to cases that are actually unhealthy but are incorrectly classified as healthy.

[0045] In medical classification tasks, sensitivity means that the model will not miss diagnoses, that is, most real patients are correctly identified. High sensitivity is very important in many disease screening tasks because missed diagnoses can lead to serious consequences. However, a model with high sensitivity may produce more false diagnoses, so it is necessary to weigh it in conjunction with specificity.

[0046] Specificity measures a model’s ability to correctly identify healthy cases, i.e., the model’s ability to exclude normal samples. Its calculation formula is as follows:

[0047] TN (True Negative) represents the number of correctly predicted negative cases, while FP represents the number of cases where a negative case was mistakenly predicted as a positive case. High specificity helps reduce the misdiagnosis rate and avoids unnecessary anxiety and treatment burden for patients.

[0048] The F1 score is the harmonic mean of precision and recall, measuring the balance between the model's ability to correctly predict the proportion of positive examples and recall positive examples. The formula is as follows:

[0049] It is particularly suitable for medical tasks with unbalanced categories, and can balance the impact of missed diagnoses and misdiagnoses.

[0050] The classification performance of ResNet18, ResNet34, MobileNetV3-S, EfficientNet_b0, EfficientNetV2-S, and ELA-EfficientNet on the original image is as follows: Table 1. Postoperative prediction performance of different algorithms on the original images using MTM.

[0051] As shown in Table 1, the ELA-EfficientNet algorithm of this invention, by introducing the ELA module, helps to improve the network's attention to the height and width dimensions of the input OCT image, which is well-suited to the distribution of macular lesion features in OCT images. Ultimately, the algorithm model of this invention achieved the highest values ​​in all four parameters—AUC, sensitivity, specificity, and F1 score—achieving a significant improvement in classification performance on the MTM postoperative prediction task.

[0052] Combining the strategy of constructing radiomics feature fusion images, the performance of ResNet18, ResNet34, MobileNetV3-S, EfficientNet_b0, EfficientNetV2-S, and ELA-EfficientNet on the fused images is as follows: Table 2. Postoperative prediction performance of different algorithms on fused images using MTM.

[0053] As shown in Table 2, the performance trends of different models remain consistent under fused images, but their performance is generally improved compared to the original images. This demonstrates the effectiveness of the proposed strategy for constructing fused radiomics feature images. In comparison, the ELA-EfficientNet algorithm of this invention still achieves the best performance in terms of AUC, sensitivity, specificity, and F1 score.

[0054] This invention addresses the challenge of small datasets in postoperative MTM prediction due to difficulties in accumulating such datasets. It proposes a postoperative MTM prediction algorithm based on radiomics feature fusion images and ELA-EfficientNet. This efficient model accurately captures key lesion features in OCT images of MTM patients' eyes, significantly improving postoperative visual acuity prediction performance with small sample sizes compared to current mainstream prediction models. This demonstrates the effectiveness of the proposed method.

[0055] The 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, and these changes should also be considered within the scope of protection of the present invention.

Claims

1. A method for predicting postoperative visual acuity after MTM surgery in a small sample size, characterized in that, This method enriches OCT image features by constructing a strategy for fusing radiomics features into images. It introduces an efficient local attention mechanism module targeting the interstitial filamentous features and traction characteristics present in the ocular OCT images of MTM patients. Utilizing its ability to construct attention along both height and width dimensions, this module helps the network better grasp key image features. Specifically, it includes the following steps: S1. Collect eye data: S2. Enrich image features by constructing a radiomics feature fusion image strategy, as detailed below: S3. Data processing of OCT images of MTM patients' eyes based on efficient local attention-efficient network; S4. Construct a loss function using the training set, and train the efficient local attention-efficient network using the binary cross-entropy loss function. After N training rounds, obtain the efficient local attention-efficient network with the best training performance. S5. Input the data from the test set described in step S1 into the trained high-efficiency local attention-high-efficiency network to perform the prediction task.

2. The method for predicting postoperative visual acuity of MTM under small sample conditions according to claim 1, characterized in that, Step 1 is implemented as follows: Optical coherence tomography (OCT) images of the eyes of different patients were acquired. Based on the patients' postoperative visual acuity test results, the images were divided into two categories: when the visual acuity test results were below a set threshold, it indicated poor visual recovery; when the visual acuity test results were above the set threshold, it indicated good visual recovery. Images are classified into two categories by setting a threshold, and the classified datasets are then divided into a test set and a training set.

3. A method for predicting postoperative visual acuity of MTM in a small sample size according to claim 1 or 2, characterized in that, The threshold set in step 1 is based on the logarithm of the minimum discernible angle. A logarithm greater than or equal to 0.3 is considered to indicate good visual recovery, while a logarithm less than 0.3 is considered to indicate poor recovery.

4. The method for predicting postoperative visual acuity of MTM under small sample conditions according to claim 2, characterized in that, Step 2 is implemented as follows: S2-1. Manually annotate the key regions of each OCT image in the dataset to obtain a unique annotated region for each image; S2-2. Extract radiomics features from the labeled regions; S2-3. The N radiomics features extracted from each image are filtered using the mutual information method. The top N1 radiomics features with the strongest relationship to the target variable are retained, and four physician-collected patient information items are introduced to obtain N1+4 features. S2-4. The N1+4 features are further filtered using a recursive feature elimination method to obtain the N2 high-order features that contribute significantly to the prediction task. S2-5. Perform logarithmic normalization on the N2 high-order features, then divide them into two groups on average. Then, perform dimensionality transformation on the N2 / 2 features in each group to obtain two feature matrices of (N2 / 2)^(1 / 2)*(N2 / 2)^(1 / 2). Then, perform unpooling on the feature matrices to obtain a feature matrix of scale 224*224. Perform dot product operation on the feature matrix with the second and third channels of the patient's eye OCT image to achieve the fusion requirement.

5. The method for predicting postoperative visual acuity of MTM in a small sample size according to claim 4, characterized in that, The four patient information items extracted by the physician include: patient gender, age, axial length, and preoperative visual acuity.

6. The method for predicting postoperative visual acuity of MTM in a small sample size according to claim 4, characterized in that, Step 3 is implemented as follows: S3-1. In the efficient local attention-efficient network front end, convolutional layers are used to expand the input image from 3 channels to 32 channels, and convolution is used to reduce the feature map size. After the convolutional layer, the BatchNorm and Swish activation functions are passed in sequence. S3-2. Efficient Local Attention - The backbone of the efficient network is composed of multiple layers of efficient local attention-moving inverse residual convolutional layers. The efficient local attention-moving inverse residual convolutional layers are used to gradually expand and compress the number of channels, while the attention mechanism is used to enhance the expressive power of features. S3-3. At the end of the multi-layer efficient local attention-moving inverse residual convolutional layer, firstly, convolution is used to expand the number of channels to 1280, then global average pooling is applied to compress the feature map to 1*1, and finally, the compressed feature map is flattened and mapped through a fully connected layer to obtain the corresponding class probability. The class with the highest probability is selected as the final prediction result.

7. The method for predicting postoperative visual acuity of MTM under small sample conditions according to claim 6, characterized in that, Multi-layer efficient local attention-moving inverse residual convolutional layers combine depthwise separable convolution, attention mechanisms, and residual connections.

8. A small-sample MTM postoperative visual acuity prediction system, characterized in that, It includes a data acquisition module, a feature processing and image fusion module, an efficient local attention-efficient network module, a training module, and a testing module; Data acquisition module: Collects eye data and divides it into training and test sets; Feature processing and image fusion module: It filters key macular radiomic features and some physician-extracted information from patients' eye OCT images, retains 32 types of high-contribution features, and performs feature fusion operations on small sample eye OCT images; The efficient local attention-efficient network module uses convolutional layers at the front end to expand the input image from 3 channels to 32 channels, and uses convolution to reduce the feature map size. After the convolutional layers, the BatchNorm and Swish activation functions are applied sequentially. The backbone of the efficient local attention-efficient network consists of multiple layers of efficient local attention-moving inverse residual convolutional layers stacked together. The efficient local attention-moving inverse residual convolutional layer module gradually expands and compresses the number of channels, while enhancing the expressive power of features through the attention mechanism. At the tail of the efficient local attention-efficient network, convolution is first used to expand the number of channels to 1280, then global average pooling is applied to compress the feature map to 1*1, and finally the compressed feature map is flattened and mapped through a fully connected layer to obtain the corresponding class probabilities. The class with the highest probability is selected as the final prediction result. Training module: Construct a loss function using the training set, and train the efficient local attention-efficiency network using the loss function. After N training rounds, the best-performing efficient local attention-efficiency network is obtained. Test module: Input the test set into the trained high-efficiency local attention-high-efficiency network to predict postoperative visual acuity after MTM.

9. A device for predicting postoperative visual acuity after MTM in a small sample size, characterized in that, The device is equipped with the small-sample MTM postoperative visual acuity prediction as described in claim 8.

Images