Cerebral blood flow diffusion assessment method based on attention guidance sparse image fusion

Abstract

The invention discloses a cerebral blood flow diffusion assessment method based on attention guidance sparse image fusion, and belongs to the technical field of medical image processing, and the method comprises the steps: collecting an original data set of a CTA image sequence of an ischemic smoke disease, constructing a plurality of groups of sparse image pairs each comprising a diffusion image and a blood vessel image, and obtaining an image training set; constructing a sparse image fusion and visualization model; training a sparse image fusion and visualization model by using the image training set to obtain a trained sparse image fusion and visualization model; and processing a to-be-detected ischemic smoke disease CTA image by using the trained sparse image fusion and visualization model to obtain a corresponding reconstructed fusion image, and completing evaluation based on a pseudo-color image of the corresponding reconstructed fusion image and a visualization result. The problems that an existing method still has limitation in the aspect of sparse image fusion precision, cross-modal feature extraction and fusion efficiency is insufficient, and the accuracy of an evaluation result is limited are solved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of medical image processing, and particularly relates to a cerebral blood flow diffusion evaluation method based on attention-guided sparse image fusion. BACKGROUND

[0002] Cerebral vascular disease is an important disease type that seriously threatens human health, among which ischemic moyamoya disease (MMD) is particularly typical. The intracranial large arteries of the disease gradually narrow or even occlude, forming an abnormal "smoke-like" vascular network, leading to insufficient blood supply to the brain. For patients, vascular reconstruction surgery is the main treatment method, and accurate evaluation of the blood flow diffusion state of the blood vessels and the surrounding area before surgery is of key significance for surgical planning and postoperative recovery.

[0003] To obtain more comprehensive image information, image fusion technology has been gradually applied to multi-source image integration tasks. Traditional methods (such as multi-scale transformation and sparse representation) and deep learning methods (such as convolutional neural networks, U-Net and its variants) have made certain progress in conventional image fusion, which can improve image clarity and information integrity to a certain extent. However, these methods are mostly designed for conventional images and are difficult to fully adapt to the characteristics of sparse images, especially in areas with small structures and weak signals.

[0004] In the specific scenario of sparse cerebral vascular image fusion, current research is still very limited. The computed tomography angiography imaging (CTA) derived vascular image (VI) and diffusion image (DI) of MMD patients usually contain a large number of sparse regions and low-contrast vascular structures, and these subtle blood vessels and diffusion information are crucial for preoperative evaluation. The existing methods lack optimization design for sparse features, which can easily lead to loss of key structural information and noise interference, thereby reducing the fusion accuracy and result interpretability.

[0005] In summary, the existing technology still has certain limitations in sparse image fusion accuracy, especially in preserving sparse region information, improving structural detail recognition ability, and enhancing the reliability of the fusion results.

[0006] Therefore, it is necessary to propose a new fusion strategy to achieve detail preservation, structure enhancement and high-quality reconstruction under sparse conditions, and to provide more scientific and reliable technical support for cerebral blood flow dynamics evaluation and related applications. SUMMARY

[0007] In view of the above-mentioned deficiencies in the prior art, the cerebral blood flow diffusion evaluation method based on attention-guided sparse image fusion provided by the present application solves the problems of limitations in sparse image fusion accuracy, insufficient cross-modal feature extraction and fusion efficiency, and limited accuracy of evaluation results of the existing methods.

[0008] To achieve the aforementioned objectives, the present invention employs the following technical solution: a method for assessing cerebral blood flow diffusion based on attention-guided sparse image fusion, comprising: Collect raw dataset of CTA image sequences for ischemic moyamoya disease; Based on the original CTA image sequence dataset, several sparse image pairs, each including diffuse and vascular images, were constructed, and the sparse image pairs were integrated to obtain the image training set. A sparse image fusion and visualization model is constructed. The sparse image fusion and visualization model includes a ConvNext encoder, a bidirectional attention fusion module, a discriminator, and a decoder, which are respectively connected to the diffusion image input end and the blood vessel image input end. A pseudo-color image construction module and an interactive three-dimensional visualization module are respectively connected to the decoder. The bidirectional attention fusion module is respectively connected to the ConvNext encoder, the discriminator, and the decoder. A sparse image fusion and visualization model is trained using an image training set to obtain the trained sparse image fusion and visualization model. The trained sparse image fusion and visualization model was used to process the CTA images of ischemic moyamoya disease to obtain the corresponding reconstructed fused images. The evaluation was then completed based on the pseudo-color images and visualization results of the corresponding reconstructed fused images.

[0009] The beneficial effects of this invention are as follows: This invention constructs the aforementioned "dual-path collaborative optimization" architecture. Unlike traditional single-output fusion networks, this invention unifies the two objectives of "pixel-level fidelity optimization" and "feature-level semantic authenticity optimization" within a single framework by introducing a multi-task reconstruction path for the decoder and a feature discrimination path for the discriminator. Each module forms a closed-loop feedback system through a unique collaborative mechanism. The bidirectional attention fusion module provides high-quality fusion features for the dual paths, while the reconstruction loss and adversarial loss generated by the dual paths propagate in the backpropagation, jointly guiding the bidirectional attention fusion module and the encoder to learn how to generate more complementary and discriminative features. The final fused image is displayed through pseudo-colorization and interactive visualization. This achieves accurate capture and efficient fusion of vascular and blood flow features, generating structurally complete and semantically consistent fused images. Combined with 3D visualization and pseudo-color mapping, it improves the accuracy and reliability of blood flow diffusion assessment, thereby providing intuitive and reliable auxiliary decision support for preoperative planning of MMD.

[0010] Furthermore, the construction of several sets of sparse image pairs, each including a diffuse image and a blood vessel image, specifically involves: By excluding sequence data in the original CTA image sequence dataset that have a consistency lower than the consistency threshold, a high-quality image dataset is obtained. For each high-quality image data in the high-quality image dataset, blood flow diffusion features and vascular structure information are extracted to obtain blood flow diffusion feature images and vascular structure information images. The blood flow diffusion feature image and the vascular structure information image were standardized to obtain diffusion image and vascular image, respectively. By integrating the diffuse image and vascular image corresponding to the same high-quality image data into a set of sparse image pairs, several sets of sparse image pairs, each including diffuse image and vascular image, are obtained.

[0011] The beneficial effects of the above-mentioned further scheme are as follows: by filtering out sequence data with "low consistency", invalid or interfering samples are removed from the source, avoiding such low-quality data from misleading the model learning, ensuring that the "high-quality image dataset" used for feature extraction has a stable anatomical structure and feature expression, and laying the foundation for the model to learn real and reliable cross-modal associations.

[0012] Furthermore, the ConvNext encoder is used to extract image features from the diffusion image and the blood vessel image respectively, to obtain a diffusion feature map and a blood vessel feature map; The bidirectional attention fusion module is used to fuse the diffusion feature map and the vascular feature map to obtain a fused feature representation; The discriminator is used to generate classification probabilities based on the diffusion feature map, the vascular feature map, and the fusion feature representation; The decoder is used to reconstruct the diffusion image, the vascular image, and the fused image based on the diffusion feature map, the vascular feature map, and the fusion feature representation, respectively. The pseudo-color image construction module is used to generate a pseudo-color image based on the reconstructed fused image; The interactive 3D visualization module is used to visualize the reconstructed fused image.

[0013] The beneficial effects of the above-mentioned further scheme are as follows: the ConvNext encoder extracts features from diffusion and vascular images in a targeted manner, fully preserving key information of each modality; the bidirectional attention fusion module can focus on complementary information, achieve information enhancement, and generate more complete fusion features; the discriminator and decoder work together to supervise and ensure the accuracy of feature extraction and fusion, avoiding information distortion; finally, through pseudo-color images and interactive 3D visualization, the spatial relationship between blood vessels and diffusion information is presented, significantly improving the accuracy and clinical practical value of medical image processing, and providing strong support for diagnosis and treatment planning.

[0014] Furthermore, the bidirectional attention fusion module includes: The first linear layer is used to linearly project the diffusion feature map into the value vector and key vector corresponding to the diffusion feature; The second linear layer is used to linearly project the blood vessel feature map into query vectors, value vectors, and key vectors corresponding to the blood vessel features; The third linear layer is used to linearly project the diffusion feature map into a query vector corresponding to the diffusion feature. The first matrix multiplication processing unit is used to perform matrix multiplication on the key vector corresponding to the diffusion feature and the query vector corresponding to the blood vessel feature to obtain the first multiplication result; The first normalization processing unit is used to normalize the first multiplication result to obtain the normalized result of the first multiplication. The first random deactivation processing unit is used to randomly deactivate the normalized result of the first multiplication to obtain the first random deactivation result; The second matrix multiplication processing unit is used to perform matrix multiplication on the key vector corresponding to the blood vessel feature and the query vector corresponding to the diffusion feature to obtain the second multiplication result; The second normalization processing unit is used to normalize the result of the second multiplication to obtain the normalized result of the second multiplication. The second random deactivation processing unit is used to randomly deactivate the normalized result of the second multiplication to obtain the second random deactivation result; The third matrix multiplication processing unit is used to multiply the value vector corresponding to the diffusion feature and the first random inactivation result to obtain the third multiplication result; The fourth linear layer is used to perform dimensionality calibration and information integration on the third multiplication result to obtain the first linearized result; The third random deactivation processing unit is used to randomly deactivate the first linearization result to obtain the third random deactivation result. The fourth matrix multiplication processing unit is used to multiply the value vector corresponding to the blood vessel features and the second random inactivation result to obtain the fourth multiplication result; The fifth linear layer is used to perform dimensionality calibration and information integration on the fourth multiplication result to obtain the second linearization result; The fourth random deactivation processing unit is used to randomly deactivate the second linearization result to obtain the fourth random deactivation result; The weighted summation processing unit is used to perform weighted summation on the third and fourth random inactivation results to obtain the fused feature representation.

[0015] The beneficial effects of the above-mentioned further scheme are as follows: by accurately constructing bimodal query, key, and value vectors through multilinear layers, a foundation for bidirectional interaction is laid; by using bidirectional matrix multiplication, normalization, and random deactivation, bidirectional modal correlation information is captured, avoiding information omissions in unidirectional attention; by combining weights and linearizing processing, effective features are strengthened and overfitting is prevented, and finally, the fused features are obtained by weighted summation, which not only preserves modal complementarity but also enhances feature correlation, thereby improving the integrity and discriminative power of the fused features.

[0016] Furthermore, the discriminator includes: The feature input end is used to obtain diffusion feature maps, blood vessel feature maps, and fusion feature representations to obtain feature sequences; The dimension rearrangement unit is used to rearrange the feature sequence according to the adapted dimension of the global average pooling unit to obtain a dimension-adapted feature sequence. The global average pooling unit is used to aggregate features from dimension-adapted feature sequences using global average pooling operations to obtain global feature descriptors. The flattening operation unit is used to flatten the global feature descriptor to obtain a compact one-dimensional feature vector; The discriminant unit is used to extract features from a compact one-dimensional feature vector using three cascaded discriminant structures to obtain discriminant features; each discriminant structure includes a fully connected layer, layer normalization, activation function, and random deactivation connected in sequence. The first fully connected layer is used to map discriminative features to classification probabilities.

[0017] The beneficial effects of the above-mentioned further scheme are as follows: by integrating multimodal features (diffusion, vascular and fusion features) at the feature input end, comprehensive discrimination of multi-source information is achieved; dimensionality rearrangement and global average pooling operations adapt feature dimensions and aggregate global information, avoiding local feature bias; flattening operation transforms global features into compact vectors, improving computational efficiency; the three-tiered discriminative structure, through a combination of fully connected layers, layer normalization, activation functions and random deactivation, refines features layer by layer and suppresses overfitting, and finally outputs classification probabilities through the first fully connected layer, which can effectively supervise the effectiveness of feature extraction and fusion, enhance the discrimination ability of lesion features, and provide reliable guidance for model optimization.

[0018] Furthermore, the loss function of the sparse image fusion and visualization model includes classification loss and reconstruction loss:

[0019] in, For sparse image fusion and visualization models; For classification loss; To rebuild the losses.

[0020] The beneficial effects of the above-mentioned further scheme are as follows: By designing the classification loss, the model can be ensured to accurately distinguish different modal features and the effectiveness of fusion features, thereby improving the ability to classify and identify key information such as lesion areas; by using the reconstruction loss to constrain the image generation process of the decoder, the model can be forced to retain the key details of the original image when reconstructing diffusion, blood vessel and fusion images, thus avoiding information distortion; the two losses work together to optimize the model from both the "discrimination accuracy" and "image fidelity" dimensions, effectively balancing feature discrimination performance and image reconstruction quality, and ultimately improving the overall stability and output reliability of the sparse image fusion and visualization model.

[0021] Furthermore, the expression for the classification loss is:

[0022] in, The binary cross-entropy classification loss is used for the diffusion feature map; For binary cross-entropy loss; The classification result obtained by the discriminator based on the diffusion feature map; The binary cross-entropy classification loss is used for the vascular feature map. The classification result obtained by the discriminator based on the vascular feature map; The binary cross-entropy classification loss is used to fuse feature representations; The classification result is obtained by the discriminator based on the fused feature representation.

[0023] The beneficial effects of the above-mentioned further scheme are as follows: by providing refined supervision of the multimodal feature discrimination process, the problem of single classification loss ignoring the discrimination bias of a certain type of feature is avoided; by summing the classification losses of the three types of features, the discrimination accuracy of each modal feature and fused feature can be balanced, ensuring that the model pays attention to both the effectiveness of single modal features and the quality of fused features; the binary cross-entropy loss is sensitive to the probability difference of classification results, which can punish the misjudgment behavior of the discriminator, help the model converge to a better discrimination state quickly, and provide more accurate supervision basis for subsequent feature fusion and image reconstruction.

[0024] Furthermore, the expression for the reconstruction loss is:

[0025] in, These are the weighting coefficients for the loss in the diffuse image reconstruction. For the loss of image reconstruction due to diffusion; These are the weighting coefficients for the blood vessel image reconstruction loss; Loss due to vascular image reconstruction; The weighting coefficient for penalizing losses; To punish the loss; This is the original diffusion image; For the reconstructed diffusion image; This is the original image of the blood vessels; Reconstructed vascular images; It is an L1 norm; For indicator functions, when =0 and If the value is 0, the value is 1; otherwise, the value is 0. The first of the original diffused image The pixel value of each pixel; The first of the original blood vessel images The pixel value of each pixel; The first fused image for reconstruction The pixel value of each pixel; For pixel index; It is an absolute value.

[0026] The beneficial effects of the above-mentioned further solutions are as follows: By setting weight coefficients for the reconstruction losses of diffusion and vascular images respectively, the constraint intensity can be flexibly adjusted according to the importance of the two modalities in clinical scenarios, avoiding the problem of insufficient reconstruction accuracy of key modal information due to a single weight; calculating the core reconstruction loss using the L1 norm can effectively reduce the interference of outliers on loss calculation, more accurately measure the pixel-level differences between the original and reconstructed images, and ensure the fidelity of image details; the additional introduction of a weighted penalty loss, through the indicator function to specifically constrain the case where "the original dual-modal pixels are all 0 but the fused image reconstruction is non-zero", can reduce the generation of meaningless artifacts and avoid false features interfering with subsequent diagnosis; the overall expression coordinates constraints from three aspects: "modal weight adaptation", "accurate measurement of pixel differences" and "targeted suppression of artifacts", comprehensively improving the reconstruction quality of diffusion, vascular and fused images, and ensuring that the output image meets the requirements of clinical diagnosis for accuracy and reliability.

[0027] Furthermore, the trained sparse image fusion and visualization model no longer reconstructs diffuse and vascular images when applied.

[0028] The beneficial effects of the above-mentioned further scheme are as follows: Each module forms a closed-loop feedback system through a unique collaborative mechanism. The bidirectional attention fusion module provides high-quality fusion features for both paths, while the reconstruction loss and adversarial loss generated by the two paths propagate in the back, jointly guiding the bidirectional attention fusion module and the encoder to learn how to generate more complementary and discriminative features. The final fused image is displayed through pseudo-colorization and interactive visualization. This achieves accurate capture and efficient fusion of vascular and blood flow features, generating structurally complete and semantically consistent fused images. Combined with 3D visualization and pseudo-color mapping, it improves the accuracy and reliability of blood flow diffusion assessment, thereby providing intuitive and reliable auxiliary decision support for preoperative planning of MMD. Attached Figure Description

[0029] Figure 1 This is a flowchart of the method of the present invention.

[0030] Figure 2 This is a schematic diagram of the original image and the preprocessed model input image pair in an embodiment of the present invention.

[0031] Figure 3 This is a structural diagram of the model of the present invention.

[0032] Figure 4 This is a structural diagram of the bidirectional attention fusion module of the present invention.

[0033] Figure 5 This is a structural diagram of the discriminator of the present invention.

[0034] Figure 6 This is a schematic diagram of the input image and the fused image in an embodiment of the present invention.

[0035] Figure 7 This is a diagram illustrating the pseudo-color mapping and interactive visualization effects in an embodiment of the present invention. Detailed Implementation

[0036] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0037] like Figure 1 As shown, in one embodiment of the present invention, a method for assessing cerebral blood flow diffusion based on attention-guided sparse image fusion includes: Collect raw dataset of CTA image sequences for ischemic moyamoya disease; Based on the original CTA image sequence dataset, several sparse image pairs, each including diffuse and vascular images, were constructed, and the sparse image pairs were integrated to obtain the image training set. A sparse image fusion and visualization model is constructed. The sparse image fusion and visualization model includes a ConvNext encoder, a bidirectional attention fusion module, a discriminator, and a decoder, which are respectively connected to the diffusion image input end and the blood vessel image input end. A pseudo-color image construction module and an interactive three-dimensional visualization module are respectively connected to the decoder. The bidirectional attention fusion module is respectively connected to the ConvNext encoder, the discriminator, and the decoder. A sparse image fusion and visualization model is trained using an image training set to obtain the trained sparse image fusion and visualization model. The trained sparse image fusion and visualization model was used to process the CTA images of ischemic moyamoya disease to obtain the corresponding reconstructed fused images. The evaluation was then completed based on the pseudo-color images and visualization results of the corresponding reconstructed fused images.

[0038] In this embodiment, the present invention aims to address two major technical challenges in assessing blood flow diffusion based on CTA images during preoperative planning for ischemic MMD: First, the sparse regions and zero-value distribution in vascular images (VI) and diffusion images (DI) result in insufficient efficiency in cross-modal feature extraction and fusion, leading to limited accuracy in assessment results; second, existing deep fusion methods are prone to feature loss when processing sparse images, resulting in insufficient structural integrity and blood flow information representation capabilities in the generated images, making it difficult to meet the clinical demand for accurate assessment. Therefore, this invention achieves accurate capture and efficient fusion of vascular and blood flow features through network architecture design and joint loss function construction, generating structurally complete and semantically consistent fused images. Combined with 3D visualization and pseudo-color mapping, the accuracy and reliability of blood flow diffusion assessment are improved, thereby providing intuitive and reliable auxiliary decision support for preoperative planning of MMD.

[0039] The construction of several sparse image pairs, each including diffuse and vascular images, is specifically as follows: By excluding sequence data in the original CTA image sequence dataset that have a consistency lower than the consistency threshold, a high-quality image dataset is obtained. For each high-quality image data in the high-quality image dataset, blood flow diffusion features and vascular structure information are extracted to obtain blood flow diffusion feature images and vascular structure information images. The blood flow diffusion feature image and the vascular structure information image were standardized to obtain diffusion image and vascular image, respectively. By integrating the diffuse image and vascular image corresponding to the same high-quality image data into a set of sparse image pairs, several sets of sparse image pairs, each including diffuse image and vascular image, are obtained.

[0040] In this embodiment, CTA image acquisition: This step is the data preparation stage. Raw CTA image sequences from ischemic MMD patients, containing multiple time series, are acquired from collaborating hospitals to provide a foundation for subsequent processing.

[0041] Image data screening: To ensure the quality of input data, patient data that conforms to the characteristics of ischemic MMD are first retained. Then, through mask construction and quantitative analysis (plotting line graphs of computed tomography values, calculating slope and dispersion), sequences with poor consistency are excluded, and finally high-quality image data are obtained for subsequent analysis.

[0042] Image preprocessing: The core of image preprocessing lies in extracting blood flow diffusion features and vascular structure information from the selected CTA sequences, providing high-quality input for subsequent sparse image construction and fusion. After preprocessing the selected images, 4160 DI-VI image pairs were obtained. These image pairs were divided into a training set of 3200 pairs and a validation set of 960 pairs. The original images and the preprocessed DI and VI pairs are shown below. Figure 2 As shown.

[0043] Sparse Image Construction: This step prepares for fusion. The DI and VI obtained after decoupling are normalized to construct regular sparse image pairs.

[0044] like Figure 3 As shown, the ConvNext encoder is used to extract image features from the diffusion image and the blood vessel image respectively, to obtain the diffusion feature map and the blood vessel feature map; The bidirectional attention fusion module is used to fuse the diffusion feature map and the vascular feature map to obtain a fused feature representation; The discriminator is used to generate classification probabilities based on the diffusion feature map, the vascular feature map, and the fusion feature representation; The decoder is used to reconstruct the diffusion image, the vascular image, and the fused image based on the diffusion feature map, the vascular feature map, and the fusion feature representation, respectively. The pseudo-color image construction module is used to generate a pseudo-color image based on the reconstructed fused image; The interactive 3D visualization module is used to visualize the reconstructed fused image.

[0045] In this embodiment, the sparse image fusion and visualization model aims to address the challenges of insufficient feature extraction accuracy and optimization in multimodal sparse image fusion. The overall architecture of the sparse image fusion and visualization model is as follows: Figure 3 As shown, the proposed model mainly consists of a ConvNext encoder, a bidirectional attention fusion module, a decoder, and a discriminator. The data processing flow of this architecture is as follows: The model takes DI and VI as inputs and first extracts features through a shared lightweight ConvNext encoder. These features are then input to the bidirectional attention fusion module, which achieves key feature interaction and fusion between different modalities through a bidirectional cross-modal attention mechanism. The fused representations are then fed into the decoder and discriminator respectively: the decoder is used to reconstruct DI, VI, and the fused image to construct the reconstruction loss; the discriminator is used to generate classification probabilities and construct the classification loss. Joint optimization of the two types of losses guides the model to generate a fused image with a more complete structure and more consistent semantics. The following details the design of the network architecture, the collaborative mechanism of key modules, the introduction of each module, and the design of the loss function.

[0046] This invention constructs the aforementioned "dual-path collaborative optimization" architecture. Unlike traditional single-output fusion networks, this architecture creatively unifies the two objectives of "pixel-level fidelity optimization" and "feature-level semantic authenticity optimization" within a single framework by introducing a multi-task reconstruction path in the decoder and a feature discrimination path in the discriminator. The modules are not simply connected in series, but rather form a closed-loop feedback system through a unique collaborative mechanism: the bidirectional attention fusion module provides high-quality fusion features for both paths, while the reconstruction loss and adversarial loss generated by the two paths propagate in reverse, jointly guiding the bidirectional attention fusion module and the encoder to learn how to generate more complementary and discriminative features. The final fused image is displayed through pseudo-colorization and interactive visualization.

[0047] The collaboration between the shared encoder and the bidirectional attention fusion module: The shared encoder is the fundamental guarantee of this architecture. It uses a weighted reuse mechanism to process DI and VI, and its core purpose is to force the features of the two modalities to be mapped to the same high-dimensional semantic space during the extraction stage. This "pre-alignment" strategy eliminates the semantic gap caused by modal differences in the subsequent bidirectional attention calculation in the bidirectional attention fusion module, allowing the attention mechanism to focus on learning the complementary relationships between modalities, rather than laboriously aligning basic semantics. The bidirectional attention fusion module, based on this, achieves accurate querying and compensation between the blood flow dynamics information of DI and the vascular structure information of VI through bidirectional cross-attention, realizing the sublimation from "feature alignment" to "feature fusion".

[0048] The collaboration between the bidirectional attention fusion module and the dual-output paths: The fused features output by the bidirectional attention fusion module are the "cornerstone" of subsequent processing in the entire system. This architecture simultaneously feeds them to two parallel paths with distinct functions: Decoder Path: This path is responsible for pixel-level fidelity optimization. It constructs a robust reconstruction loss by simultaneously reconstructing the DI, VI, and fused images. This design forces the fusion features generated by the bidirectional attention fusion module to contain sufficiently rich and accurate source image information to ensure reconstruction accuracy, thus fundamentally constraining the fusion result from deviating from the original anatomical and functional information.

[0049] Discriminator Path: This path is responsible for optimizing the semantic plausibility at the feature level. The discriminator does not directly process image pixels, but rather performs a plausibility assessment on the fused features output by the bidirectional attention fusion module. This process ensures the naturalness and clinical plausibility of the fusion result from a higher-dimensional semantic level, avoiding potential structural distortions or semantic inconsistencies.

[0050] Contribution of dual output paths to the joint loss function and closed-loop optimization: The outputs of the dual paths mentioned above together constitute the joint loss function proposed in this invention.

[0051] The decoder path generates a reconstruction loss, which is explicit and strongly constrained. It focuses on the pixel-level accuracy of the fusion result, ensuring the authenticity of the anatomical structure reconstruction.

[0052] The discriminator path generation adversarial loss is implicit and guided. It focuses on the reasonable distribution of the fusion result in the feature space, ensuring the visual naturalness and semantic consistency of the result.

[0053] The combined effect of these two losses forms a closed-loop optimization system: the gradient signals generated by the reconstruction loss and the adversarial loss propagate back through the system, jointly guiding the bidirectional attention fusion module (and the preceding encoder) to learn how to generate optimal fusion features that are both faithful to the source image and conform to clinical visual expectations. This synergistic mechanism is the key to the high-quality fusion results generated in this invention.

[0054] Encoder: This invention employs ConvNext as the core backbone network of the encoder, primarily based on its superior performance in computer vision and its deep compatibility with the needs of medical image processing. The ConvNext architecture, through deep optimization of convolutional networks, maintains the inherent advantages of convolutional operations—such as translational variability and local feature extraction capabilities—while achieving representational capabilities comparable to Transformers. Its hierarchical feature pyramid structure is particularly suitable for processing multi-scale, multi-level anatomical and functional information in medical images. Compared to traditional convolutional neural networks (CNNs), ConvNext's introduction of deep separable convolution, inverse bottleneck structure, and refined activation function design achieves a better balance between parameter efficiency and feature extraction performance, making it especially suitable for processing 512×512 high-resolution input image data. Its core value lies in providing a high-quality feature foundation for subsequent modules. This encoder processes the 512×512 input image through four stages of downsampling, sequentially generating multi-scale feature maps with [96, 192, 384, 768] channels. Specifically, this invention selects the 384-dimensional features output from the third stage as the input to the fusion module. This design achieves an optimal balance between receptive field and computational efficiency, ensuring that subsequent bidirectional attention fusion can simultaneously capture local details and global contextual information. The encoder employs a shared weight mechanism to process both DI and VI modalities, ensuring consistency in feature extraction while significantly reducing the number of model parameters, thus laying a solid foundation for the efficient training of the entire fusion network.

[0055] Decoder: This invention employs a lightweight dynamic upsampling decoder. Its core function is to reconstruct the high-level semantic features (384 dimensions, 32×32 spatial size) output by the fusion module into a high-resolution fused image of 512×512, and simultaneously reconstruct the corresponding reconstructed image from the input image feature map extracted by the encoder. This decoder achieves progressive resolution restoration through multiple cascaded upsampling modules. Each module integrates transposed convolution, batch normalization, and the GELU activation function, progressively increasing the feature map size and gradually compressing the channel dimensions at a stride of 2. Its unique dynamic architecture adaptively calculates the required number of upsampling levels, ensuring that the output size precisely matches the target resolution. Finally, a 3×3 convolutional layer maps the number of channels to the target output channels, and bilinear interpolation is introduced to achieve pixel-level size calibration, thereby generating a high-quality reconstructed image with complete structure and clear details.

[0056] like Figure 4 As shown, the bidirectional attention fusion module includes: The first linear layer is used to linearly project the diffusion feature map into the value vector and key vector corresponding to the diffusion feature; The second linear layer is used to linearly project the blood vessel feature map into query vectors, value vectors, and key vectors corresponding to the blood vessel features; The third linear layer is used to linearly project the diffusion feature map into a query vector corresponding to the diffusion feature. The first matrix multiplication processing unit is used to perform matrix multiplication on the key vector corresponding to the diffusion feature and the query vector corresponding to the blood vessel feature to obtain the first multiplication result; The first normalization processing unit is used to normalize the first multiplication result to obtain the normalized result of the first multiplication. The first random deactivation processing unit is used to randomly deactivate the normalized result of the first multiplication to obtain the first random deactivation result; The second matrix multiplication processing unit is used to perform matrix multiplication on the key vector corresponding to the blood vessel feature and the query vector corresponding to the diffusion feature to obtain the second multiplication result; The second normalization processing unit is used to normalize the result of the second multiplication to obtain the normalized result of the second multiplication. The second random deactivation processing unit is used to randomly deactivate the normalized result of the second multiplication to obtain the second random deactivation result; The third matrix multiplication processing unit is used to multiply the value vector corresponding to the diffusion feature and the first random inactivation result to obtain the third multiplication result; The fourth linear layer is used to perform dimensionality calibration and information integration on the third multiplication result to obtain the first linearized result; The third random deactivation processing unit is used to randomly deactivate the first linearization result to obtain the third random deactivation result. The fourth matrix multiplication processing unit is used to multiply the value vector corresponding to the blood vessel features and the second random inactivation result to obtain the fourth multiplication result; The fifth linear layer is used to perform dimensionality calibration and information integration on the fourth multiplication result to obtain the second linearization result; The fourth random deactivation processing unit is used to randomly deactivate the second linearization result to obtain the fourth random deactivation result; The weighted summation processing unit is used to perform weighted summation on the third and fourth random inactivation results to obtain the fused feature representation.

[0057] In this embodiment, the fourth and fifth linear layers have similar functions. Their core function is to perform dimensional calibration and information integration on the attention-weighted features obtained from the matrix multiplication in the previous step, so as to ensure that the output features can be adapted to the subsequent fusion process.

[0058] Dimension alignment: Ensures that the feature dimensions after attention calculation are consistent with the original input feature dimensions, thus enabling direct connection to subsequent fusion operations such as weighted summation.

[0059] Information re-encoding: The output of the attention mechanism is "the result of calculating the value features according to the attention weights", but there may be redundancy in feature representation. By using a learnable weight matrix, these weighted features are linearly combined to strengthen key information (such as the interaction information between diffusion features and vascular features) and suppress noise, making the features more discriminative.

[0060] In this embodiment, the bidirectional attention fusion module is the core module of the network architecture. Its function is to overcome the limitations of existing fusion methods in terms of the depth and breadth of feature interaction, and to achieve global, adaptive fusion between the diffusion feature map and the vascular feature map. The specific structural design of this module is as follows: Figure 4 As shown, its input is the feature map output by the third stage of the encoder, namely the diffusion feature map and the vascular feature map with a channel dimension of 384 and a spatial size of 32×32.

[0061] This module abandons traditional convolutional or simple concatenation strategies, adopting a dual-path parallel architecture based entirely on an attention mechanism. Specifically, the module contains two structurally symmetrical but functionally complementary cross-attention paths. In the first path, the vascular feature map is linearly projected to generate a query vector, while the diffusion feature map generates key and value vectors respectively; the second path does the reverse. In each path, the projection dimensions of the query, key, and value vectors are strictly maintained at 384 dimensions. Subsequently, the precise attention weights are calculated through matrix multiplication of the query and key vectors, scaling factor adjustment, and Softmax normalization. This invention sets the number of attention heads to 8, with each attention head having a dimension of 48. This configuration enables the model to capture different types of feature dependencies from multiple representation subspaces in parallel.

[0062] After weighted summation using the aforementioned attention mechanism, each path outputs an enhanced feature with a channel dimension of 384. These two features are stabilized through residual connections and layer normalization during training, and then integrated element-wise. The integrated features are further transformed nonlinearly through a feedforward neural network. This network first expands the feature dimension from 384 to 1536 to enhance representational power, and then projects it back to 384 dimensions after passing through the GELU activation function. Finally, the module outputs a fused feature representation with a spatial resolution of 32×32 and maintaining 384 channels.

[0063] like Figure 5 As shown, the discriminator includes: The feature input end is used to obtain diffusion feature maps, blood vessel feature maps, and fusion feature representations to obtain feature sequences; The dimension rearrangement unit is used to rearrange the feature sequence according to the adapted dimension of the global average pooling unit to obtain a dimension-adapted feature sequence. The global average pooling unit is used to aggregate features from dimension-adapted feature sequences using global average pooling operations to obtain global feature descriptors. The flattening operation unit is used to flatten the global feature descriptor to obtain a compact one-dimensional feature vector; The discriminant unit is used to extract features from a compact one-dimensional feature vector using three cascaded discriminant structures to obtain discriminant features; each discriminant structure includes a fully connected layer, layer normalization, activation function, and random deactivation connected in sequence. The first fully connected layer is used to map discriminative features to classification probabilities.

[0064] In this embodiment, the core function of the discriminator is to ensure the rationality and authenticity of the fusion result from the feature semantic level. Its specific structure is as follows: Figure 5 As shown. Unlike traditional discriminators that directly determine the authenticity of images, the input of this discriminator is the fused feature sequence output by the bidirectional attention fusion module and the input feature sequence extracted by the encoder. This enables it to make discriminations in a high-dimensional feature space, guiding the generator to produce fused features that are not only pixel-accurate but also have a semantic distribution that meets expectations.

[0065] The discriminator's data processing flow begins with the structured reconstruction of the input feature sequence: first, the 1024×384 sequence format is adjusted to 384×1024 through dimensional rearrangement to fit the input requirements of the pooling layer; then, global average pooling is used to aggregate all feature vectors in the sequence, compressing them into a global feature descriptor of dimension 384×1; finally, a flattening operation is performed to obtain a compact one-dimensional feature vector of dimension 384. This process effectively converts the feature sequence into a fixed-dimensional representation while preserving the most critical global information.

[0066] Subsequently, the 384-dimensional global feature vector is fed into an MLP core consisting of three cascaded discriminative modules. Each discriminative module strictly follows the sequence of "fully connected layer → layer normalization → activation function → random deactivation." The fully connected layer is responsible for feature transformation and dimension mapping; layer normalization is used to stabilize the training process and accelerate convergence; the activation function introduces nonlinearity to enhance the model's discriminative ability; and random deactivation serves as an effective regularization method to prevent overfitting. The repeated stacking of these three modules allows the discriminator to progressively refine features and construct a robust nonlinear decision boundary.

[0067] Finally, the features refined by the three discriminative modules are mapped to a single discriminative logits value by an independent fully connected output layer. This logits is calculated using a binary cross-entropy loss function to distinguish whether the distribution of the fused features is consistent with that of the real image features. Through this adversarial training in the feature domain rather than the pixel domain, the discriminator and the generator (encoder-fusion module-decoder) form an effective game, jointly optimizing the network to generate fused images with both high fidelity and high semantic plausibility.

[0068] The loss function of the sparse image fusion and visualization model includes classification loss and reconstruction loss:

[0069] in, For sparse image fusion and visualization models; For classification loss; To rebuild the losses.

[0070] In this embodiment, the present invention addresses the core challenges of incomplete feature preservation and significant background artifacts in sparse cerebral vascular image fusion tasks by designing a joint loss function based on multi-task learning. This loss function, through the synergistic optimization of reconstruction loss and classification loss, simultaneously imposes constraints on both pixel-level fidelity and feature-level semantic consistency, significantly improving the structural integrity and visual credibility of the fused image.

[0071] The expression for the classification loss is:

[0072] in, The binary cross-entropy classification loss is used for the diffusion feature map; For binary cross-entropy loss; The classification result obtained by the discriminator based on the diffusion feature map; The binary cross-entropy classification loss is used for the vascular feature map. The classification result obtained by the discriminator based on the vascular feature map; The binary cross-entropy classification loss is used to fuse feature representations; The classification result is obtained by the discriminator based on the fused feature representation.

[0073] In this embodiment, to further improve the quality of the fused image, a classification loss is constructed, and the discrimination results are calculated for the three types of images respectively. The classification loss is used to measure the accuracy of the discriminator in classifying the generated fused image, diffuse image, and blood vessel image.

[0074] In this embodiment, This is the classification output of the discriminator on the diffusion feature map, with the true label being 1; The goal is to constrain the discriminator to "identify the diffused image as the true source image" through binary cross-entropy loss. This constraint ensures that the features of the diffused image are accurately preserved during the fusion process. If the structural and semantic information of the diffused image is lost during fusion, the discriminator will misclassify it. Consequently, backpropagation of the loss forces the generator (fusion model) to enhance the restoration of the diffused features, ensuring the integrity of the anatomical structure and functional information of the diffused modality in the fused image.

[0075] This is the classification output of the discriminator on the vascular feature map, with the true label being 0; The discriminator is constrained by a binary cross-entropy loss function to "identify blood vessel images as real source images" (label 0 is a task-defined category identifier used to distinguish different modalities). This constraint ensures that the features of the blood vessel images are accurately preserved during fusion. If the morphology, distribution, or other information of the blood vessels in the images is distorted during fusion, the discriminator's classification error will drive the generator to optimize, ensuring that the core information of the blood vessel modalities in the fused image is not lost.

[0076] This is the classification output of the discriminator on the fused feature map, with a true label of 0.5; The goal is to constrain the discriminator to either explicitly classify the fused image as a diffusion or vascular image using a binary cross-entropy loss. This design forces the fused image to possess "cross-modal semantic consistency and fusion specificity." If the fused image is biased towards a single modality (such as resembling only a diffusion image or only a vascular image), the discriminator can easily classify it, and the loss will drive the generator to adjust, ultimately allowing the fused image to simultaneously contain the key information of both modalities and present a natural fusion effect (neither pure diffusion nor pure vascular, but an effective combination of the two). While learning to distinguish between pure diffusion (1) and pure vascular (0), the discriminator's requirement for "ambiguity" in the fused image (0.5) essentially forces the fusion module to retain the core features of both modalities during fusion, without bias towards either, ultimately achieving effective cross-modal fusion of "1+0→0.5". Here, "real labels" are artificially defined category identifiers. By assigning label 1 to the diffused image, label 0 to the vascular image, and label 0.5 to the fused image, a constraint system of "modal differentiation-fusion specificity" is constructed.

[0077] The expression for the reconstruction loss is:

[0078] in, These are the weighting coefficients for the loss in the diffuse image reconstruction. For the loss of image reconstruction due to diffusion; These are the weighting coefficients for the blood vessel image reconstruction loss; Loss due to vascular image reconstruction; The weighting coefficient for penalizing losses; To punish the loss; This is the original diffusion image; For the reconstructed diffusion image; This is the original image of the blood vessels; Reconstructed vascular images; It is an L1 norm; For indicator functions, when =0 and If the value is 0, the value is 1; otherwise, the value is 0. The first of the original diffused image The pixel value of each pixel; The first of the original blood vessel images The pixel value of each pixel; The first fused image for reconstruction The pixel value of each pixel; For pixel index; It is an absolute value.

[0079] In this embodiment, the reconstruction loss aims to ensure that the fusion result accurately restores the key anatomical structures and functional information of the source image. Considering the characteristics of sparse images, such as a high proportion of zero-value regions and sparse distribution of effective signals, this invention abandons the L2 norm, which is sensitive to outliers, and adopts the more robust L1 norm as a distance metric to effectively suppress impulse noise interference. To suppress the generation of artifacts in the background region of the fused image, this invention introduces a penalty loss into the reconstruction loss. The penalty loss is calculated for zero-value regions, ensuring that the model suppresses the generation of significant noise in zero-value regions.

[0080] The trained sparse image fusion and visualization model no longer needs to reconstruct diffuse and vascular images when applied.

[0081] In this embodiment, loss calculation refers to the process during training. During training, the model continuously calculates the loss to minimize it, ultimately arriving at an optimal loss. The image reconstructed using the optimal parameters obtained at this point is the desired reconstructed fused image. This reconstructed image can be considered the model-generated fused image, or the optimal fused image. The fused feature map, diffusion feature map, and vascular feature map obtained during training can be seen as intermediate products of the model training process, intended to calculate and optimize the loss to ultimately generate the optimal fused image. Therefore, in practical applications, once the model is trained, these loss calculation operations are no longer needed; the optimal parameters can be directly used to generate the optimal fused image.

[0082] In this embodiment, based on the above design, the model is trained and validated to generate high-quality fused images. During the training phase, the Adam optimizer is used, with a learning rate set to 1×10⁻⁶. 4 The batch size was 12, and the training run consisted of 80 rounds. During the validation phase, four metrics—Peak Signal-to-Noise Ratio (PSNR), Structural Similarity (SSIM), Mutual Information (MI), and Vessel Sensitivity—were used to evaluate the fusion results, measuring the fidelity and reliability of the fused image in terms of blood flow diffusion features and vascular structure features. Some input images and the fusion results are shown below. Figure 6 As shown in the figure. Some experimental results are shown in Table 1.

[0083] Table 1. Comparison with sparse model methods on self-built sparse datasets.

[0084] Visualization and Pseudo-color Mapping: After sparse image fusion, to comprehensively reveal blood flow diffusion characteristics, pseudo-color mapping and 3D interactive visualization are used to display these characteristics based on the fused images. Pseudo-color mapping, after normalization, uses a blue-to-red gradient to represent the diffusion speed from slow to fast, highlighting vascular structures to maintain anatomical clarity. 3D interactive visualization supports mouse rotation for model observation, coordinate tracking, and dynamic color switching, intuitively depicting the spatial relationship between vascular morphology and blood flow diffusion, providing support for lesion identification and preoperative planning. The visualization effect is as follows: Figure 7 As shown.

Claims

1. A method for assessing cerebral blood flow diffusion based on attention-guided sparse image fusion, characterized in that, include: Collect raw dataset of CTA image sequences for ischemic moyamoya disease; Based on the original CTA image sequence dataset, several sparse image pairs, each including diffuse and vascular images, were constructed, and the sparse image pairs were integrated to obtain the image training set. A sparse image fusion and visualization model is constructed. The sparse image fusion and visualization model includes a ConvNext encoder, a bidirectional attention fusion module, a discriminator and a decoder, which are respectively connected to the diffusion image input end and the blood vessel image input end, as well as a pseudo-color image construction module and an interactive three-dimensional visualization module, which are respectively connected to the decoder. The bidirectional attention fusion module is connected to the ConvNext encoder, discriminator, and decoder, respectively. A sparse image fusion and visualization model is trained using an image training set to obtain the trained sparse image fusion and visualization model. The trained sparse image fusion and visualization model was used to process the CTA images of ischemic moyamoya disease to obtain the corresponding reconstructed fused images. The evaluation was then completed based on the pseudo-color images and visualization results of the corresponding reconstructed fused images.

2. The method for assessing cerebral blood flow diffusion based on attention-guided sparse image fusion according to claim 1, characterized in that, The construction of several sparse image pairs, each including diffuse and vascular images, is specifically as follows: By excluding sequence data in the original CTA image sequence dataset that have a consistency lower than the consistency threshold, a high-quality image dataset is obtained. For each high-quality image data in the high-quality image dataset, blood flow diffusion features and vascular structure information are extracted to obtain blood flow diffusion feature images and vascular structure information images. The blood flow diffusion feature image and the vascular structure information image were standardized to obtain diffusion image and vascular image, respectively. By integrating the diffuse image and vascular image corresponding to the same high-quality image data into a set of sparse image pairs, several sets of sparse image pairs, each including diffuse image and vascular image, are obtained.

3. The method for assessing cerebral blood flow diffusion based on attention-guided sparse image fusion according to claim 1, characterized in that, The ConvNext encoder is used to extract image features from the diffusion image and the blood vessel image respectively, to obtain the diffusion feature map and the blood vessel feature map; The bidirectional attention fusion module is used to fuse the diffusion feature map and the vascular feature map to obtain a fused feature representation; The discriminator is used to generate classification probabilities based on the diffusion feature map, the vascular feature map, and the fusion feature representation; The decoder is used to reconstruct the diffusion image, the vascular image, and the fused image based on the diffusion feature map, the vascular feature map, and the fusion feature representation, respectively. The pseudo-color image construction module is used to generate a pseudo-color image based on the reconstructed fused image; The interactive 3D visualization module is used to visualize the reconstructed fused image.

4. The method for assessing cerebral blood flow diffusion based on attention-guided sparse image fusion according to claim 1, characterized in that, The bidirectional attention fusion module includes: The first linear layer is used to linearly project the diffusion feature map into the value vector and key vector corresponding to the diffusion feature; The second linear layer is used to linearly project the blood vessel feature map into query vectors, value vectors, and key vectors corresponding to the blood vessel features; The third linear layer is used to linearly project the diffusion feature map into a query vector corresponding to the diffusion feature. The first matrix multiplication processing unit is used to perform matrix multiplication on the key vector corresponding to the diffusion feature and the query vector corresponding to the blood vessel feature to obtain the first multiplication result; The first normalization processing unit is used to normalize the first multiplication result to obtain the normalized result of the first multiplication. The first random deactivation processing unit is used to randomly deactivate the normalized result of the first multiplication to obtain the first random deactivation result; The second matrix multiplication processing unit is used to perform matrix multiplication on the key vector corresponding to the blood vessel feature and the query vector corresponding to the diffusion feature to obtain the second multiplication result; The second normalization processing unit is used to normalize the result of the second multiplication to obtain the normalized result of the second multiplication. The second random deactivation processing unit is used to randomly deactivate the normalized result of the second multiplication to obtain the second random deactivation result; The third matrix multiplication processing unit is used to multiply the value vector corresponding to the diffusion feature and the first random inactivation result to obtain the third multiplication result; The fourth linear layer is used to perform dimensionality calibration and information integration on the third multiplication result to obtain the first linearized result; The third random deactivation processing unit is used to randomly deactivate the first linearization result to obtain the third random deactivation result. The fourth matrix multiplication processing unit is used to multiply the value vector corresponding to the blood vessel features and the second random inactivation result to obtain the fourth multiplication result; The fifth linear layer is used to perform dimensionality calibration and information integration on the fourth multiplication result to obtain the second linearization result; The fourth random deactivation processing unit is used to randomly deactivate the second linearization result to obtain the fourth random deactivation result; The weighted summation processing unit is used to perform weighted summation on the third and fourth random inactivation results to obtain the fused feature representation.

5. The method for assessing cerebral blood flow diffusion based on attention-guided sparse image fusion according to claim 1, characterized in that, The discriminator includes: The feature input end is used to obtain diffusion feature maps, blood vessel feature maps, and fusion feature representations to obtain feature sequences; The dimension rearrangement unit is used to rearrange the feature sequence according to the adapted dimension of the global average pooling unit to obtain a dimension-adapted feature sequence. The global average pooling unit is used to aggregate features from dimension-adapted feature sequences using global average pooling operations to obtain global feature descriptors. The flattening operation unit is used to flatten the global feature descriptor to obtain a compact one-dimensional feature vector; The discriminant unit is used to extract features from a compact one-dimensional feature vector using three cascaded discriminant structures to obtain discriminant features; each discriminant structure includes a fully connected layer, layer normalization, activation function, and random deactivation connected in sequence. The first fully connected layer is used to map discriminative features to classification probabilities.

6. The method for assessing cerebral blood flow diffusion based on attention-guided sparse image fusion according to claim 1, characterized in that, The loss function of the sparse image fusion and visualization model includes classification loss and reconstruction loss: in, For sparse image fusion and visualization models; For classification loss; To rebuild the losses.

7. The method for assessing cerebral blood flow diffusion based on attention-guided sparse image fusion according to claim 6, characterized in that, The expression for the classification loss is: in, The binary cross-entropy classification loss is used for the diffusion feature map; For binary cross-entropy loss; The classification probability obtained by the discriminator based on the diffusion feature map; The binary cross-entropy classification loss is used for the vascular feature map. The classification result obtained by the discriminator based on the vascular feature map; The binary cross-entropy classification loss is used to fuse feature representations; The classification result is obtained by the discriminator based on the fused feature representation.

8. The method for assessing cerebral blood flow diffusion based on attention-guided sparse image fusion according to claim 6, characterized in that, The expression for the reconstruction loss is: in, These are the weighting coefficients for the loss in the diffuse image reconstruction. For the loss of image reconstruction due to diffusion; These are the weighting coefficients for the blood vessel image reconstruction loss; Loss due to vascular image reconstruction; The weighting coefficient for penalizing losses; To punish the loss; This is the original diffusion image; For the reconstructed diffusion image; This is the original image of the blood vessels; Reconstructed vascular images; It is an L1 norm; For indicator functions, when =0 and If the value is 0, the value is 1; otherwise, the value is 0. The first of the original diffusion image The pixel value of each pixel; The first of the original blood vessel images The pixel value of each pixel; The first fused image for reconstruction The pixel value of each pixel; For pixel index; It is an absolute value.

9. The method for assessing cerebral blood flow diffusion based on attention-guided sparse image fusion according to claim 1, characterized in that, The trained sparse image fusion and visualization model no longer needs to reconstruct diffuse and vascular images when applied.

Images