Parameter efficient fine-tuning method for medical image segmentation model based on transform domain learnable low-rank factors

Abstract

This invention presents an efficient fine-tuning method for medical image segmentation model parameters based on learnable low-rank factors in the transform domain. Building upon the framework of efficient fine-tuning of tensor singular value decomposition (t-SVD) parameters, this method constructs modules for a transform domain low-rank adapter, frequency-adaptive rank allocation, randomized SVD initialization, and dynamic weight reconstruction. By replacing the exact SVD decomposition with learnable low-rank factors in the Fourier transform domain, and employing real-number FFT conjugate symmetry compression and frequency-point energy-guided rank allocation strategies, it achieves synergistic optimization of computational efficiency and model performance. An rFFT parameter compression technique and a frequency-point energy-guided adaptive rank allocation algorithm are designed, utilizing conjugate symmetry and energy contribution analysis to achieve intelligent configuration of low-rank parameters at different frequencies. This method effectively solves the problems of high computational overhead, unstable parameter initialization, and insufficient utilization of frequency domain information in traditional t-SVD methods for medical image segmentation fine-tuning. Experiments show that this model exhibits excellent performance in both objective and subjective evaluation metrics.

Description

Technical Field

[0001] This invention relates to the interdisciplinary field of artificial intelligence, deep learning, medical image analysis, and transfer learning technologies. Specifically, it relates to a parameter-efficient fine-tuning system and method based on transform-domain learnable low-rank factorization, frequency-point adaptive rank allocation, and randomized SVD initialization. This system achieves efficient transfer of pre-trained medical image segmentation models to target tasks by constructing a transform-domain low-rank adapter, a frequency-point energy-guided rank allocation algorithm, a randomized SVD fast initialization strategy, and a dynamic weight reconstruction mechanism. This provides technical support for small-sample learning and computationally limited scenarios in the field of medical image analysis. Background Technology

[0002] The hippocampus is a crucial brain structure involved in cognitive functions such as memory and spatial navigation. In neuroimaging research and clinical applications, morphological changes in the hippocampus are important biomarkers for neurodegenerative diseases such as Alzheimer's disease; therefore, accurate segmentation is essential for quantitative analysis and pathological studies. However, due to the hippocampus's complex anatomy, blurred boundaries, and low contrast with surrounding tissues, accurate automated segmentation remains challenging. These factors exhibit significant anatomical variability among individuals, particularly under pathological conditions, making it difficult for universal segmentation algorithms to maintain consistent performance across different subjects.

[0003] In recent years, deep learning methods have made significant progress in the field of medical image segmentation. However, these methods have limitations when applied to hippocampal segmentation. Deep learning models typically require a large amount of labeled data for training, but medical image annotation is time-consuming and labor-intensive, and it is difficult to obtain sufficient labeled data. In addition, training these models requires substantial computational resources, which poses a significant challenge to research institutions and clinical environments with limited computing power.

[0004] Transfer learning improves the learning efficiency and performance of a model in one domain by leveraging knowledge gained from training it in another, while reducing the demand for data and computational resources to some extent. Traditional transfer learning methods typically fine-tune all parameters of a pre-trained model, requiring significant computational resources and training time, and are prone to overfitting. To address these issues, parameter-efficient fine-tuning methods have been proposed, which improve fine-tuning efficiency by adjusting only a small subset of the model's parameters, thus reducing the number of parameter updates.

[0005] Existing efficient parameter fine-tuning methods mainly include the following categories:

[0006] (1) Adapter method, which embeds an adapter module with a small number of trainable parameters into a pre-trained model;

[0007] (2) The LoRA method applies low-rank decomposition to the weight matrix and only updates the parameters of the low-rank part;

[0008] (3) Tensor-based decomposition methods, such as LoTR using tensor column decomposition and Fact using Tucker decomposition;

[0009] (4) LoRA-PT method, a parameter-efficient fine-tuning method based on tensor singular value decomposition.

[0010] Existing systems generally suffer from the following technical bottlenecks:

[0011] (1) The computational overhead is large. t-SVD requires performing exact SVD on each frequency domain slice, which results in high computational complexity.

[0012] (2) The parameter initialization is unstable. The initialization strategy of traditional methods is prone to causing training instability in scenarios with few samples.

[0013] (3) Frequency domain information is not fully utilized. Existing methods use the same rank for all frequency points, which fails to make full use of the non-uniform distribution characteristics of frequency domain information;

[0014] (4) Limited storage efficiency, a large number of frequency points generated by a complete FFT, resulting in high storage overhead. Summary of the Invention

[0015] This invention addresses the core technical bottlenecks of existing efficient parameter fine-tuning techniques, such as high computational overhead, unstable initialization, and insufficient utilization of frequency domain information. It proposes an efficient parameter fine-tuning system and method for medical image segmentation models based on transform-domain learnable low-rank factors. This system innovatively constructs five core technical modules:

[0016] (1) Transform domain low-rank adapter module, which innovatively uses learnable low-rank factors. It replaces the exact SVD decomposition, while maintaining the theoretical framework of t-SVD, and significantly reduces the computational complexity, achieving an efficient low-rank approximation of frequency domain slices;

[0017] (2) The rFFT parameter compression module utilizes the conjugate symmetry of the Fourier transform of real signals to store and calculate parameters at only half of the frequency points, achieving nearly 50% parameter compression and computation acceleration.

[0018] (3) Frequency point adaptive rank allocation module: By analyzing the energy contribution of each frequency point, it realizes the adaptive configuration of low frequency high rank and high frequency low rank under the fixed rank budget constraint, making full use of the non-uniform distribution characteristics of frequency domain information;

[0019] (4) Randomized SVD fast initialization module, which quickly obtains the main singular subspace through random projection and power iteration, improves the computation speed by about 16 times compared with the exact SVD, while maintaining the initialization quality and training stability;

[0020] (5) The weight dynamic reconstruction module periodically transforms the low-rank parameters in the frequency domain back to the original domain and updates the model weights during the training process to ensure the effectiveness of end-to-end training.

[0021] The core technological innovation of this invention lies in: proposing for the first time an efficient fine-tuning method for parameters in the transform domain using learnable low-rank factors to replace exact SVD decomposition; and innovatively designing a transform domain low-rank adapter module, through... Factorization enables flexible low-rank representation of frequency domain slices; an original adaptive rank allocation algorithm based on frequency point energy is developed, dynamically configuring the rank budget according to the information importance of each frequency point; randomized SVD is introduced for the first time into the initialization process of efficient parameter fine-tuning, and stable and efficient master subspace capture is achieved through exponential iteration; a complete rFFT parameter compression framework is constructed, achieving nearly 50% parameter saving by utilizing conjugate symmetry; and a dynamic weight reconstruction mechanism is established to ensure real-time updates of model weights during training.

[0022] The system was validated on three publicly available hippocampal segmentation datasets: EADC-ADNI, LPBA40, and HFH. In few-shot scenarios, compared to full parameter fine-tuning, this method improves the Dice coefficient by 0.57% and reduces HD95 by 0.15 mm on the EADC-ADNI dataset; improves the Dice coefficient by 0.72% on the LPBA40 dataset; and improves the Dice coefficient by 2.34% on the HFH dataset. The parameter update amount is only 3.16% of that of full parameter fine-tuning, and the initialization speed is improved by approximately 16 times. Experimental results demonstrate that this system exhibits excellent performance in terms of segmentation accuracy, parameter efficiency, and computational efficiency, demonstrating significant technical benefits and industrial application value. Attached Figure Description

[0023] Figure 1 This is the overall flowchart of the system of the present invention.

[0024] Figure 2 This is a diagram of the low-rank adapter module architecture in the transform domain of the present invention.

[0025] Figure 3 This is a diagram of the frequency-adaptive rank allocation structure of the present invention.

[0026] Figure 4 This is a schematic diagram of the frequency point energy distribution and cumulative energy threshold of the present invention.

[0027] Figure 5 This is an example diagram illustrating the segmentation effect of the present invention.

[0028] Figure 6 This is a graph showing the training process of the present invention. Detailed Implementation

[0029] The following detailed description, with reference to specific embodiments, illustrates the efficient fine-tuning method for medical image segmentation model parameters based on transform-domain learnable low-rank factors according to the present invention. This embodiment is based on the EADC-ADNI hippocampal segmentation dataset, which contains 135 MRI brain scan samples.

[0030] Example 1: Efficient Parameter Fine-Tuning System Based on Transform Domain Learnable Low-Rank Factor

[0031] This embodiment constructs a complete system integrating five core modules: tensor quantization, transform domain low-rank decomposition, adaptive rank allocation, fast initialization, and weight reconstruction. The system adopts an end-to-end design concept, where each module can be optimized independently or work collaboratively, exhibiting good scalability and maintainability.

[0032] Step 1: Data Acquisition and Preprocessing

[0033] (1) MRI data acquisition: T1-weighted MRI brain scan images were obtained from the EADC-ADNI database with an original resolution of 218×182×182 voxels to capture the fine anatomical features of the brain structure.

[0034] (2) Image preprocessing process: First, the skull is dissected to remove non-brain tissue areas; then, bias field correction is performed to eliminate intensity inhomogeneity during MRI acquisition; next, spatial standardization is performed to register all images to the MNI standard space; finally, intensity normalization is performed to scale the image intensity values ​​to the [0,1] interval.

[0035] (3) Hippocampal annotation standards: Establish hippocampal region annotation standards, which are accurately annotated by professional neuroimaging experts. The annotation range includes three anatomical subregions of the hippocampus: head, body, and tail. Professional annotation tools are used to delineate the boundaries to ensure annotation quality.

[0036] (4) Data augmentation techniques: Expand the size of the training dataset and improve the model's generalization ability by using data augmentation techniques such as rotation (-15° to +15°), flipping (horizontal and vertical), scaling (0.9 to 1.1 times), and intensity transformation (±10%).

[0037] (5) Dataset partitioning: A dataset containing 135 high-quality labeled samples was constructed, including 5 training samples, 30 validation samples, and 100 test samples. The effectiveness of the method in data-scarce scenarios was verified by using a small sample set.

[0038] Step 2: Loading the pre-trained model and tensor quantization of weights

[0039] A medical image segmentation model was built based on the UNETR architecture, with weights pre-trained on the BraTS2021 brain tumor segmentation dataset. UNETR is a hybrid architecture combining Transformer and U-Net, specifically designed for 3D medical image segmentation tasks.

[0040] (1) UNETR encoder structure: The encoder adopts a 12-layer Transformer structure with hidden dimensions. The system has 12 attention heads and a feedforward network dimension of 3072. The input image is first processed through a 3D convolutional layer for patch embedding, dividing the image into 16×16×16 patches, with each patch mapped to a 768-dimensional vector.

[0041] (2) UNETR Decoder Structure: The decoder consists of multiple upsampling modules, each containing a transposed convolutional layer, a batch normalization layer, and a ReLU activation function. Features from different layers of the encoder are fused with corresponding layers of the decoder through skip connections, achieving multi-scale feature fusion.

[0042] (3) Weight Matrix Extraction: Extract the weight matrix of the Transformer encoder from the pre-trained model. Each layer of the self-attention module contains four weight matrices: Query, Key, Value, and Output, for a total of 48. matrix.

[0043] (4) Self-attention weight tensor quantization: Stack the 48 self-attention weight matrices along the third dimension to form a tensor. This tensor contains the Q, K, V, and O projection matrices for all layers.

[0044] (5) MLP weight tensor quantization: 12 MLP upprojections Matrix stacking ; 12 projections under MLP Matrix stacking .

[0045] Step 3: Construction of the transform domain low-rank adapter module

[0046] The transform domain low-rank adapter module is the core innovative module of this invention. It achieves an efficient parameterized representation of pre-trained weights by constructing a learnable low-rank factor decomposition in the Fourier transform domain.

[0047] (1) Principle of rFFT: The real-number fast Fourier transform (rFFT) utilizes the conjugate symmetry of real-number signals to calculate and store only half of the frequency points. For a length of... The real number sequence, generated by rFFT A complex number of frequency points.

[0048] (2) Tensor rFFT transform: For example, perform a real fast Fourier transform along the third dimension. ,produce A frequency domain slice.

[0049] (3) Frequency domain slice structure: Each frequency domain slice for The complex matrix contains two components: a real part and an imaginary part. Utilizing conjugate symmetry, only 25 frequency points need to be stored and computed instead of 48, reducing the number of parameters by approximately 48%.

[0050] (4) Learnable low-rank factorization: for each frequency domain slice Learnable factors are used instead of exact SVD decomposition: ,in .

[0051] (5) Low-rank factor parameterization: and For learnable low-rank complex factors, For the frozen residual matrix, Let be the rank of that frequency point. This indicates the conjugate transpose.

[0052] (6) Complex parameter separation strategy: The real part-imaginary part separation method is used to process complex parameters. , .

[0053] (7) Implementation of complex matrix multiplication: .

[0054] (8) Parameter storage optimization: By separating the real and imaginary parts, each complex parameter is represented by two real parameters, which facilitates gradient calculation and parameter update in the standard deep learning framework.

[0055] Step 4: Randomized SVD initialization module

[0056] The randomized SVD initialization module quickly obtains the principal singular subspace of the weight matrix through random projection and exponential iteration, which is about 16 times faster than the exact SVD, while maintaining initialization quality and training stability.

[0057] (1) Generating random projection matrix: generating random Gaussian matrix ,in For the target rank, This is the oversampling parameter (usually 5-10). The elements of the random matrix follow a standard normal distribution. .

[0058] (2) Initial projection calculation: Calculate the projection This involves projecting the high-dimensional matrix onto a low-dimensional subspace. The computational complexity of this step is O(n log n). Far lower than accurate SVD .

[0059] (3) Power iteration enhancement: through The power iteration enhances the ability to capture the main singular direction. ,repeat Power iteration can significantly improve the suppression effect on directions corresponding to small singular values.

[0060] (4) Orthogonalization treatment: for Perform QR decomposition to obtain orthogonal basis Orthogonalization ensures the numerical stability of subsequent SVD calculations.

[0061] (5) Projected SVD calculation: Perform exact SVD in the low-dimensional subspace. , .

[0062] (6) Low-rank factor initialization: Learnable factors are initialized using approximate SVD results. .

[0063] (7) Initialization quality assurance: ensures that the initial quality is guaranteed. The best rank approximating the original matrix Approximation ensures the quality of the training starting point.

[0064] (8) Calculation of residual matrix: Calculate the frozen residuals ,Will Set them as non-trainable parameters.

[0065] Step 5: Frequency Adaptive Rank Allocation Module

[0066] The frequency-adaptive rank allocation module analyzes the energy contribution of each frequency point and achieves adaptive rank allocation under a fixed rank budget constraint, making full use of the non-uniform distribution characteristics of frequency domain information.

[0067] (1) Frequency point energy calculation: The Frobenius norm is calculated as the energy index for each frequency domain slice. .

[0068] (2) Energy normalization: Calculate the energy percentage at each frequency point. This makes the sum of the weights of all frequency points equal to 1.

[0069] (3) Rank budget allocation: in the total rank budget Under constraints, the rank of each frequency point is allocated according to its energy proportion. .

[0070] (4) Budget constraints are satisfied: all constraints are satisfied. The total rank is ensured to not exceed the budget through iterative adjustments.

[0071] (5) Low-frequency priority strategy: Considering that low-frequency components usually contain more structural information, a rank quota is added to the low-frequency points. in This is the low-frequency enhancement factor.

[0072] (6) Typical allocation result: Frequency point 0, which is the DC component. The energy content is the highest, containing average information between layers; frequency points 1-5, which are low frequencies. It contains the main modes that change slowly between layers; the frequency points are 6-24, which are high frequencies. It contains detailed information about rapid changes between layers.

[0073] (7) Physical meaning of adaptive rank allocation: Low-frequency components correspond to the slow change pattern of the weight matrix between layers, which usually contains more task-related information; high-frequency components correspond to the rapid change between layers, which may contain more noise.

[0074] (8) Parameter efficiency optimization: By adaptive rank allocation, more parameters are allocated to low-frequency components with large information content under the same total number of parameters, thereby improving parameter utilization efficiency.

[0075] Step 6: Training Process

[0076] Parameter freezing strategy: Freeze all parameters in the encoder except for the low-rank adapter in the transform domain, including the patchbedding layer, position encoding, layer normalization parameters, etc.

[0077] Decoder parameter settings: Keep the decoder parameters trainable, including the parameters of the transposed convolutional layer, batch normalization layer, and final classification layer.

[0078] Skip connection processing: Freeze the convolutional layer parameters in skip connections, retaining only the necessary feature transfer functionality.

[0079] Optimizer configuration: Collect all trainable parameters, including decoder parameters and transform domain low-rank adapter parameters. Factors; Adam optimizer was used with an initial learning rate of 0.001 and weight decay. .

[0080] Adam optimizer parameters: It employs the default momentum and adaptive learning rate mechanism.

[0081] Dynamic weight reconstruction mechanism: Frequency domain reconstruction is performed after each training step. .

[0082] Conjugate symmetry completion: Utilizing the conjugate symmetry of rFFT, from Recover the complete spectrum at each frequency point. .

[0083] Inverse FFT: Perform an inverse real FFT to return to the original domain. .

[0084] Destacking update: Tensor Split along the third dimension and update the weight matrix corresponding to each layer. .

[0085] Loss function design: Softmax Dice loss is used, and the Dice loss for the background and foreground is calculated separately. .

[0086] Advantages of Dice loss: Dice loss is naturally robust to class imbalance problems and is suitable for tasks such as hippocampal segmentation where the foreground region is small.

[0087] Learning rate scheduling: employs polynomial decay. ,in The total number of training rounds. This is the initial learning rate.

[0088] Training rounds setting: The total number of training rounds is set to 1000 rounds. Validation and evaluation are performed every 100 rounds, and the model with the best performance on the validation set is saved.

[0089] Early stopping strategy: If there is no performance improvement after 50 consecutive rounds of validation, training is terminated early to prevent overfitting.

[0090] Gradient clipping: Gradient norm clipping is used, with the maximum norm set to 1.0 to prevent gradient explosion.

[0091] Step 7: Deployment of Reasoning

[0092] Weight merging: After training is complete, the final weight reconstruction is performed, permanently merging the low-rank parameters in the frequency domain into the model weights.

[0093] Final reconfiguration process: Perform on all frequency points Then, perform an inverse FFT to obtain the final weight tensor.

[0094] Module cleanup: Delete all intermediate parameters of the transform domain low-rank adapter module, including... Factors and residual matrix Free up storage space.

[0095] Inference Deployment: The merged model is completely consistent with the original UNETR structure, requiring no additional computational overhead during inference, achieving zero-overhead inference deployment.

[0096] Model export: Supports exporting to standard formats such as ONNX and TorchScript, facilitating deployment on different platforms.

[0097] Inference optimization: The merged model can use standard inference optimization techniques, such as TensorRT acceleration and quantization.

[0098] The above description is merely a preferred embodiment of a method for efficient fine-tuning of medical image segmentation model parameters based on learnable low-rank factors in the transform domain. The scope of protection of this method is not limited to the above embodiments; all technical solutions falling within this conceptual framework are within the scope of protection of this invention. It should be noted that for those skilled in the art, any improvements and variations made without departing from the principles of this invention should also be considered within the scope of protection of this invention. This invention can be widely applied to medical image segmentation, transfer learning, efficient parameter fine-tuning, few-sample learning, and other related fields, possessing significant practical value and promising prospects for wider application.

Claims

1. A method for efficient fine-tuning of parameters in a medical image segmentation model based on a transform-domain learnable low-rank factor, characterized in that, Includes the following steps: Step 1: Obtain the weight matrix of the pre-trained medical image segmentation model and perform tensor quantization stacking operation; Step 2: Construct an efficient parameter fine-tuning framework that includes a transform domain low-rank adapter module, a frequency point adaptive rank allocation module, a randomized SVD initialization module, and a weight dynamic reconstruction module; Step 3: Perform real fast Fourier transform on the tensor weights along the third dimension in the transform domain low-rank adapter module, construct a learnable low-rank factor decomposition in the frequency domain, and integrate the residual freezing mechanism and complex parameter separation strategy to achieve an efficient low-rank approximate representation of the frequency domain slice; Step 4: In the frequency-adaptive rank allocation module, the rank budget is dynamically allocated according to the energy contribution of each frequency point. Through energy normalization and low-frequency priority strategy, the intelligent configuration of low-rank parameters at different frequencies is realized. Step 5: Use randomized SVD for low-rank factor initialization, and quickly obtain the principal singular subspace through random projection and power iteration to achieve stable and efficient parameter initialization; Step 6: Implement the periodic update of frequency domain parameters to original domain weights through a dynamic weight reconstruction mechanism, and establish an end-to-end efficient parameter fine-tuning system.

2. The method according to claim 1, characterized in that, The tensor stacking operation in step 1 is specifically as follows: The system extracts weight matrices from the Transformer encoder of the pre-trained UNETR model and stacks them into three categories according to their matrix size. Weights are extracted from the 12-layer Transformer structure to construct a complete set of weight tensors covering the self-attention module and the multilayer perceptron module. (1) Tensor quantization of self-attention module weights: Stack the 48 d×d matrices (query, key, value, and output projection matrices of the 12 Transformer layers) along the third dimension to form a tensor. ; (2) MLP Upprojection Weight Tensorization: Stack the 12-layer upprojection matrices to form a tensor. ; (3) MLP Downprojection Weight Tensorization: Stack the downprojection matrices of the 12 layers to form a tensor. ; Where d is the hidden dimension, the scattered weight matrix is ​​organized into a structured tensor through tensor quantization, laying the foundation for subsequent transform domain processing.

3. The method according to claim 1, characterized in that, The transform domain low-rank adapter module in step 3 is specifically as follows: A low-rank adapter system based on the Fourier transform domain is constructed. The purpose of this system is to efficiently represent and train the low-rank principal components of the model weights in the frequency domain, providing a high-quality low-rank approximate representation for subsequent efficient parameter fine-tuning. (1) rFFT Transform Design: Perform real fast Fourier transform on the tensor quantized weights along the third dimension: (Dimensionality Reduction Convolution) Utilizing the conjugate symmetry of real-valued signals, rFFT only generates Each frequency point saves nearly half the parameter storage and computation compared to the full FFT. (2) Learnable low-rank factorization: for each frequency domain slice Learnable factors are used instead of exact SVD: in and For learnable low-rank complex factors, For the frozen residual matrix, Let be the rank of that frequency point. This indicates the conjugate transpose. (3) Complex parameter separation strategy: The real part-imaginary part separation method is used to process complex parameters: Complex matrix multiplication is performed according to the following rules: 。 4. The method according to claim 1, characterized in that, The frequency point adaptive rank allocation module in step 4 is specifically as follows: This module analyzes the energy contribution of each frequency point and achieves adaptive rank allocation under a fixed rank budget constraint. It makes full use of the non-uniform distribution characteristics of frequency domain information to realize an intelligent configuration strategy of high rank at low frequencies and low rank at high frequencies. (1) Frequency point energy calculation: Calculate the Frobenius norm as an energy index for each frequency domain slice: (2) Energy normalization: Calculate the energy percentage at each frequency point: (3) Rank budget allocation, in the total rank budget Under constraints, the rank of each frequency point is allocated according to its energy proportion: Simultaneously satisfy the constraints: (4) Low-frequency priority strategy: Considering that low-frequency components usually contain more structural information, a rank quota is added to the low-frequency points: in This is the low-frequency enhancement factor. This represents the total number of frequency points. This strategy prioritizes the retention of important frequency domain information.

5. The method according to claim 1, characterized in that, The randomized SVD initialization strategy in step 5 is as follows: This strategy rapidly obtains the principal singular subspace of the weight matrix through random projection and exponential iteration, improving computation speed by approximately 16 times compared to exact SVD, while maintaining initialization quality and training stability. (1) Random projection: generating a random Gaussian matrix ,in For the target rank, For oversampling parameters: (2) Power iteration enhancement: through Power-law iteration enhances the ability to capture the main singular direction: , repeat Second-rate (3) Orthogonalization: for Perform QR decomposition to obtain orthogonal basis: (4) Projective SVD, performing exact SVD in a low-dimensional subspace: (5) Low-rank factor initialization: Initialize learnable factors using approximate SVD results: Make the initial time The best rank approximating the original matrix approximate.

6. The method according to claim 1, characterized in that, The dynamic weight reconstruction mechanism in step 6 is specifically as follows: This mechanism periodically reconstructs the low-rank parameters in the frequency domain back to the original domain weights during training, ensuring that the model uses the updated weights during forward propagation, thus achieving efficient end-to-end parameter fine-tuning. (1) Frequency domain reconstruction: combining frozen residuals and learnable low-rank terms: (2) Conjugate symmetric completion: Utilizing the conjugate symmetry of rFFT, the complete spectrum is recovered from $K$ frequency points: (3) Inverse FFT transformation: Perform an inverse real FFT to return to the original domain: (4) Destacking update: update the tensor Split along the third dimension and update the weight matrix corresponding to each layer: (5) Update frequency control: Set the update interval ,Every Perform a reconstruction update once per training step to balance computational overhead and training effectiveness.

7. The method according to claim 1, characterized in that, The method also includes a weight merging module for the inference phase, specifically: After training, the low-rank parameters are permanently merged into the model weights, eliminating the need for additional computational overhead during inference. (1) Final Reconstruction: Perform the complete frequency domain reconstruction and inverse transform process to obtain the final weight tensor. ; (2) Weight write-back: The weights of each layer after destacking are written into the corresponding parameters of the model; (3) Module cleanup: Delete all intermediate parameters of the transform domain low-rank adapter module and release storage space; \ (4) Inference deployment: The merged model has the same structure as the original model, and the inference speed and storage usage are the same as the original model, achieving inference deployment with zero additional overhead.

8. The method according to claim 1, characterized in that, The specific parameter efficiency index of the method is as follows: In the task of fine-tuning hippocampal segmentation in the UNETR model, the system parameter efficiency reached: (1) Original number of trainable parameters: approximately 28.31 million (full parameter fine-tuning); (2) Number of trainable parameters for the transform domain low-rank adapter: approximately 895,000 (rank) hour); (3) Parameter compression ratio: 31.6 times (only 3.16% of parameters need to be updated); (4) Parameter distribution of each module: 73776 parameters for self-attention module, 9228 parameters for MLP upward projection, and 9228 parameters for MLP downward projection; (5) Initialization speedup: Randomized SVD is about 16 times faster than exact SVD; (6) rFFT compression ratio: saves about 50% of frequency parameters compared to the full FFT.

9. The method according to claim 1, characterized in that, The training strategy of the method is as follows: An end-to-end training strategy and multi-stage learning rate scheduling are adopted to achieve rapid model convergence and performance optimization. Training parameter settings: batch size 4, initial learning rate 0.001, number of training epochs 1000, optimizer is Adam; Learning rate scheduling: Polynomial decay scheduling is used, the formula is: in The total number of training rounds. The initial learning rate; Loss function: Softmax Dice loss is used, and the Dice losses for the background and foreground are calculated separately and then summed with weights. Weight decay: ; Validation interval: Validation evaluation is conducted every 5 rounds; Weight update frequency: Perform transform domain low-rank adapter weight reconstruction once after each training batch.

10. A highly efficient fine-tuning system for medical image segmentation model parameters based on transform-domain learnable low-rank factors, characterized in that: This system employs an end-to-end parameter-efficient fine-tuning architecture, comprising five core components: a tensor quantization module, a transform-domain low-rank adapter module, a frequency-adaptive rank allocation module, a randomized SVD initialization module, and a weight dynamic reconstruction module. Through innovative transform-domain learnable low-rank factorization and frequency-adaptive rank allocation mechanisms, it achieves efficient transfer of pre-trained models to the target task. The system is based on a deeply improved t-SVD theoretical framework, using learnable factors... It replaces exact SVD decomposition and saves half of the parameters by using conjugate symmetry through rFFT transformation; it adopts randomized SVD for stable and efficient initialization and captures the main singular subspace quickly through power iteration; it combines an adaptive rank allocation strategy with frequency point energy analysis to achieve intelligent configuration of low-frequency high-rank and high-frequency low-rank; and it ensures real-time updating of model weights during training through a periodic weight reconstruction mechanism. The system can dynamically adjust the rank allocation based on the information importance of different frequency points: more ranks are allocated to low-frequency components to preserve structural information, and fewer ranks are allocated to high-frequency components to filter noise. On the EADC-ADNI hippocampus segmentation dataset, the system achieves a Dice coefficient of 87.46% and an HD95 distance of 5.31 mm. Compared with full-parameter fine-tuning, this represents a 0.57% improvement in Dice and a 0.15 mm reduction in HD95 distance. The parameter update amount is only 3.16% of that of full-parameter fine-tuning, and the initialization speed is improved by 16 times. This effectively solves the key technical problems of high computational cost, unstable initialization, and insufficient utilization of frequency domain information in traditional efficient parameter fine-tuning methods for medical image segmentation tasks, demonstrating significant technical innovation and practical value.

Images