Artery input function automatic selection method based on attention mechanism in DCE-MRI

Abstract

The invention discloses an automatic artery input function selection method based on an attention mechanism in DCE-MRI, and the method comprises the steps: carrying out the fast Fourier transform of data, and obtaining the corresponding amplitude and phase; inputting double-domain data into an HDC Block residual convolutional network layer, capturing multi-scale features, and fusing shallow-layer details and deep-layer semantic features through residual connection; data feature representation output by the HDC Block layer enters the RWK layer to realize feature time sequence dependent modeling, and parameter learning of the RWKV layer is completed in advance through a sequence reconstruction task; classifying all data by utilizing a classification head; and according to a category priority principle, performing score calculation to obtain a final highest score AIF. According to the method, accurate and efficient AIF automatic selection with clinical practicability can be realized, the ability of capturing blood flow dynamic details is remarkably enhanced, the strong robustness of an algorithm to scanning parameter differences is kept, and the clinical uncertainty of perfusion parameter interpretation is effectively reduced.

Description

Technical Field

[0001] The present invention belongs to the technical field of medical image analysis, and in particular relates to a method for automatically selecting arterial input functions based on an attention mechanism in DCE-MRI. Background Art

[0002] In direct coronary artery perfusion imaging (DCE), accurate selection of the arterial input function (AIF) for calculating key physiological parameters such as cerebral blood flow (CBF), cerebral blood volume (CBV), and mean transit time (MTT) is crucial. By quantifying the dynamic concentration-time curve of contrast agents within arterial vessels, the AIF provides important prior information for perfusion imaging. However, AIF exhibits significant inter-individual variability due to individual physiological differences, scanning noise, and pathological factors. Traditional manual selection methods are not only time-consuming but also highly subjective, with inter-operator judgment errors, which seriously compromises the robustness and reproducibility of the results. In contrast, automated AIF selection methods, through an algorithm-driven, objective process, can significantly improve the accuracy and consistency of perfusion analysis. Therefore, the development of an automated AIF selection method that combines high accuracy, high efficiency, and high applicability has become an urgent need in the field of medical image analysis.

[0003] Existing methods can be broadly categorized into two mainstream techniques: clustering-based methods and deep learning-based methods. Clustering-based methods, such as the K-Means algorithm, utilize shape features (such as tracer arrival time, peak intensity, and clearance rate) to screen candidate curves for initial voxel selection. Subsequently, a two-stage clustering process is employed to refine the candidate curves: the first stage aims to maximize the similarity within clusters to generate a candidate set, while the second stage further filters the curves with the lowest first moment as the final AIF.

[0004] As an important branch of machine learning, deep learning is centered on automatically extracting high-level features of data through multi-layer nonlinear transformations. Compared with traditional machine learning methods, deep learning effectively solves the limitations of manual feature design through an end-to-end training model, and shows significant advantages in image processing and time series data analysis. As a typical architecture of deep learning, convolutional neural network (CNN) has been widely used in the field of medical image analysis due to its local connection, weight sharing and spatial downsampling characteristics. The Winner team innovatively applied CNN to the automatic detection of AIF. This method uses a three-layer convolutional pooling structure to classify the concentration-time curves and uses the cross-entropy loss of the Adam optimizer to maximize the cross-correlation between the predicted AIF and the annotated data. During the inference process, the top 150 curves are selected according to the scores, and the geometric mean method is used to synthesize the final AIF.

[0005] Using classic clustering algorithms to solve the problem is undoubtedly feasible, but it also has obvious limitations. On the one hand, K-Means relies on a manually preset number of clusters, which can easily lead to underfitting or overfitting. On the other hand, when the signal-to-noise ratio is low, local outliers will seriously affect the stability of clustering, and the reliance on shallow shape features makes it difficult to capture the deep dynamic features of the curve. The CNN-based method requires large-scale annotated data, which results in high labeling costs, and the differences in scanning parameters also limit its versatility. More importantly, the local receptive field of CNN makes it difficult to simulate the long-term temporal dependence of the AIF curve, and the average synthesis strategy may lead to the loss of dynamic details, thereby affecting the accuracy of the perfusion parameters.

[0006] To break through the bottleneck of the above methods in time series modeling, researchers began to explore the Transformer architecture with global perception capabilities. This architecture establishes a dynamic weight association between any two time points through the self-attention mechanism. Its theoretical memory span can cover the entire time series, which provides a new paradigm for solving the long-term dependency modeling of the AIF curve. The Swin Transformer further introduces a hierarchical window attention mechanism, which reduces the computational complexity from O(n 2 ) is reduced to O(n), enabling it to process high-resolution DCE image data. Specifically, this method divides the time dimension into overlapping local windows, performs dense attention calculations within the windows, and then realizes cross-region feature interaction through window sliding.

[0007] However, directly migrating the visually optimized Swin Transformer to AIF representation still faces significant adaptation barriers: First, fixed-size window divisions produce artificial segmentation boundaries on the time axis, destroying the continuity characteristics of the hemodynamic curve; second, although local attention is used to reduce the computational load, GPU memory consumption still increases quadratically when processing high-temporal resolution DCE data above 256 frames. These limitations make it difficult for existing Transformer variants to balance computational efficiency and physiological feature fidelity in AIF detection tasks. Summary of the Invention

[0008] The purpose of the present invention is to propose a two-stage technical framework for the problem of automatic selection of arterial input function (AIF) in dynamic contrast-enhanced (DCE) imaging. The function is realized by combining a hybrid architecture (Dil-RWKV) based on dilated convolution and recursive weighted key-value (RWKV) network with a multi-dimensional scoring task to solve the defects of the existing technology, such as insufficient multi-scale feature modeling capability, low efficiency in capturing long-term temporal dependencies, and strong dependence on labeled data.

[0009] To achieve the purpose of the present invention, the present invention discloses a method for automatically selecting arterial input functions in DCE-MRI based on an attention mechanism, comprising the following steps:

[0010] Step 1: Perform fast Fourier transform on the data t to convert it into the frequency domain to obtain the corresponding amplitude m and phase p;

[0011] Step 2: The dual-domain data is input into the HDC Block residual convolutional network layer, which uses a layered dilated convolutional structure to capture multi-scale features and fuse shallow details with deep semantic features through residual connections;

[0012] Step 3: The data feature representation output by the HDC Block layer enters the RWK layer to implement feature temporal dependency modeling, and the parameter learning of the RWKV layer is completed in advance through a sequence reconstruction task;

[0013] Step 4: Use the classification head to classify all data into categories;

[0014] Step 5: Calculate the scores according to the category priority principle to obtain the final highest score AIF.

[0015] Furthermore, in step 1, a fast Fourier transform is performed on the given data t to convert it into the frequency domain to obtain the corresponding amplitude m and phase The final input data consists of three parts:

[0016]

[0017] Where b is the batch size and L is the time series length.

[0018] Furthermore, in step 2, the input data dimension is adjusted to (b,n,1,p) by performing patch embedding, and finally reshaped to (b×n,1,p), where b represents the batch size and P represents the patch size. represents the number of patches; then, the channel dimension is expanded to c through the convolution layer and average pooling is performed to generate feature representation Where B = b × n, d corresponds to the feature dimension; the processed features will be sent to the layered dilated convolution block for further processing.

[0019] Furthermore, the Hierarchical Dilated Convolution Block (HDC Block) enhances the representation of time series features through multi-scale perceptual fields, while simultaneously addressing the problems of long-range dependency modeling and gradient vanishing. This block employs an incremental hierarchical architecture, gradually fusing local details and global contextual information through three stages. Each stage contains two basic submodules, each consisting of dilated convolutional layers with different kernel sizes and dilation rates. Convolutions with a stride of 2 are used between adjacent submodules to downsample feature maps, gradually compressing spatial resolution and expanding the perceptual area. The first stage focuses on capturing local transient features; the second stage simulates mid-range fluctuations; and the third stage further captures global trends. The output is added to the original input via skip connections, forming a residual learning framework:

[0020] Out i =X+F i (X)

[0021] in, represents the composite convolution operation at stage i;

[0022] To handle variable-length feature sequences, an adaptive average pooling layer is applied, followed by a dimensionality resizing operation; the feature Out3 is compressed into a fixed-dimensional representation and the shape is converted to (B, C, 1) → (b, n, C), where C represents the output channel dimension and N is the predefined output length.

[0023] Furthermore, in step 3, after multi-scale feature learning of the original signal, a hierarchical embedding strategy is used to jointly encode the three modal data of timing, amplitude and phase. At the same time, in order to capture the relative position relationship in the sequence, the model uses a trainable position encoding matrix to inject hierarchical position information into the entire sequence.

[0024] Furthermore, the encoded spatiotemporal feature tensor of dimension b×n×C is fed into the RWKV Block, where long-range dependencies are learned through a gated recurrent mechanism and a state transfer model. To enhance the model's understanding of time-frequency domain representations, a cross-domain reconstruction task is used to pre-train the RWKV Block. A structured dropout strategy is employed to preserve the integrity of the time-frequency representation while maintaining temporal dynamic continuity.

[0025] The construction of the multimodal tensor is The model will then randomly discard some of the segments in the training with a probability of r. If the amplitude of the segment is discarded, the corresponding phase will also be discarded to ensure that the complete frequency information is ignored. Finally, the retained (1-r)×N slices will be used for reconstruction;

[0026]

[0027] The retained tensor slices are fed into a transformer-based encoder-decoder architecture, where the encoder extracts latent representations by aggregating cross-modal correlations, while the decoder reconstructs the original input by adaptively fusing temporal, spectral, and spatial features. To enhance robustness to partial data loss, a masked attention mechanism is employed to dynamically adjust feature weights based on the probability of patch retention.

[0028] The pre-training framework uses a hybrid loss function:

[0029] L=L MSE +L freq

[0030] Given input data x and reconstructed output x', the reconstruction loss is defined as the normalized mean square error NMSE:

[0031]

[0032] Where L represents the length of the time series, which represents the total number of time steps, and d represents the feature dimension. To enhance the consistency in the frequency domain, a spectrum constraint is introduced through the fast Fourier transform (FFT):

[0033]

[0034] in, represents the Fourier transform operator, ||·|| represents the complex amplitude spectrum, and k represents the frequency component in the spectral domain.

[0035] Furthermore, in step 4, for the AIF pre-screening task, the representations learned from the RWKV Block are first aggregated through global average pooling to compress the dimension b×C, and then input into a lightweight multi-layer perceptron (MLP) classifier with three output units corresponding to the clinical diagnosis categories; the MLP head consists of two hidden layers, each of which contains GeLU activation and BatchNorm regularization, and finally produces the output

[0036] The cross entropy loss calculation formula used is:

[0037]

[0038] Where N is the total number of samples, C is the number of categories, and y i,c represents the true label of sample i in category c, It represents the predicted probability that the i sample belongs to category c estimated by the model.

[0039] Furthermore, in step 5, to ensure physiological plausibility and robustness, a scoring system evaluating three key features was implemented in the second phase, incorporating key indicators: baseline stability S baseline , peak amplitude S peak and the stationarity of the downward trend S smoothness ;

[0040]

[0041] The final AIF selection is determined by the above formula, where w1, w2, and w3 represent weight parameters.

[0042] Compared with the existing technology, the significant progress of the present invention lies in: 1) the arterial input function generated by the model is highly close to the manually labeled results in terms of waveform morphology and timing characteristics, significantly improving the capture accuracy and stability of key hemodynamic parameters; 2) efficient and lightweight computing, greatly reducing the complexity of the model through architectural innovation, achieving clinical-level real-time processing efficiency while maintaining optimal performance, and providing feasibility for bedside instant diagnosis; 3) in terms of structural consistency, natural statistical characteristics and local distortion detection, the model output is highly consistent with the manually labeled results; 4) multi-parameter diagnostic verification shows that the model's evaluation performance on core hemodynamic indicators is highly consistent with clinical standards, and has clinical effectiveness and potential practicality.

[0043] In order to more clearly illustrate the functional characteristics and structural parameters of the present invention, further description is given below with reference to the accompanying drawings and specific embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of this application. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0045] Figure 1 This is a schematic diagram of the overall architecture of an automatic arterial input function selection method based on an attention mechanism in DCE-MRI;

[0046] Figure 2 This is a comparison chart of AIF selected by different methods;

[0047] Figure 3 It is a schematic diagram of the diagnostic effect of the multi-parameter hemodynamic assessment model. DETAILED DESCRIPTION

[0048] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments; based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0049] In response to the core defects of existing AIF detection methods, the present invention proposes a hybrid spatiotemporal architecture Dil-RWKV. This design systematically solves the bottleneck problems of previous research from three dimensions: First, a multi-scale feature extraction module is constructed by dilated causal convolution and a learnable exponential decay mechanism. While maintaining the sensitivity of CNN local features, a time domain dependency chain covering the complete blood flow cycle is established, breaking through the problem of temporal continuity fragmentation caused by the traditional Transformer fixed window division; secondly, the self-supervised time series reconstruction task and the semi-supervised contrastive learning feature decoupling module are innovatively integrated to form a dual-task joint optimization paradigm, so that the model can still maintain robustness in clinical scenarios with insufficient labeled data and variable scanning protocols; finally, a dynamic sequence length adaptation mechanism is developed to achieve seamless adaptation from multiple types of inputs through spatiotemporal feature decoupling and reconstruction technology. This architecture achieves accurate pre-screening of arterial input functions through a multi-dimensional feature fusion strategy, providing reliable decision support for subsequent scoring mechanisms. The specific implementation methods of this scheme are as follows:

[0050] Step 1: Perform fast Fourier transform on the data t, convert it to the frequency domain, and obtain the corresponding amplitude m and phase p

[0051] The present invention first performs a fast Fourier transform on the given data t, converts it into the frequency domain, and obtains the corresponding amplitude m and phase The final input data consists of three parts:

[0052]

[0053] Where b is the batch size and L is the time series length.

[0054] Step 2: The dual-domain data is input into the HDC Block residual convolutional network layer, which adopts a layered dilated convolution structure to capture multi-scale features and fuse shallow details with deep semantic features through residual connections.

[0055] First, the input data dimension is adjusted to (b,n,1,p) by performing patch embedding, and finally reshaped to (b×n,1,p), where b represents the batch size and P represents the patch size. Represents the number of patches. Subsequently, the channel dimension is expanded to c through the convolution layer and average pooling is performed to generate feature representation Here, B = b × n, and d corresponds to the feature dimension. The processed features are then fed into a layered dilated convolution block for further processing.

[0056] This paper proposes a Hierarchical Dilated Convolution Block (HDCBlock), which aims to enhance the representation of time series features through a multi-scale perceptual field while simultaneously addressing the problems of long-range dependency modeling and vanishing gradients. This block employs an incremental hierarchical architecture, gradually fusing local details with global contextual information over three stages.

[0057] Each stage consists of two basic submodules, each composed of dilated convolutional layers with varying kernel sizes and dilation rates. Convolutions with a stride of 2 between adjacent submodules downsample the feature maps, gradually compressing the spatial resolution and expanding the receptive area. The first stage focuses on capturing local transient features; the second stage simulates mid-range fluctuations; and the third stage further captures global trends. The output is appended to the original input via skip connections, forming a residual learning framework:

[0058] Out i =X+F i (X)

[0059] in, Represents the composite convolution operation at stage i. This design directly transfers shallow layer detail information to the deep network, effectively alleviating the gradient vanishing problem while promoting the complementary fusion of multi-scale features.

[0060] To handle variable-length feature sequences, an adaptive average pooling layer is applied, followed by a dimension resizing operation. This compresses the feature Out3 into a fixed-dimensional representation, transforming the shape to (B, C, 1) →

[0061] (b,n,C), where C represents the output channel dimension and N is the predefined output length. This normalization ensures robustness to changes in input length, thereby promoting stable performance of downstream tasks.

[0062] Step 3: The data feature representation output by the HDC Block layer enters the RWK layer to implement feature temporal dependency modeling, and the parameter learning of the RWKV layer is completed in advance through a sequence reconstruction task.

[0063] After learning multi-scale features of the original signal, a hierarchical embedding strategy is used to jointly encode the three modal data of timing, amplitude and phase. At the same time, in order to capture the relative position relationship in the sequence, the model uses a trainable position encoding matrix to inject hierarchical position information into the entire sequence.

[0064] The encoded spatiotemporal feature tensor (dimension b×n×C) is input into the RWKV Block, and long-range dependency learning is achieved through a gated recurrent mechanism and a state transfer model. To enhance the model's understanding of time-frequency domain representations, we use a cross-domain reconstruction task to pre-train the RWKV Block. In the self-supervised learning paradigm, the practice of masking part of the data for reconstruction is widely adopted. Traditional methods usually mask time segments by clearing specific areas to zero, which distorts the original signal pattern and introduces spectral leakage artifacts. To address this limitation, we adopt a structured dropping strategy that preserves the integrity of the time-frequency representation while maintaining temporal dynamic continuity.

[0065] The construction of the multimodal tensor is The model will then randomly discard some of the segments in training with a probability of r. If the amplitude of the segment is discarded, the corresponding phase will also be discarded to ensure that the complete frequency information is ignored. Finally, the retained (1-r)×N slices will be used for reconstruction.

[0066]

[0067] The reconstruction task aims to promote comprehensive feature learning through self-supervised training. The retained tensor slices are fed into a Transformer-based encoder-decoder architecture, where the encoder extracts latent representations by aggregating cross-modal correlations, while the decoder reconstructs the original input by adaptively fusing temporal, spectral, and spatial features. To enhance robustness to partial data loss, a masked attention mechanism is employed that dynamically adjusts feature weights based on the probability of patch retention.

[0068] The pre-training framework uses a hybrid loss function:

[0069] L=L MSE +L freq

[0070] Given input data x and reconstructed output x', the reconstruction loss is defined as the normalized mean squared error (NMSE):

[0071]

[0072] Where L represents the length of the time series, which represents the total number of time steps, and d represents the feature dimension. To enhance the consistency in the frequency domain, a spectrum constraint is introduced through the fast Fourier transform (FFT):

[0073]

[0074] in, represents the Fourier transform operator, ||·|| represents the complex amplitude spectrum, and k represents the frequency component in the spectral domain.

[0075] Step 4: Use the classification head to classify all data into categories.

[0076] For the AIF pre-screening task, the representations learned from the RWKV Block are first aggregated through global average pooling to compress the dimension b×C, and then input into a lightweight multi-layer perceptron (MLP) classifier with three output units corresponding to the clinical diagnosis categories. The MLP head consists of two hidden layers, each of which contains GeLU activation and BatchNorm regularization, and finally produces the output

[0077] The cross entropy loss calculation formula used is:

[0078]

[0079] Where N is the total number of samples, C is the number of categories, and y i,c represents the true label of sample i in category c, represents the predicted probability estimated by the model that sample i belongs to category c. This representation effectively measures the difference between the predicted probability distribution and the true distribution across all samples and categories.

[0080] Step 5: Calculate the score based on the category priority principle to obtain the final highest score AIF

[0081] To ensure physiological plausibility and robustness, we implemented a scoring system that assessed three key features in the second phase, incorporating key indicators: baseline stability S baseline , peak amplitude S peak and the stationarity of the downward trend S smoothness The final AIF selection is determined by the following factors:

[0082]

[0083] Where w1, w2, and w3 represent weight parameters.

[0084] Example

[0085] This paper proposes a hybrid architecture (Dil-RWKV) based on dilated convolution and recursive weighted key value (RWKV) network, combined with a multi-dimensional scoring mechanism to achieve accurate, efficient and clinically practical automatic selection of AIF. The overall architecture is as follows: Figure 1 As shown. Multi-dimensional optimization is achieved on the core hemodynamic indicators, maintaining a high degree of consistency with the manual reference standard, and significantly enhancing the ability to capture dynamic details of blood flow. More importantly, this method maintains the algorithm's strong robustness to differences in scanning parameters while keeping the number of parameters at only 45.9% of Swin Transformer, effectively reducing the clinical uncertainty of perfusion parameter interpretation. The Dil-RWKV hybrid architecture proposed in this invention, combined with the scoring mechanism, has the following significant advantages in the automatic selection of arterial input function (AIF):

[0086] (1) High consistency between waveform similarity and hemodynamic characteristics

[0087] Table 1 Quantitative index results

[0088] method Normalized cross-correlation Peak Peak time Full Width Half Height Parameter quantity Manual reference - 119.08±27.31 35.83±6.46 34.02±38.53 - K-Means 0.795±0.138 99.09±32.05 37.02±10.02 55.14±61.85 - CNN 0.852±0.120 92.23±9.54 38.13±8.29 73.32±38.52 52,151 Swin-T 0.836±0.101 92.78±10.67 36.34±6.77 37.51±36.53 1,318,186 Dil-RWKV 0.867±0.033 93.58±3.31 36.94±6.36 34.62±25.23 605,314

[0089] Table 1 shows the quantitative indicator results, including normalized cross correlation (NCC), peak height (PH), time to peak (TTP), full width at half maximum (FWHM) and total trainable parameters (Parameters).

[0090] As shown in Table 1, the AIF generated by Dil-RWKV is closest to the manually annotated results in terms of waveform similarity (NCC = 0.867 ± 0.033), and its NCC stability is 72.5% higher than that of the second-best method (CNN). In terms of morphological parameters, the absolute error of peak height (21.5%) is significantly lower than that of K-Means and Swin-T methods, the error of peak time is 3.2% lower than that of CNN, and the coefficient of variation of half-width (FWHM) estimation is only 13.5% of that of K-Means method. A comparison of AIFs selected by different methods (all manually selected AIFs are completed by medical imaging experts with many years of clinical perfusion analysis experience) is shown in the figure. Figure 2 shown.

[0091] (2) Computational efficiency and model lightweight

[0092] Compared with the Swin-Transformer architecture, the number of parameters of Dil-RWKV is reduced by 54.1% (605,314 vs 1,318,186) while maintaining the optimal comprehensive performance indicators, significantly improving the feasibility of the algorithm in clinical real-time applications.

[0093] (3) Reliability verification of hemodynamic parameter maps

[0094] Table 2 Consistency of AIF selection methods on CBV parameter map quality

[0095] Image quality assessment metrics Intraclass correlation coefficient F-number P-value 95% confidence interval consistency BRISQUE 0.9304 14.372 1.38E-12 [0.860.96] - NIQE 0.8608 7.186 3.21E-08 [0.730.93] - PIQE - - - - 91.67％

[0096] The intraclass correlation coefficient (ICC) and consistency analysis (Table 2) confirmed that:

[0097] BRISQUE evaluation ICC = 0.930 (95% CI: 0.86-0.96)

[0098] NIQE evaluation ICC = 0.861 (95% CI: 0.73-0.93)

[0099] PIQE percentage consistency reached 91.67%

[0100] The above results exceeded the clinical reliability threshold (ICC>0.75) and were statistically significant (p<1×10 -7 ), indicating that the automatic selection results are highly consistent with manual annotations in terms of structural similarity, natural scene statistical characteristics, and sensitivity to local distortion.

[0101] (4) Quantitative analysis and evaluation of the diagnostic effect of the multi-parameter ROC curve model

[0102] The CBV parameter map quality consistency results are as follows Figure 3 As shown in Figure 3, we performed a comprehensive ROC analysis to evaluate the diagnostic performance of our proposed Dil-RWKV framework (b) compared with the traditional manual reference standard (a) on multiple hemodynamic parameters. The robust agreement observed across all evaluated parameters further emphasizes the clinical validity and potential utility of our approach.

[0103] In summary, the core of the present invention is to construct a hybrid architecture that integrates multi-scale feature extraction and long-term temporal dependency modeling. Through the synergistic effect of the time-frequency domain joint modeling architecture of the layered dilated convolution module and the RWKV network, accurate feature learning of the arterial input function is achieved. It effectively solves the problem of feature truncation caused by the limited receptive field of the traditional CNN model and realizes the modeling of long-term temporal dependencies under O(N) computational complexity. In addition, the amplitude spectrum and phase spectrum are generated as frequency domain representations through fast Fourier transform to enhance the modeling ability. At the same time, the dimension normalization of the variable-length feature sequence is achieved through the adaptive average pooling layer, so that the model can adapt to the time resolution differences under different scanning protocols. The synergistic effect of the above-mentioned technological breakthroughs enables the present invention to achieve or even exceed the accuracy level of the supervised learning method while maintaining the data efficiency advantage of the unsupervised method. In particular, through the combination of frequency domain feature enhancement and self-supervised pre-training, the model's dependence on labeled data is reduced, the feature learning ability of the data is significantly improved, and accurate and efficient pre-screening and classification tasks are achieved. At the same time, a multi-dimensional scoring system was defined. Based on the core requirements of clinical practice for an ideal AIF, a quantitative scoring mechanism including a three-stage physiological rationality verification was constructed, and the optimal AIF was finally dynamically selected through a weighted scoring function.

[0104] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus.

[0105] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A method for automatic selection of arterial input functions in DCE-MRI based on attention mechanism, characterized by: The following steps are involved: Step 1: Perform fast Fourier transform on the data t to convert it into the frequency domain to obtain the corresponding amplitude m and phase p; Step 2: The dual-domain data is input into the HDC Block residual convolutional network layer, which uses a layered dilated convolutional structure to capture multi-scale features and fuse shallow details with deep semantic features through residual connections; Step 3: The data feature representation output by the HDC Block layer enters the RWK layer to implement feature temporal dependency modeling, and the parameter learning of the RWKV layer is completed in advance through a sequence reconstruction task; Step 4: Use the classification head to classify all data into categories; Step 5: Calculate the scores according to the category priority principle to obtain the final highest score AIF.

2. The method for automatically selecting arterial input functions based on an attention mechanism in DCE-MRI according to claim 1, characterized in that: In step 1, a fast Fourier transform is performed on the given data t to convert it into the frequency domain to obtain the corresponding amplitude m and phase The final input data consists of three parts: Where b is the batch size and L is the time series length.

3. The method for automatically selecting arterial input functions based on an attention mechanism in DCE-MRI according to claim 1, characterized in that: In step 2, the input data dimension is adjusted to (b,n,1,p) by performing patch embedding, and finally reshaped to (b×n,1,p), where b represents the batch size and P represents the patch size. represents the number of patches; then, the channel dimension is expanded to c through the convolution layer and average pooling is performed to generate feature representation Where B = b × n, d corresponds to the feature dimension; the processed features will be sent to the layered dilated convolution block for further processing.

4. The method for automatically selecting arterial input functions based on an attention mechanism in DCE-MRI according to claim 3, characterized in that: The Hierarchical Dilated Convolution Block (HDC Block) enhances the representation of time series features through multi-scale perceptual fields, while simultaneously addressing the problems of long-range dependency modeling and vanishing gradients. This block employs an incremental hierarchical architecture that gradually fuses local details and global contextual information through three stages. Each stage contains two basic submodules, each consisting of dilated convolutional layers with different kernel sizes and dilation rates. Convolutions with a stride of 2 are used between adjacent submodules to downsample feature maps, gradually compressing spatial resolution and expanding the perceptual area. The first stage focuses on capturing local transient features. The second stage simulates mid-range fluctuations; the third stage further captures global trends. The output is added to the original input via skip connections, forming a residual learning framework: Out i =X+F i (X) in, represents the composite convolution operation at stage i; To handle variable-length feature sequences, an adaptive average pooling layer is applied, followed by a dimensionality resizing operation; the feature Out3 is compressed into a fixed-dimensional representation and the shape is converted to (B, C, 1) → (b, n, C), where C represents the output channel dimension and N is the predefined output length.

5. The method for automatically selecting arterial input functions based on an attention mechanism in DCE-MRI according to claim 1, characterized in that: In step 3, after multi-scale feature learning of the original signal, a hierarchical embedding strategy is used to jointly encode the three modal data of timing, amplitude and phase. At the same time, in order to capture the relative position relationship in the sequence, the model uses a trainable position encoding matrix to inject hierarchical position information into the entire sequence.

6. The method for automatically selecting arterial input functions based on an attention mechanism in DCE-MRI according to claim 5, characterized in that: The encoded spatiotemporal feature tensor with dimensions of b×n×C is input into the RWKV Block, and long-range dependency learning is achieved through the gated loop mechanism and state transfer model; To enhance the model's understanding of time-frequency domain representation, a cross-domain reconstruction task is used to pre-train the RWKV Block; A structured discarding strategy is used to preserve the integrity of time-frequency representation while maintaining temporal dynamic continuity; The construction of the multimodal tensor is The model will then randomly discard some of the segments in the training with a probability of r. If the amplitude of the segment is discarded, the corresponding phase will also be discarded to ensure that the complete frequency information is ignored. Finally, the retained (1-r)×N slices will be used for reconstruction; The retained tensor slices are fed into a transformer-based encoder-decoder architecture, where the encoder extracts latent representations by aggregating cross-modal correlations, while the decoder reconstructs the original input by adaptively fusing temporal, spectral, and spatial features. To enhance robustness to partial data loss, a masked attention mechanism is employed to dynamically adjust feature weights based on the probability of patch retention. The pre-training framework uses a hybrid loss function: L=L MSE +L freq Given input data x and reconstructed output x', the reconstruction loss is defined as the normalized mean square error NMSE: Where L represents the length of the time series, which represents the total number of time steps, and d represents the feature dimension. To enhance the consistency in the frequency domain, a spectrum constraint is introduced through the fast Fourier transform (FFT): in, represents the Fourier transform operator, ||·|| represents the complex amplitude spectrum, and k represents the frequency component in the spectral domain.

7. The method for automatically selecting arterial input functions based on an attention mechanism in DCE-MRI according to claim 1, characterized in that: In step 4, for the AIF pre-screening task, the representations learned from the RWKV Block are first aggregated through global average pooling to compress the dimension b×C, and then input into a lightweight multi-layer perceptron (MLP) classifier with three output units corresponding to the clinical diagnosis categories; The MLP head consists of two hidden layers, each of which contains GeLU activation and BatchNorm regularization, and finally produces the output The cross entropy loss calculation formula used is: Where N is the total number of samples, C is the number of categories, and y i,c represents the true label of sample i in category c, It represents the predicted probability that the i sample belongs to category c estimated by the model.

8. The method for automatically selecting arterial input functions based on an attention mechanism in DCE-MRI according to claim 1, characterized in that: In step 5, to ensure physiological plausibility and robustness, a scoring system evaluating three key features was implemented in the second phase, incorporating key indicators: baseline stability S baseline , peak amplitude S peak and the stationarity of the downward trend S smoothness ; The final AIF selection is determined by the above formula, where w1, w2, and w3 represent weight parameters.

Images