Pharmacokinetic parameter estimation method and device, system and storage medium

Abstract

The invention discloses a pharmacokinetic parameter estimation method, device and system and a storage medium. The pharmacokinetic parameter estimation method comprises the steps that S1, an obtained breast tumor data sample is preprocessed; s2, according to the preprocessed breast tumor data sample, constructing and training a space-time separation fusion model oriented to DCE-MRI data; and S3, inputting to-be-processed breast tumor data into the trained time-space separation fusion model to obtain a pharmacokinetic parameter estimation value. By adopting the technical scheme provided by the invention, the PK parameter of the breast tumor can be accurately estimated.

Description

Technical Field

[0001] This invention belongs to the field of information processing technology, specifically relating to a method, apparatus, system, and storage medium for estimating pharmacokinetic parameters. Background Technology

[0002] DCE-MRI is an important medical imaging technique that uses contrast agents injected and acquires continuous time-series images to assess the hemodynamic characteristics of tissues. PK parameters (e.g.) K trans , V e , V p Accurate estimation of tumors is of great significance for clinical applications such as tumor diagnosis and treatment.

[0003] Traditional PK parameter estimation methods are mainly based on the Tofts model or extended Tofts model, which estimates PK parameters by fitting the time-signal intensity curve. However, these methods have the following problems: (1) Traditional methods require nonlinear fitting pixel by pixel, which is computationally inefficient, computationally intensive, and time-consuming; (2) DCE-MRI usually has a low signal-to-noise ratio, and traditional methods are sensitive to noise and easily affected by noise; (3) In regions with low signal-to-noise ratio or low contrast agent concentration, the parameter estimation error is large.

[0004] In recent years, deep learning technology has made significant progress in the field of medical image processing. Some studies have attempted to estimate PK parameters directly from DCE-MRI time series using convolutional neural networks (CNNs). However, existing methods have the following shortcomings: (1) Most methods treat the time dimension as a channel dimension, which does not make full use of time information and fails to effectively model the dynamic changes of time series; (2) There is a lack of effective mechanisms to capture features in both spatial and temporal dimensions, resulting in insufficient fusion of spatial and temporal features; (3) Frequency domain feature enhancement methods use the same processing strategy for local and global features, resulting in a single feature enhancement strategy and failing to fully utilize the characteristics of different types of features. Summary of the Invention

[0005] To address the problems existing in the prior art, this invention provides a method, apparatus, system, and storage medium for estimating pharmacokinetic parameters. It can effectively utilize time series information, achieve spatial-temporal feature separation processing, and employ a deep learning method with a differentiated feature enhancement strategy to improve the accuracy and efficiency of pharmacokinetic parameter estimation.

[0006] To achieve the above objectives, the present invention provides the following solution: A method for estimating pharmacokinetic parameters includes: Step S1: Preprocess the acquired breast tumor data samples; Step S2: Based on the preprocessed breast tumor data samples, construct and train a spatiotemporal separation and fusion model for DCE-MRI data; Step S3: Input the breast tumor data to be processed into the trained spatiotemporal separation and fusion model to obtain the estimated pharmacokinetic parameters.

[0007] As a preferred approach, the spatiotemporal separation fusion model for DCE-MRI data adopts a spatiotemporal separation encoder-decoder architecture, including: STR-UNet spatiotemporal residual module, STDD spatiotemporal decoupling dual-path module, and FDDAG spatiotemporal enhancement module.

[0008] As a preferred option, the total loss L for training the spatiotemporal separation and fusion model for DCE-MRI data is... total for: in, i ∈ { K trans , V e , V p}, L i For the first i The estimated loss for each parameter, σ i For learnable parameters optimized via backpropagation along with network weights, logσ i This is a regularization term.

[0009] The present invention also provides a pharmacokinetic parameter estimation device, comprising: The first processing module is used to preprocess the acquired breast tumor data samples; The second processing module is used to construct and train a spatiotemporal separation and fusion model for DCE-MRI data based on the preprocessed breast tumor data samples. The third processing module is used to input the breast tumor data to be processed into the trained spatiotemporal separation and fusion model to obtain pharmacokinetic parameter estimates.

[0010] As a preferred approach, the spatiotemporal separation fusion model for DCE-MRI data adopts a spatiotemporal separation encoder-decoder architecture, including: STR-UNet spatiotemporal residual module, STDD spatiotemporal decoupling dual-path module, and FDDAG spatiotemporal enhancement module.

[0011] As a preferred option, the total loss L for training the spatiotemporal separation and fusion model for DCE-MRI data is... total for: in,i ∈ { K trans , V e , V p}, L i For the first i The estimated loss for each parameter, σ i For learnable parameters optimized via backpropagation along with network weights, logσ i This is a regularization term.

[0012] The present invention also provides a pharmacokinetic parameter estimation system, comprising: a memory and a processor, wherein the memory stores a computer program executed by the processor, and the computer program performs a pharmacokinetic parameter estimation method when executed by the processor.

[0013] The present invention also provides a storage medium storing a computer program, which executes a pharmacokinetic parameter estimation method when running.

[0014] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention constructs a spatiotemporal separation and fusion model for DCE-MRI data. This model introduces a dual-path adaptive gating mechanism combining temporal residual connectivity and frequency domain enhancement, thereby significantly enhancing the representation capability of the temporal dimension and complex structures while maintaining computational efficiency. The innovations of this invention are mainly reflected in the following three aspects: This invention uses a 3D time encoder to extract deep spatiotemporal features, then restores spatial resolution through a 2D upsampling path, and introduces temporal residual connections at key locations: the original time series is projected onto a high-dimensional feature space using 3D convolution, and residual fusion is performed with the upsampled features on the complete time axis, effectively compensating for the information loss caused by time compression and enhancing the perception of dynamic process details.

[0015] Based on this, the present invention designs a branch-differentiated frequency domain processing strategy based on Fourier transform. In local branches, more aggressive high-frequency enhancement is used to highlight edges and textures, while in global branches, mild enhancement is used to maintain overall smoothness and robustness, thereby achieving differentiated amplification of information at different scales in the frequency domain.

[0016] Finally, through a dual-path adaptive gating module, based on the joint features in the spatial and frequency domains, the importance weights of the local and global branches at different spatial locations and channels are explicitly learned, achieving refined fusion and dynamic adjustment of the two feature paths. This design enables the local and global branches to better perceive each other's information, thus improving the overall accuracy and robustness of PK parameter estimation. Attached Figure Description

[0017] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart of the pharmacokinetic parameter estimation method according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the STDD module in this invention; Figure 3 This is a schematic diagram of the FDDAG module in this invention; Figure 4 This is a schematic diagram of the structure of the BDFDP module in this invention; Figure 5 This is a schematic diagram of the structure of the DAG module in this invention; Figure 6 This is a schematic diagram of the PK parameter estimation results obtained by the method of the present invention on synthetic data. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0021] Example 1 For ease of understanding, the English abbreviations and technical terms appearing in the embodiments of this invention are explained as follows: DCE-MRI: The full English name is dynamic contrast enhanced magnetic resonance imaging, which is abbreviated as dynamic contrast enhanced magnetic resonance imaging in this invention.

[0022] PK: The full English name is pharmacokinetics, which is abbreviated as pharmacokinetic parameters in this invention.

[0023] STSF: The full English name is spatial-temporal separation and fusion. This invention is simply referred to as spatial-temporal separation and fusion.

[0024] STR-UNet: The full English name is spatial-temporal residual UNet, and this invention is simply referred to as spatiotemporal residual UNet.

[0025] STDD stands for spatial-temporal decoupled dual-path, which is abbreviated as spatiotemporal decoupled dual-path in this invention.

[0026] FDDAG: The full English name is Fourier dual-path differential adaptive gating, and this invention is simply referred to as Fourier dual-path differential adaptive gating module.

[0027] BDFDP: The full English name is branch differentiated frequency domain processing, and this invention is simply referred to as branch differentiated frequency domain processing.

[0028] DAG: The full English name is dual-channel adaptive gating, which is abbreviated as dual-channel adaptive gating in this invention.

[0029] SDP: The full English name is spatial domain projection, which is abbreviated as spatial domain projection in this invention.

[0030] RMSE: Root mean square error, commonly used to measure the difference between predicted and true values. In image processing, it is calculated by taking the square root of the sum of the squares of the errors at each pixel between the predicted and true images. A smaller value indicates a closer pixel-level match between the predicted and true images, and a more accurate prediction.

[0031] SSIM stands for Structural Similarity Index Measure. It is used to comprehensively evaluate the similarity between a predicted image and a real image based on three aspects: brightness, contrast, and structure. The value ranges from 0 to 1. A value closer to 1 indicates greater similarity in structure and visual effect between the two images.

[0032] PSNR: Peak signal-to-noise ratio. It measures the consistency between a created or predicted image and the real image at the pixel intensity level. A higher value indicates less noise or distortion and higher overall image quality.

[0033] CNN: The full English name is convolutional neural network, which is abbreviated as CNN in this invention. like Figure 1 As shown, the present invention provides a method for estimating pharmacokinetic parameters, comprising: Step S1: Preprocess the acquired breast tumor data samples; Step S2: Based on the preprocessed breast tumor data samples, construct and train a spatiotemporal separation and fusion model for DCE-MRI data; Step S3: Input the breast tumor data to be processed into the trained spatiotemporal separation and fusion model to obtain the estimated pharmacokinetic parameters.

[0034] As one embodiment of the present invention, step S1 includes: S1.1 Obtaining breast tumor data The Extended Tofts Model is a classic model widely used in the field of quantitative analysis of PK parameters. Based on the Extended Tofts Model, a dataset was generated using the population mean arterial input function (AIF) for model training and validation. This dataset contains 1000 synthetic DCE-MRI samples, totaling 20,000 image time series and corresponding PK parameter maps. Each sample contains 9 image time series (20-second intervals, total duration approximately 2.67 minutes), along with the corresponding PK parameter map and region of interest (ROI) mask. Here, the spatial resolution of the images is 100×100.

[0035] To simulate the spatial continuity and heterogeneity of tumor internal parameters, this invention uses a Gaussian smoothed random field to generate the simulated PK parameter map, with the parameter range set as follows: K trans ∈ [0.10, 0.35] min -1 , V e ∈ [0.25, 0.60]、 V p ∈ [0.02, 0.08]. This method ensures a smooth transition of parameter values ​​between adjacent pixels, which is more consistent with real physiological conditions. The two-dimensional Gaussian kernel function used is as follows: The background region also contains signal values, simulating the perfusion characteristics of normal tissue. All samples use realistically extracted tumor masks as ROIs to ensure that the spatial morphology of the synthetic data is consistent with that of the real data.

[0036] S1.2 Data Preprocessing and Partitioning Strategies In the data preprocessing stage, the DCE-MRI time series were standardized: first, the images of each sample at 9 time points were uniformly scaled to a preset size (in this experiment, the original pixel size was kept at 100×100). Then, the signal intensity S(t) at each time point was normalized by Z-score using the following formula to obtain S_norm(t): For the PK parameter map, Min-Max normalization is used to map the parameters to the [0, 1] interval to eliminate the impact of differences in parameter dimensions on model training. For each parameter P, its normalized value P_norm is calculated using the following formula: Time series data is organized along the time dimension in the format [T, H, W], where T=9 represents the number of time points, and H and W represent the height and width of the image, respectively.

[0037] For data partitioning, a five-fold cross-validation strategy was adopted, randomly dividing the 20,000 samples into five mutually exclusive subsets, each containing approximately 4,000 samples. This ensured that all 20 layers of data for each participant belonged to the same fold, preventing data leakage. In each fold cross-validation, one subset was selected sequentially as the test set, and the remaining four subsets were used as the training set. Finally, the mean and standard deviation of the five-fold test results were obtained to comprehensively evaluate the model's generalization performance.

[0038] As one embodiment of the present invention, in step S2, the spatiotemporal separation fusion model for DCE-MRI data adopts a spatiotemporal separation encoder-decoder architecture, which consists of three main parts: STR-UNet spatiotemporal residual module, STDD spatiotemporal decoupling dual-path module, and FDDAG spatiotemporal enhancement module.

[0039] First, the input DCE-MRI time series is reconstructed into a five-dimensional tensor and fed into a 3D temporal encoder. This encoder progressively compresses the temporal and spatial resolution through three layers of 3D convolutions and downsampling, preserving intermediate features at each scale for subsequent skip connections. In the bottleneck layer, average pooling is performed on the temporal dimension to achieve temporal compression, resulting in the 2D spatiotemporal encoded features and multi-scale skip features of the bottleneck layer. At the decoding end, the network uses upsampling to progressively restore the spatial resolution. First, the bottleneck layer uses a three-layer transposed convolution upsampling path, concatenating and fusing with skip features from each scale of the encoder to form progressively refined high-resolution feature maps. After the final upsampling stage, the resulting 2D features are first extended through temporal recovery and temporal residual connections. This expands the features and maps the original time series through 1×1×1 convolutions, adding them to the extended features, thus explicitly injecting the original temporal information into the top layer of STR-UNet to compensate for the information loss caused by temporal compression. Subsequently, the features with restored complete spatiotemporal dimensions are fed into the STDD module. STDD first employs 1×3×3 convolutions to construct local paths and dilated convolutions to construct global paths in the spatial stage, performing multi-scale modeling of the spatial structure of each time frame. Then, in the temporal stage, 3×1×1 convolutions are used to similarly construct local and global paths, decoupling the modeling of short-term and long-term dynamics across the entire time axis, ultimately outputting spatiotemporal fusion features with doubled channel count. After compressing the temporal dimension into 2D features using mean pooling, the FDDAG module is introduced. Differentiated enhancement strategies are applied to the local and global branches in the frequency domain (aggressive high-frequency enhancement for the local branch and moderate enhancement for the global branch). A spatially aware adaptive gating network then assigns position- and channel-related weights to the two feature paths, achieving Fourier dual-path differential adaptive fusion. Finally, the fused features are mapped to a three-channel parameter map by the final convolutional head, thus completing the end-to-end estimation of PK parameters from DCE-MRI time series.

[0040] Furthermore, the STR-UNet spatiotemporal residual module performs the following operations: The encoder uses 3D convolution to process the temporal dimension, progressively compressing the temporal and spatial resolution through three layers of downsampling. The first layer downsamples the input [B, 1, T, H, W] to [B, base_channels, T / 2, H / 2, W / 2] using two 3×3×3 convolutions and max pooling. The second layer doubles the number of channels and continues downsampling to [B, 2·base_channels, T / 4, H / 4, W / 4]. The third layer doubles the number of channels again and downsamples to [B, 4·base_channels, T / 8, H / 8, W / 8]. Intermediate features (x11, x21, x31) are saved at each layer for subsequent skip connections. Finally, temporal average pooling compresses the 3D features into a 2D bottleneck layer feature [B, 4·base_channels, H / 8, W / 8], completing the initial encoding and compression of temporal information.

[0041] The decoder employs a three-layer upsampling approach to progressively restore spatial resolution. The first layer upsamples the bottleneck layer features from H / 8 to H / 4, concatenates them with the encoder's third-layer skip features (x31), and then performs convolutional fusion. The second layer upsamples from H / 4 to H / 2, fusing the encoder's second-layer skip features (x21). The third layer upsamples from H / 2 to H, fusing the encoder's first-layer features (x11). Each layer uses a transposed convolution for 2x upsampling and adjusts the number of channels in the skip features using a 1×1 convolution to ensure channel consistency after concatenation. This UNet-style skip connection preserves multi-scale detail information, contributing to the accurate reconstruction of spatial structure.

[0042] After upsampling, STR-UNet performs temporal dimension restoration and temporal residual connection: the 2D feature [B, ngf, H, W] is expanded to [B, ngf, T, H, W]; simultaneously, the original time series [B, T, H, W] is projected to [B, ngf, T, H, W] through a 1×1×1 convolution, and then added to the expanded feature. The mathematical expression is as follows: Where Y is the spatiotemporal feature of the final output of STR-UNet, X is the original DCE-MRI time series input, Z is the 2D feature output by the encoder at the bottleneck layer, UpFuse(·) represents the upsampling path of the decoder, ChannelMap(·) represents the channel mapping of the original input, and Expand(·) is the operation of expanding the 2D feature map output by the decoder into a 3D tensor along the time dimension.

[0043] The design injects the original time information into the upsampled features, and its feasibility and necessity are reflected in the following two aspects: (1) Compensation for information loss: Although the time dimension average pooling operation in the encoder effectively compresses the features, it also leads to the loss of time dynamic information. The time residual connection designed in this invention can effectively recover most of the lost time information by directly projecting the original time series into the decoder, providing a complete basis for dynamic changes for subsequent modules. This is crucial for accurately capturing the rapid infusion and clearing stages of the contrast agent. (2) Optimization of gradient propagation: From the perspective of network training, this residual connection creates a direct path from input to output, allowing the gradient to flow directly to the shallow layers of the network. This effectively alleviates the gradient vanishing problem that may be caused by network deepening and 3D to 2D dimension transformation, thereby accelerating the model convergence speed and improving the final performance. Therefore, STR-UNet achieves end-to-end mapping from time series to high-resolution spatiotemporal features through multi-scale feature fusion and time residual connection.

[0044] Furthermore, such as Figure 2 As shown, the STDD spatiotemporal decoupling dual-path module performs the following operations: The STDD module decouples spatial and temporal dimensions in its modeling, extracting multi-scale features separately in the spatial and temporal stages, thus possessing stronger modeling capabilities. The spatial stage employs 1×3×3 convolutions, processing the spatial dimension while preserving the temporal dimension. First, a 1×3×3 convolutional projection is performed through spatial domain projection, mapping the input features to the spatial processing space.

[0045] Two parallel paths are then constructed: a spatial local path uses standard 1×3×3 convolutions to capture local spatial patterns; and a spatial global path uses dilated 1×3×3 convolutions (spatial dilation rates = 2, 4, 8) to expand the receptive field and capture the global spatial context. The two paths are fused through residual connections to form spatially enhanced features. In the temporal phase, 1×3×3 convolutions are used to process the temporal dimension while maintaining the spatial dimension. First, the spatially enhanced features are mapped to the temporal processing space through temporal projection. Then, two parallel paths are constructed: a temporal local path uses standard 3×1×1 convolutions to capture short-term temporal dependencies; and a temporal global path uses dilated 3×1×1 convolutions (temporal dilation rates = 1, 2, 4) to expand the temporal receptive field and capture long-term temporal dependencies. The two paths are concatenated and fused to output a spatiotemporal fusion feature [B, 2·ngf, T, H, W] with doubled channel count. This decoupled design optimizes spatial and temporal modeling separately, improving the representation ability of complex spatiotemporal patterns while maintaining computational efficiency. Its core computational flow can be summarized as follows: Where Y is the spatiotemporal feature of the final output of STR-UNet, Local_s(·) and Global_s(·) represent the local and global paths in the spatial stage, respectively, Local_t(·) and Global_t(·) represent the local and global paths in the temporal stage, respectively, and Concat(·) means concatenating the outputs of the temporal stage in the channel dimension to form the final output F_out.

[0046] Furthermore, such as Figure 3 As shown, the FDDAG spatiotemporal enhancement module performs the following operations: The FDDAG module includes a spatial domain projection (SDP), a branched differential frequency domain processing (BDFDP) module, and a dual-path adaptive gating (DAG) module. These three modules enable refined fusion of local and global features.

[0047] First, spatial domain projection is performed on both the local and global branches. The local branches map the input features to the processing space using 1×1 convolutions, batch normalization, and ReLU activation. Similarly, the global branches are projected using the same structure. This projection provides normalized features for subsequent frequency domain processing and enables preliminary channel transformation and nonlinear activation.

[0048] like Figure 4 As shown, the BDFDP submodule employs different frequency domain enhancement strategies for the projected local and global branches. The theoretical basis for this differentiated enhancement strategy is that the frequency domain energy of medical images is mainly concentrated in the low-frequency region, while the high-frequency region contains detailed information such as edges and textures, but is also more susceptible to noise interference. The linear gain function proposed in this invention, boost = a + b·r (where r is the radial frequency), can enhance high-frequency components and image details while controlling the degree of suppression of low-frequency information and noise by adjusting parameters a and b.

[0049] First, the spatial domain features are converted to the frequency domain using Fast Fourier Transform (FFT), separating the amplitude and phase spectra. For local branches, an aggressive high-frequency enhancement strategy is employed to highlight fine structures such as tumor boundaries: boost = 0.5 + 1.0 This function effectively suppresses low-frequency background signals while significantly amplifying high-frequency edge information. A relatively high limiting threshold (3.0 times the mean) is set to prevent noise from being excessively amplified. For global branches, a gentle enhancement strategy is adopted to maintain the integrity and smoothness of the organizational structure: boost = 1.0 + 0.5 The function r moderately enhances high-frequency components while preserving the original low-frequency information. A more conservative amplitude limiting threshold (2.0 times the mean) is set to ensure the robustness of the features. The processed amplitude spectrum is reconstructed from the original phase spectrum, transformed back to the spatial domain via IFFT, and then subjected to 1×1 convolutional projection and residual connections to obtain the frequency domain enhanced features.

[0050] like Figure 5 As shown, the DAG submodule adaptively fuses the local and global features enhanced by BDFDP. First, the two feature paths are concatenated and fed into a spatially aware gating network: the first layer uses a 3×3 convolution to capture the neighborhood context, generating a spatially aware intermediate representation; the second layer uses a 1×1 convolution for channel compression; the third layer uses a 1×1 convolution to generate gating weights, and the weights are restricted to the [0, 1] interval using a Sigmoid activation function. The generated gating weights are divided into local and global gating parts, which are multiplied element-wise with the corresponding enhanced features to achieve adaptive weighting related to position and channel. Finally, the weighted two feature paths are concatenated and fused through a fusion convolutional layer (3×3 convolution + 1×1 convolution) to output a fused feature of [B, 2·ngf, H, W]. The gating weighting and fusion process can be represented by the following formula: Where F_local and F_global are the local and global features after frequency domain enhancement by BDFDP, and f_gate(·) represents a gated weight generation network consisting of three convolutional layers. (·) is the Sigmoid activation function, which normalizes the output to gate weights in the interval [0, 1]. G_local and G_global are weights separated from G, corresponding to local and global paths, respectively. The first layer represents element-wise multiplication, i.e., applying gated weights. `f_fuse(·)` is the final fusion convolutional layer, used to integrate the weighted features and produce the final output `F_out`. The visualization of the output is shown below. Figure 6 As shown.

[0051] The spatiotemporal separation and fusion model for DCE-MRI data is trained as follows: This invention employs an end-to-end deep learning network framework for training and uses five-fold cross-validation to evaluate the model's PK parameter estimation performance. The dataset is divided into five non-overlapping subsets. Each fold uses four subsets as the training set and the remaining subset as the test set. This process is repeated five times to ensure that each subset is used as a test set once. Each fold is trained from scratch for 100 epochs.

[0052] To optimize PK parameter estimation, this invention defines a combination loss L.param This includes L1 loss and L2 loss (i.e., mean squared error). L1 loss focuses on the parameter estimation error of each pixel and is more robust to outliers; L2 loss optimizes the smoothness of the overall parameter distribution and emphasizes large errors. Combined loss L... param The specific form is: Where λ1 and λ2 are the balancing weights. For a single pixel, L1 and L2 are defined as follows: Where y pred The parameter values ​​predicted by the model, y true The parameter values ​​are based on synthetic data. To address the weighting imbalance in parameter estimation and avoid a single parameter dominating the optimization, a dynamic weighting strategy based on task uncertainty is adopted. This is achieved by introducing a learnable parameter σ related to the three parameter estimation tasks. K trans σ V e σ Vp The model automatically adjusts the contribution of each parameter to the total loss. These parameters represent the uncertainty of the model's estimates for each parameter. During training, the model assigns smaller weights to parameters with higher uncertainty (larger loss values) and larger weights to parameters with lower uncertainty. Total loss L total Defined as: in, i ∈ { K trans , V e , V p}, L i For the first i The estimated loss for each parameter, σ i These are learnable parameters that, along with network weights, are optimized through backpropagation. The regularization term is logσ. i To prevent excessive increases in uncertainty parameters and ensure reasonable weight allocation, the Adam optimizer is used to update all learnable parameters during training. The initial learning rate is set to 1e-5, and the weight decay is 1e-4 to prevent overfitting. A LambdaLR learning rate regulator is employed, following a cosine annealing strategy: the learning rate smoothly decreases from its initial value over the training period, aiding model convergence in later stages. This strategy ensures both rapid early learning and fine-tuning in later stages, improving convergence stability and final performance.

[0053] To monitor training progress, prevent overfitting, and select the final model for final testing, this invention employs a complete validation and model selection process. During each training epoch, the model is evaluated on an independent validation set after each training cycle. The validation phase switches to evaluation mode, without gradient calculation or parameter updates. To evaluate the performance of PK parameter estimation, multiple metrics are used: Root Mean Square Error (RMSE) measures the PK parameter estimation error; Structural Similarity Index (SSIM) assesses the structural similarity between the predicted and true parameter maps; and Peak Signal-to-Noise Ratio (PSNR) evaluates the overall quality.

[0054] At the end of each epoch, if the current validation loss is lower than the lowest value recorded in all previous epochs, the current model is saved as the optimal model. This process continues throughout the entire training epoch, and the finally saved model is used as the optimal model for that epoch in the final testing phase. In the final testing phase, SSIM and PSNR are used to evaluate the model's performance on PK parameter estimation. These metrics are used to quantify model performance, but model selection is based on the validation loss.

[0055] The proposed STSF model was compared with several existing PK parameter estimation models, as shown in Table 1. Experimental results show that the STSF model achieves better performance on the PK parameter estimation task. As can be seen from the table, this model demonstrates superiority in all core metrics. In the SSIM metric, which measures parameter estimation accuracy, thanks to the temporal residual connection design of the STR-UNet decoder, which preserves original temporal information and compensates for information loss caused by temporal compression, this model achieves a significant improvement, indicating that the predicted parameter map is structurally closer to the true parameter distribution. Simultaneously, in the PSNR metric, which measures overall quality, due to the spatial-temporal decoupling dual-path design of the STDD module, which optimizes feature extraction in both spatial and temporal dimensions, this model also achieves a significant improvement, indicating less noise and distortion in the prediction results. Thanks to the differentiated frequency domain processing and dual-path adaptive gating mechanism of the FDDAG module, the model can perform refined feature enhancement and fusion tailored to the characteristics of different parameters.

[0056] Experimental results show that the STSF model achieves more accurate and robust predictions in the PK parameter estimation task through techniques such as temporal residual connection, spatial-temporal decoupling dual-path, and Fourier dual-path differentiated adaptive gating, providing more reliable parameter estimation results for clinical diagnosis.

[0057] Table 1 Example 2 The present invention also provides a pharmacokinetic parameter estimation device, comprising: The first processing module is used to preprocess the acquired breast tumor data samples; The second processing module is used to construct and train a spatiotemporal separation and fusion model for DCE-MRI data based on the preprocessed breast tumor data samples. The third processing module is used to input the breast tumor data to be processed into the trained spatiotemporal separation and fusion model to obtain pharmacokinetic parameter estimates.

[0058] As one embodiment of the present invention, the spatiotemporal separation fusion model for DCE-MRI data adopts a spatiotemporal separation encoder-decoder architecture, including: STR-UNet spatiotemporal residual module, STDD spatiotemporal decoupling dual-path module, and FDDAG spatiotemporal enhancement module.

[0059] As one embodiment of the present invention, the total loss L for training a spatiotemporal separation and fusion model for DCE-MRI data is... total for: in, i ∈ { K trans , V e , V p}, L i For the first i The estimated loss for each parameter, σ i For learnable parameters optimized via backpropagation along with network weights, logσ i This is a regularization term.

[0060] Example 3 The present invention also provides a pharmacokinetic parameter estimation system, comprising: a memory and a processor, wherein the memory stores a computer program executed by the processor, and the computer program performs a pharmacokinetic parameter estimation method when executed by the processor.

[0061] Example 4 The present invention also provides a storage medium storing a computer program, which executes a pharmacokinetic parameter estimation method when running.

[0062] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for estimating pharmacokinetic parameters, characterized in that, include: Step S1: Preprocess the acquired breast tumor data samples; Step S2: Based on the preprocessed breast tumor data samples, construct and train a spatiotemporal separation and fusion model for DCE-MRI data; Step S3: Input the breast tumor data to be processed into the trained spatiotemporal separation and fusion model to obtain the estimated pharmacokinetic parameters.

2. The method for estimating pharmacokinetic parameters as described in claim 1, characterized in that, The spatiotemporal separation and fusion model for DCE-MRI data adopts a spatiotemporal separation encoder-decoder architecture, including: STR-UNet spatiotemporal residual module, STDD spatiotemporal decoupling dual-path module, and FDDAG spatiotemporal enhancement module.

3. The method for estimating pharmacokinetic parameters as described in claim 2, characterized in that, The total loss L for training a spatiotemporal separation and fusion model for DCE-MRI data total for: in, i ∈ { K trans , V e , V p }, L i For the first i The estimated loss for each parameter, σ i For learnable parameters optimized via backpropagation along with network weights, logσ i This is a regularization term.

4. A pharmacokinetic parameter estimation device, characterized in that, include: The first processing module is used to preprocess the acquired breast tumor data samples; The second processing module is used to construct and train a spatiotemporal separation and fusion model for DCE-MRI data based on the preprocessed breast tumor data samples. The third processing module is used to input the breast tumor data to be processed into the trained spatiotemporal separation and fusion model to obtain pharmacokinetic parameter estimates.

5. The pharmacokinetic parameter estimation device as described in claim 4, characterized in that, The spatiotemporal separation and fusion model for DCE-MRI data adopts a spatiotemporal separation encoder-decoder architecture, including: STR-UNet spatiotemporal residual module, STDD spatiotemporal decoupling dual-path module, and FDDAG spatiotemporal enhancement module.

6. The pharmacokinetic parameter estimation device as described in claim 5, characterized in that, The total loss L for training a spatiotemporal separation and fusion model for DCE-MRI data total for: in, i ∈ { K trans , V e , V p }, L i For the first i The estimated loss for each parameter, σ i For learnable parameters optimized via backpropagation along with network weights, logσ i This is a regularization term.

7. A pharmacokinetic parameter estimation system, characterized in that, include: A memory and a processor, wherein the memory stores a computer program executed by the processor, the computer program performing the pharmacokinetic parameter estimation method as described in any one of claims 1-3 when executed by the processor.

8. A storage medium, characterized in that, The storage medium stores a computer program that, when executed, performs the pharmacokinetic parameter estimation method as described in any one of claims 1-3.

Images