Medical image super-resolution reconstruction method based on incremental learning

Abstract

The invention discloses a medical image super-resolution reconstruction method based on incremental learning, and aims to solve the problem that a super-resolution reconstruction network is easy to have disastrous forgetting during incremental learning of a multi-modal medical image through a cooperative strategy of adaptive memory buffer management, relational knowledge distillation and dynamic gradient equilibrium. And efficient learning of the new mode and stable retention of the reconstruction capability of the old mode are realized. According to the method, adaptive gradient weighted difficulty scoring is combined with farthest point sampling to efficiently screen playback samples, a characteristic space geometrical relationship is maintained by utilizing relational knowledge distillation, and conflicts are monitored in real time and weights are automatically adjusted through a dynamic gradient balance mechanism. According to the method, the disastrous forgetting degree is remarkably reduced, a single model can learn multiple modes in sequence and continuously keep the super-resolution reconstruction capacity for all the learned modes, and the method can be widely applied to multi-mode medical image super-resolution reconstruction tasks such as MRI and chest X-rays and has high clinical application value.

Description

Technical Field

[0001] This invention relates to the fields of medical image super-resolution reconstruction and incremental learning, and particularly to a method for medical image super-resolution reconstruction based on incremental learning. Background Technology

[0002] Image super-resolution reconstruction is one of the fundamental tasks in computer vision, with significant value in medical diagnosis. It enables the reconstruction of high-resolution images from low-resolution medical images, improving diagnostic accuracy. However, most existing models are designed for specific modalities (such as MRI or X-ray). In clinical practice, medical image data typically occurs as a continuous multimodal data stream. Training and maintaining independent models for each new modality is not only inefficient and costly but also fails to leverage prior knowledge. Therefore, there is an urgent need for a super-resolution reconstruction method that supports incremental learning, allowing a single model to learn different tasks sequentially to adapt to continuous data streams.

[0003] The core challenge of incremental learning lies in overcoming catastrophic forgetting, where a model rapidly loses its retention of knowledge from older modalities when learning new modalities, leading to a significant decline in performance on older tasks. Therefore, striking a balance between efficiently learning new modalities and stably retaining old ones is crucial for achieving clinical applications.

[0004] Existing solutions for incremental learning in medical image super-resolution reconstruction face several challenges: regularization-based methods may over-constrain the model, making it difficult to adapt to new data distributions and impairing its adaptability. Playback-based methods are highly dependent on sample selection strategies; random or simple sampling can lead to inefficient memory buffers. Furthermore, when combined with knowledge distillation techniques, existing methods typically employ hard constraints to match feature outputs, potentially limiting the flexibility required for learning new tasks. Therefore, a collaborative design approach that balances sample selection, knowledge preservation, and the integration of new and old tasks to achieve efficient and flexible incremental learning for medical image super-resolution reconstruction remains an open challenge. Summary of the Invention

[0005] The goal of this invention is to provide an incremental medical image super-resolution reconstruction method that addresses the problems of low buffer efficiency, knowledge distillation affecting plasticity, and insufficient balance between new and old tasks in existing incremental learning techniques.

[0006] This invention achieves its goals through the following technical solutions: First, to address the buffer efficiency problem, an adaptive memory buffer management strategy is designed. This strategy evaluates sample value through adaptive gradient-weighted difficulty scoring and combines feature space farthest point sampling for diversity selection, enabling efficient screening and dynamic replacement of high-difficulty, high-impact, and high-diversity samples. Second, to address the issue of knowledge distillation affecting plasticity, a relational knowledge distillation strategy is designed. This strategy constrains the geometric relationship of the feature space between samples within a batch, rather than the absolute value of the features, thereby improving the flexibility of learning new features while retaining old knowledge. Third, to address the issue of dynamically changing conflict levels between new and old tasks, a dynamic gradient balancing training strategy is designed. This strategy monitors the cosine similarity of the gradients of new and old tasks in real time, assesses the degree of conflict, and automatically adjusts the knowledge retention weights, thereby intelligently balancing the stability and plasticity of the model.

[0007] The specific technical solution for achieving the objective of this invention is as follows:

[0008] A medical image super-resolution reconstruction method based on incremental learning enhances the incremental learning capability of the super-resolution reconstruction network in multimodal data streams through adaptive memory buffer management, relational knowledge distillation, and dynamic gradient balancing training strategies. Specifically, it includes the following steps:

[0009] Step S1: Dataset Construction

[0010] A multimodal medical image dataset was constructed, treating each modality as an independent learning task. The task sequence arrived in order, and data from old tasks could not be accessed when learning a new task. The dataset was divided into a training set and a test set. Each task in the training set contained 4000 samples, and each task in the test set contained 1000 samples.

[0011] Step S2: Dataset Preprocessing

[0012] The multimodal medical image dataset is preprocessed by generating corresponding low-resolution samples for each high-resolution sample using bicubic interpolation downsampling, forming image pairs for supervised training of the super-resolution reconstruction network. In addition, a variety of data augmentation methods, including random cropping, random flipping, random scaling, and random adjustment of image brightness, saturation, and chroma, are applied to the image pairs to ensure that the anatomical structure information is not distorted.

[0013] Step S3: Construction of the Super-Resolution Reconstruction Network

[0014] The super-resolution reconstruction network adopts a Transformer-based encoder-bottleneck-decoder architecture. Low-resolution images are processed by convolutional layers to extract shallow features. Simultaneously, a routing instruction module generates routing instruction vectors through a convolutional encoder, pooling layers, and a softmax layer to guide different modalities in selecting processing paths. The encoder contains multiple layers, each with a spatial routing module preceding the Transformer block. The spatial routing module employs a hybrid expert mechanism to select expert networks to process features based on the routing instruction vectors. After the Transformer block extracts deep features, it reduces spatial resolution through downsampling. The bottleneck part sets up a channel routing module before the Transformer block, which, based on the routing... The instruction vector generates a channel mask to achieve channel-level feature selection. After the Transformer block extracts features, it is upsampled and then fed into the decoder. The decoder contains multiple layers. The feature input layers of different layers of the encoder are spatially aligned and weighted by the attention module, and then passed to the corresponding layer of the decoder after channel concatenation and convolution adjustment. Each layer sets a channel routing module before the Transformer block to perform channel-level selection. After the Transformer block extracts features, it restores the spatial resolution by upsampling and fuses the features from the corresponding layer of the encoder after being processed by the layer attention module. The final output of the decoder is processed by a convolutional layer to generate a residual image, which is added to the upsampled low-resolution input to generate a high-resolution image.

[0015] Step S4: Collaborative Strategies for Incremental Learning

[0016] The incremental learning strategies are designed, including adaptive memory buffer management, relational knowledge distillation, and dynamic gradient balancing training strategies, specifically as follows:

[0017] 4-1: Adaptive Memory Buffer Management Strategy

[0018] A fixed-size memory buffer is maintained, and the value of samples is evaluated by an adaptive gradient-weighted difficulty score, which integrates the reconstruction difficulty and learning influence of the samples. During training, samples are preferentially sampled according to the scores. After the task is completed, the buffer is updated by selecting samples with both high value and diversity by combining the scores with the feature space farthest point sampling algorithm, and the deep feature representation of the samples is stored for knowledge distillation.

[0019] 4-2: Relational Knowledge Distillation Strategy

[0020] When training a new task, the replay samples are used to constrain the feature geometric relationship structure of the current model on the old samples; the pairwise distance matrix of the new feature representation batch extracted by the current model and the pairwise distance matrix of the old feature representation batch are calculated, and the knowledge distillation loss is the difference between the new and old distance matrices, thereby constraining the relationship structure of the feature space and preserving the old knowledge.

[0021] 4-3 Dynamic Gradient Balancing Training Strategy

[0022] During training, the cosine similarity between the gradients of the new task and the gradients of the old task is calculated to assess the degree of gradient conflict. A dynamic scaling factor is calculated based on the degree of conflict. The weights of the replay sample loss and relation distillation loss are automatically adjusted using the dynamic scaling factor, and the protection of old knowledge is automatically increased when gradient conflicts occur.

[0023] Step S5: Incremental learning training and use of the model

[0024] Using the data from step S1, train the network from step S3 in task order, and perform incremental learning using the strategy from step S4. In each training iteration, sample new task samples from the current task dataset, sample replay samples from the memory buffer, and update the model using the total loss function, which includes new task super-resolution reconstruction loss, replay sample super-resolution reconstruction loss, and relation distillation loss. Update the memory buffer after task training is completed. After all tasks are trained, use the final model to perform super-resolution reconstruction on low-resolution images in the test set.

[0025] Furthermore, the adaptive gradient-weighted difficulty scoring described in step 4-1 pass Calculations; among which, reconstruction difficulty The influence is measured by the Structural Similarity Index (SSIM) between the current model reconstruction result and the real high-resolution image. The gradient norm of the super-resolution reconstruction loss is used as a metric. For balance coefficient, This represents the min-max normalization operation; the feature space farthest point sampling algorithm iteratively selects samples from the candidate pool with high scores, and each time selects the sample that is farthest from the selected sample set in the deep feature space in the Euclidean distance, so as to ensure that the samples stored in the buffer have both high value and diversity.

[0026] Furthermore, the knowledge distillation loss described in step 4-2 Defined as the Frobenius norm between the normalized distance matrices representing batches of new and old features; the normalization process involves dividing each element in the matrix by the maximum value of all off-diagonal elements of the matrix; by minimizing the loss... The relative distance between sample pairs within a batch must be kept consistent.

[0027] Furthermore, the dynamic scaling factor described in step 4-3 The gradient is adaptively determined by calculating the cosine similarity between the gradient of the new task and the gradient of the old task; when gradient conflicts occur, the... Automatically enlarge; using the described Real-time adjustment of replay sample loss weights And the relationship between distillation loss weight ,in and These are the basic weighting coefficients.

[0028] Compared with the prior art, the present invention has the following beneficial technical effects and significant technical progress:

[0029] 1) Efficient Adaptive Buffer Management: The adaptive gradient-weighted difficulty scoring proposed in this invention combines the reconstruction difficulty and learning influence of samples, while using sampling at the farthest point in the feature space to ensure sample diversity. By replacing the worst-scoring old sample, the buffer ensures that it always retains the most valuable samples for the current model, significantly improving replay efficiency.

[0030] 2) Balancing stability and plasticity in knowledge preservation: This invention employs relational knowledge distillation, constraining only the geometric relationships between features, rather than the absolute values ​​of features. This flexible knowledge preservation method, compared to traditional feature distillation, prevents forgetting and provides greater adaptability for the model to learn new modalities, especially those with significant differences.

[0031] 3) Intelligent Adaptive Training Strategy: The dynamic gradient balancing mechanism designed in this invention can detect gradient conflicts between old and new tasks in real time. When the task differences intensify, such as transitioning from MRI to chest X-ray modality, the protective weights of old knowledge are automatically increased. When the tasks are similar, the weights are automatically reduced to accelerate the learning of new knowledge. This mechanism enables the model to intelligently adapt to task transitions without the need for manual parameter adjustments.

[0032] 4) Medical Image Application Value and Practical Prospects: This invention is specifically designed for multimodal medical image super-resolution reconstruction and can be directly applied to tasks involving continuous multimodal data streams such as MRI and chest X-rays. This method ensures that the model retains the reconstruction capabilities of the old modalities when learning new modalities, demonstrating high potential for clinical application. Attached Figure Description

[0033] Figure 1 This is a schematic diagram of the process of the present invention;

[0034] Figure 2 This is a schematic diagram of the super-resolution reconstruction network constructed in this invention;

[0035] Figure 3 A schematic diagram of the task adaptive routing strategy;

[0036] Figure 4 A schematic diagram of the layer attention module;

[0037] Figure 5 This is a schematic diagram of the incremental learning process;

[0038] Figure 6 This is a super-resolution reconstruction result of a medical image. Detailed Implementation

[0039] See Figure 1 The present invention performs incremental learning of a medical image super-resolution reconstruction network according to the following steps:

[0040] Step S1: Dataset Construction

[0041] A multimodal medical image dataset was constructed, treating each modality (e.g., Task 1: Modality A, Task 2: Modality B) as an independent learning task, with tasks arriving sequentially. When learning a new task, the complete training data for previous tasks is not accessible. The dataset is divided into a training set and a test set, with 4000 samples per task in the training set and 1000 samples per task in the test set.

[0042] Step S2: Dataset Preprocessing

[0043] The multimodal dataset constructed in step S1 is preprocessed. A corresponding low-resolution sample is generated for each high-resolution sample using bicubic interpolation downsampling, forming image pairs for supervised training of the super-resolution reconstruction network. Data augmentation methods, including random cropping, random flipping, random scaling, and random adjustment of image brightness, saturation, and chroma, are employed. To preserve anatomical structural information, a consistent transformation is applied to the image pairs to achieve data augmentation.

[0044] Step S3: Construction of the Super-Resolution Reconstruction Network

[0045] See Figure 2 , Figure 3 and Figure 4 A medical image super-resolution reconstruction network was constructed. This network adopts a Transformer-based encoder-bottleneck-decoder architecture.

[0046] The low-resolution image is passed through a 3×3 convolutional layer to obtain shallow features. At the same time, the low-resolution image is input in parallel into the instruction routing module. This module uses a convolutional encoder, a global average pooling layer, and a softmax layer to perform a weighted summation of the learnable instruction dictionary to generate a routing instruction vector. This routing instruction vector is repeatedly used during the forward propagation of the entire network to guide data of different modalities to select different processing paths in the network.

[0047] The shallow features then enter the encoder. The encoder consists of three layers, each with a spatial routing module preceding the Transformer block. The spatial routing module employs a hybrid expert mechanism, calculating gating weights based on the routing instruction vector and using a Top-K operator to select K expert networks (composed of multilayer perceptrons) to process the features. The output is a weighted sum of the outputs of the K expert networks. The spatially routed features then enter the Transformer block for deep feature extraction. Each Transformer block contains a multi-head self-attention mechanism and a feedforward neural network. After the Transformer block output, pixel rearrangement is performed to downsample and reduce spatial resolution while increasing the number of channels. The encoder repeats the above operations at multiple layers to progressively extract deep feature representations.

[0048] The deep features output from the last layer of the encoder enter the bottleneck section. The bottleneck section sets up a channel routing module before the Transformer block. This module also uses the routing instruction vector, passing it through a fully connected layer and a sigmoid activation function to generate a channel mask, which is then multiplied element-wise with the input features to achieve channel-level selection. The channel-routed features are then processed by the Transformer block for feature extraction, and after upsampling, they enter the decoder.

[0049] The decoder employs a multi-level structure symmetrical to the encoder. Features from different encoder levels are first input to the layer attention module, where spatial alignment, correlation matrix calculation, and weighted fusion are performed. The fused features integrate complementary information from shallow and deep layers. After channel-wise concatenation and 1×1 convolution adjustment, the fused features are passed to the corresponding decoder level. Each level also has a channel routing module before the Transformer block, and the working mechanism of the channel routing module is the same as that of the bottleneck part. The channel-routed features are extracted by the Transformer block, then upsampled through pixel rearrangement, and fused with the features processed by the layer attention module from the corresponding encoder level.

[0050] The decoder finally outputs a residual image generated by a 3×3 convolutional layer. The residual image is then added element-wise to the upsampled low-resolution input to generate a high-resolution image.

[0051] Step S4: Collaborative Strategies for Incremental Learning

[0052] The incremental learning strategies are designed, including adaptive memory buffer management, relational knowledge distillation, and dynamic gradient balancing training strategies, specifically as follows:

[0053] 4-1: Adaptive Memory Buffer Management Strategy

[0054] Maintain a fixed-size memory buffer. And manage it at each stage of incremental learning. Define an adaptive gradient-weighted difficulty score. Used to quantify any sample Importance to the current model. (Score) This incorporates information from two dimensions. The first is the difficulty of reconstructing the sample. The structural similarity (SSIM) between the current model's reconstruction result and the real high-resolution image is measured. Secondly, the learning influence of the samples is considered. The super-resolution reconstruction loss is measured by the gradient norm of the model parameters. These two dimensions are then combined using a weighted sum. ,in For balance coefficient, This is a min-max normalization operation. When training a new task, based on the stated score... Perform a priority sampling strategy for playback. After training, perform a buffer update process: (a) Calculate the scores for all new samples. and its deep feature representation The (a) Obtained by inputting low-resolution images into the encoder and bottleneck part of the current model. (b) Selecting scores. Top The new samples constitute the candidate pool. ,in , For the number of replacements. (c) In the candidate pool Sample The farthest point sampling algorithm is used in the feature representation space to select... Update the sample with the furthest distance between each feature. The specific steps of the farthest point sampling algorithm are as follows: from the candidate pool... Select rating The largest sample is added to the set as the initial selected sample. Then iteratively from Select with Minimum Euclidean distance of all samples Largest sample added until Include One sample. Finally, middle The quadruplets of the newly selected samples (low-resolution image, high-resolution image, score) and deep feature representation Store in buffer and remove middle rating Use the lowest number of old samples to keep the buffer size constant.

[0055] 4-2 Relational Knowledge Distillation Strategy

[0056] When training on a new task, the replayed old samples are used to constrain the feature geometric relationship structure of the current model on the old samples. When training on a new task, a batch of samples of size [number missing] is sampled from the memory buffer. The replay samples consist of low-resolution images, high-resolution images, and old deep feature representations. The old deep feature representations are feature vectors extracted and stored by the old model (encoder and bottleneck parts) when the task to which this sample belongs was trained. The low-resolution images in this batch are processed using the currently trained model to obtain the corresponding new feature representation batch. Based on the new and old feature representation batches, the Euclidean distance between all sample pairs within each batch is calculated, constructing two... The old and new distance matrices are obtained. The two distance matrices are normalized by dividing each element by the maximum value of all off-diagonal elements. The relational distillation loss is defined. The Frobenius norm is used to characterize the normalized distance matrix between batches of new and old features. This is achieved by minimizing... Constraining the relative distance relationship between sample pairs within a batch to remain consistent between the old and new models helps avoid forgetting.

[0057] 4-3 Dynamic Gradient Balancing Training Strategy

[0058] During incremental learning training, the weights of knowledge retention terms are adaptively adjusted in real time based on the degree of gradient conflict between new and old tasks. In each optimization step, three basic loss components are calculated: one is the super-resolution reconstruction loss calculated for the current batch of new task samples. Secondly, the super-resolution reconstruction loss calculated for the current batch of playback samples. Thirdly, the relationship distillation loss calculated for the current batch of playback samples. To assess the conflict between the old and new tasks, the gradients of the old and new tasks are calculated, and the cosine similarity between these two gradient vectors is also calculated. A dynamic scaling factor is then calculated based on this cosine similarity. Using the dynamic scaling factor To adjust the base weights of playback loss and relational distillation loss and To obtain dynamic weights and When a significant conflict is detected between the learning objectives of new and old tasks, the retention rate of old knowledge will be automatically increased. and The weights are allocated to prioritize the protection of the model's existing capabilities and mitigate catastrophic forgetting.

[0059] Step S5: Incremental learning training and use of the model

[0060] See Figure 5 This invention employs an incremental learning paradigm to train and update the network model constructed in step S3. During the network training phase, the AdamW optimizer is used for parameter optimization, which effectively handles weight decay and momentum issues. Specific parameter settings are as follows: the optimizer's weight decay is set to... , The parameters are set to (0.9, 0.99), and the initial learning rate is set to... A cosine annealing strategy was employed to dynamically adjust the learning rate. Two NVIDIA GeForce RTX 3090 graphics cards were used for training, with a 1:1 ratio of new task samples to replay samples in each training batch. The training epochs for each task were set to 100. This phased training strategy, alternating between intra-task training and inter-task buffer updates, ensures the stability and effectiveness of the model during incremental learning, making it particularly suitable for scenarios in medical image analysis that require processing multimodal medical data.

[0061] In each training iteration, from the current task New task sample batches are obtained by sampling from the training set. Simultaneously, according to the priority sampling method described in step 4-1, playback sample batches are sampled from the memory buffer. Total usage loss Backpropagation is performed to update the model parameters. The total loss function... It consists of three parts: new task super-resolution reconstruction loss Playback sample super-resolution reconstruction loss And relational knowledge distillation loss According to the dynamic gradient balancing strategy in step 4-3, and The weights are dynamic weights. and The total loss is .

[0062] The model is saved after all tasks have been trained. This can be used for super-resolution reconstruction of low-resolution images from all learned tasks in a sequence. In practical applications, for any low-resolution input image from a learned task, it is input into the model. Generate high-resolution output images. When new tasks emerge in the future... If necessary, the training process in this step can be repeated, starting from the task. New task samples are sampled from the training set, while replay samples are sampled from the memory buffer, using the total loss function. The trained model obtains updated parameters After training is complete, the buffer update process will be executed to update the task. High-scoring samples are added to the buffer to enable incremental learning and capacity expansion of the model, allowing it to continuously adapt to new medical image modalities without forgetting the knowledge it has already learned.

[0063] The present invention will be further described below through specific embodiments:

[0064] Embodiments of this invention utilize the publicly available IXI dataset and Chest X-ray dataset to construct a multi-stage, multimodal medical image super-resolution reconstruction task sequence. This sequence is divided into four stages: Task 1 (PD-weighted MRI), Task 2 (T1-weighted MRI), Task 3 (T2-weighted MRI), and Task 4 (chest X-ray images). The goal of this invention is to train a single model that, by sequentially learning these four tasks, can perform high-quality reconstruction of low-resolution images across all four modalities after incremental learning.

[0065] To verify the effectiveness of the method of this invention, a qualitative evaluation is performed on the final trained model. (See also...) Figure 6 The present invention presents super-resolution reconstruction results on random test samples of the latter three tasks (T1-weighted MRI, T2-weighted MRI, and chest X-ray images) in the sequence. By comparing the reconstruction results of the present invention side-by-side with the input low-resolution images, it can be clearly seen that the present invention successfully recovers fine structural and texture details, significantly improving image clarity. Especially after learning the subsequent chest X-ray image task, the model still maintains high-quality reconstruction capabilities on the older MRI task, confirming that the present invention effectively alleviates the catastrophic forgetting problem.

[0066] The above embodiments provide a detailed explanation of the incremental medical image super-resolution reconstruction method based on adaptive knowledge retention proposed in this invention. This invention successfully solves the catastrophic forgetting problem in medical image super-resolution reconstruction through the synergistic operation of three core strategies: adaptive memory buffer management, relational knowledge distillation, and dynamic gradient balancing. It is suitable for processing continuously emerging and complex medical images of different modalities. It should be noted that the above embodiments are merely specific application examples of this invention and do not constitute a limitation on the technical solution of this invention. All modifications, equivalent substitutions, or improvements made to the technical solution disclosed in this invention under the guidance of the technical ideas and basic principles of this invention, without departing from the spirit and scope of this invention, should be included within the protection scope of this invention.

Claims

1. A method for super-resolution reconstruction of medical images based on incremental learning, characterized in that, The method includes the following specific steps: Step S1: Dataset Construction A multimodal medical image dataset is constructed, where each modality is treated as an independent learning task, and the task sequence arrives in order, making it impossible to access data from old tasks when learning a new task; The dataset is divided into a training set and a test set. Each task in the training set contains 4,000 samples, and each task in the test set contains 1,000 samples. Step S2: Dataset Preprocessing The multimodal medical image dataset is preprocessed by generating corresponding low-resolution samples for each high-resolution sample using bicubic interpolation downsampling, forming image pairs for supervised training of the super-resolution reconstruction network. In addition, a variety of data augmentation methods, including random cropping, random flipping, random scaling, and random adjustment of image brightness, saturation, and chroma, are applied to the image pairs to ensure that the anatomical structure information is not distorted. Step S3: Construction of the Super-Resolution Reconstruction Network The super-resolution reconstruction network adopts a Transformer-based encoder-bottleneck-decoder architecture. Low-resolution images are processed by convolutional layers to extract shallow features. Simultaneously, the input routing module generates routing instruction vectors through a convolutional encoder, pooling layers, and a softmax layer to guide different modalities in selecting processing paths. The encoder contains multiple layers, each with a spatial routing module preceding the Transformer block. This spatial routing module employs a hybrid expert mechanism to select expert networks to process features based on the routing instruction vectors. After extracting deep features, the Transformer block reduces spatial resolution through downsampling. The bottleneck part has a channel routing module preceding the Transformer block, generating channel masks based on the routing instruction vectors to achieve channel-level feature selection. After feature extraction by the Transformer block, features are upsampled before entering the decoder. The decoder contains multiple layers. Features from different layers of the encoder are spatially aligned and weighted by the input layer attention module, then channel-wise concatenated and convolutionally adjusted before being passed to the corresponding layer of the decoder. Each layer has a channel routing module preceding the Transformer block for channel-wise selection. After feature extraction by the Transformer block, spatial resolution is restored through upsampling, and features from the corresponding layer of the encoder, processed by the layer attention module, are fused together. The final output of the decoder is passed through a convolutional layer to generate a residual image, which is then added to the upsampled low-resolution input to generate a high-resolution image. Step S4: Collaborative Strategies for Incremental Learning The incremental learning strategies are designed, including adaptive memory buffer management, relational knowledge distillation, and dynamic gradient balancing training strategies, specifically as follows: 4-1: Adaptive Memory Buffer Management Strategy A fixed-size memory buffer is maintained, and the value of samples is evaluated by an adaptive gradient-weighted difficulty score, which integrates the reconstruction difficulty and learning influence of the samples. During training, samples are preferentially sampled according to the scores. After the task is completed, the buffer is updated by selecting samples with both high value and diversity by combining the scores with the feature space farthest point sampling algorithm, and the deep feature representation of the samples is stored for knowledge distillation. 4-2: Relational Knowledge Distillation Strategy When training a new task, use replay samples to constrain the feature geometric relationship structure of the current model on old samples; Calculate the pairwise distance matrix of the new feature representation batch extracted by the current model and the pairwise distance matrix of the stored old feature representation batch. The knowledge distillation loss is the difference between the new and old distance matrices, thereby constraining the relational structure of the feature space to retain old knowledge. 4-3 Dynamic Gradient Balancing Training Strategy During training, the cosine similarity between the gradients of the new task and the gradients of the old task is calculated to assess the degree of gradient conflict. A dynamic scaling factor is calculated based on the degree of conflict. The weights of the replay sample loss and relation distillation loss are automatically adjusted using the dynamic scaling factor, and the protection of old knowledge is automatically increased when gradient conflicts occur. Step S5: Incremental learning training and use of the model Using the data from step S1, train the network from step S3 in task order, and perform incremental learning using the strategy from step S4. In each training iteration, sample new task samples from the current task dataset, sample replay samples from the memory buffer, and update the model using the total loss function, which includes new task super-resolution reconstruction loss, replay sample super-resolution reconstruction loss, and relation distillation loss. Update the memory buffer after task training is completed. After all tasks are trained, use the final model to perform super-resolution reconstruction on low-resolution images in the test set.

2. The medical image super-resolution reconstruction method based on incremental learning according to claim 1, characterized in that, The adaptive gradient-weighted difficulty scoring described in step 4-1 pass Calculations; among which, reconstruction difficulty The influence is measured by the Structural Similarity Index (SSIM) between the current model reconstruction result and the real high-resolution image. The gradient norm of the super-resolution reconstruction loss is used as a metric. For balance coefficient, This represents the min-max normalization operation; the feature space farthest point sampling algorithm iteratively selects samples from the candidate pool with high scores, and each time selects the sample that is farthest from the selected sample set in the deep feature space in the Euclidean distance, so as to ensure that the samples stored in the buffer have both high value and diversity.

3. The medical image super-resolution reconstruction method based on incremental learning according to claim 1, characterized in that, Knowledge distillation loss described in step 4-2 Defined as the Frobenius norm between the normalized distance matrices representing batches of new and old features; the normalization process involves dividing each element in the matrix by the maximum value of all off-diagonal elements of the matrix; by minimizing the loss... The relative distance between sample pairs within a batch must be kept consistent.

4. The medical image super-resolution reconstruction method based on incremental learning according to claim 1, characterized in that, The dynamic scaling factor described in step 4-3 The algorithm adaptively determines the gradient by calculating the cosine similarity between the gradient of the new task and the gradient of the old task; when gradient conflicts occur, the algorithm... Automatically enlarge; using the described Real-time adjustment of replay sample loss weights And the relationship between distillation loss weight ,in and These are the basic weighting coefficients.

Images