An intraoperative magnetic resonance image reconstruction method and system based on medical registration guidance

Abstract

The present application relates to the field of medical magnetic resonance image reconstruction and intraoperative navigation, and particularly relates to an intraoperative magnetic resonance image reconstruction method and system based on medical registration guidance, comprising: acquiring preoperative high-field magnetic resonance images and intraoperative low-field magnetic resonance images of the same patient; inputting the preoperative high-field magnetic resonance images and the intraoperative low-field magnetic resonance images into a trained image processing model; performing registration by taking the preoperative images as the moving images and the intraoperative images as the fixed images through a registration network, generating a deformation field from the preoperative to the intraoperative, and correcting the brain structure displacement caused by the surgery; performing spatial transformation on the preoperative images based on the deformation field to obtain high-field images aligned with the intraoperative images; and fusing the aligned high-field prior information and the intraoperative low-field image features by using a reconstruction network to reconstruct high-quality target images. The present application effectively improves the definition and detail resolution of the intraoperative images by introducing individualized preoperative prior, and provides more accurate image support for surgical navigation.

Description

Technical Field

[0001] This invention relates to the field of medical magnetic resonance image reconstruction and intraoperative navigation technology, specifically to an intraoperative magnetic resonance image reconstruction method and system based on medical registration guidance. Background Technology

[0002] Intraoperative magnetic resonance imaging (MRI) is a crucial support for the precision advancement of neurosurgery. This technology provides surgeons with dynamic anatomical and pathological information by acquiring MRI images in real time during surgical procedures, playing a key role, especially in delicate surgeries such as brain tumor resection. Its core value is mainly reflected in three aspects: First, it enables real-time visualization of tumor boundaries, helping to remove lesions to the greatest extent possible while protecting important functional areas; Secondly, it corrects brain structural displacement (i.e., "brain drift") caused by cerebrospinal fluid loss, tissue resection or traction after craniotomy, making up for the limitations of relying solely on preoperative image navigation. Third, it serves as a tool for intraoperative quality control, enabling timely detection of residual lesions or complications such as bleeding, thereby improving surgical safety.

[0003] However, due to limitations in operating room environment, cost, and safety regulations, most intraoperative MRI systems used in clinical practice employ low-field open magnets, resulting in significant deficiencies in image quality. Such images typically exhibit low signal-to-noise ratios, limited spatial resolution, and poor tissue contrast, affecting the identification of subtle lesion boundaries and critical neural structures, potentially hindering the accuracy of surgical decisions.

[0004] To improve the clarity and usability of intraoperative MRI images, current enhancement techniques primarily focus on post-processing of acquired images. These techniques can be broadly categorized into two approaches: Firstly, data-driven models or traditional signal processing techniques can be used to directly enhance the resolution and suppress noise in low-quality images acquired during surgery. However, these methods rely entirely on the limited information contained in the low-quality images themselves. When processing lesion areas, they often fail to restore true anatomical details and may instead generate false textures lacking anatomical basis or cause blurred edges, thus reducing the reliability of the results in clinical diagnosis.

[0005] Secondly, a strategy of introducing external reference information is adopted, such as spatially aligning general brain anatomy atlases or image templates of healthy volunteers with intraoperative images of patients. However, such standard atlases reflect the common anatomical structures of a group or in a healthy state, and have systematic deviations from the unique brain anatomy and specific pathological changes of a particular patient (such as the space-occupying effect of tumors, the distribution of surrounding edema zones, etc.), making it difficult to provide prior knowledge that truly fits the individual situation.

[0006] The above reasons together result in the limited performance and insufficient reliability of existing enhancement technologies when faced with complex intraoperative scenarios. They are unable to stably generate high-definition and high-reliability intraoperative images while ensuring anatomical authenticity, thus failing to provide surgeons with accurate and safe decision support. Summary of the Invention

[0007] To address the above technical problems, this invention provides a technical solution for an intraoperative magnetic resonance image reconstruction method and system based on medical registration guidance.

[0008] The technical problem solved by this invention can be achieved by the following technical solutions: A method for intraoperative magnetic resonance imaging reconstruction based on medical registration guidance, comprising: Step S1: Acquire preoperative high-field magnetic resonance imaging (MRI) images and intraoperative low-field MRI images of the same patient. Step S2: Input the preoperative high-field magnetic resonance image and the intraoperative low-field magnetic resonance image into the trained image processing model; Step S3: Using the registration network in the image processing model, the preoperative high-field magnetic resonance image is used as the moving image and the intraoperative low-field magnetic resonance image is used as the fixed image for registration, thereby generating a deformation field from the preoperative image space to the intraoperative image space. Step S4: Based on the deformation field, perform spatial transformation on the preoperative high-field magnetic resonance image to obtain a registered preoperative high-field image that is spatially aligned with the intraoperative low-field magnetic resonance image. Step S5: Using the reconstruction network in the image processing model, feature fusion and reconstruction are performed on the registered preoperative high-field image and the intraoperative low-field magnetic resonance image to obtain the target image.

[0009] Preferably, the registration network is an unsupervised learning network, which includes a feature encoder, a dual attention module, and a deformation field prediction layer; The dual attention module includes a spatial attention mechanism and a channel attention mechanism.

[0010] Preferably, step S3 includes: Step S31: Use the feature encoder to perform joint feature encoding on the preoperative high-field magnetic resonance image and the intraoperative low-field magnetic resonance image to generate a multi-scale joint feature map; Step S32: Use the spatial attention mechanism to perform feature processing and weighting operations on the multi-scale joint feature map to obtain the first weighted feature map; Step S33: Use the channel attention mechanism to perform feature processing and weighting operations on the multi-scale feature map to obtain a second weighted feature map; Step S34: Fuse the first weighted feature map and the second weighted feature map to obtain a weighted fused feature map; Step S35: Process the weighted fusion feature map using the deformation field prediction layer to generate the deformation field.

[0011] Preferably, the deformation field is constrained by an adaptive multi-scale regularization loss function, the specific formula of which is expressed as: L_reg=λ1 * L_smooth+λ2 * L_Jacobian L_smooth=∑ p∈Ω ( || Φ(p)||2) 2 L_Jacobian=∑ p∈Ω φ( |J Φ (p)| ) / |Ω| Where L_reg is the regularization loss function, L_smooth is the global smoothing loss, L_Jacobian is the local Jacobian determinant penalty, λ1 and λ2 are adjustable weight coefficients, p is a spatial coordinate in the image spatial domain Ω, and φ is an adaptive penalty function based on the Jacobian determinant value. Let J be the Laplace operator, Φ be the deformation field, and J be the deformation field. Φ Let be the determinant of the Jacobian matrix of the deformation field.

[0012] Preferably, the reconstruction network is a deep learning network based on a dual encoder-single decoder architecture, and the reconstruction network includes a high-field encoder, a low-field encoder, a feature fusion module, and a decoder; The feature fusion module is a multi-level feature modulation network that includes spatial attention mechanism and channel attention mechanism.

[0013] Preferably, step S5 includes: Step S51: Use the high-field encoder to perform multi-level feature encoding on the registered preoperative high-field image to obtain a preoperative high-field multi-scale feature map. Step S52: Use the low-field encoder to perform multi-level feature encoding on the intraoperative low-field magnetic resonance image to obtain an intraoperative low-field multi-scale feature map. Step S53: Based on the intraoperative low-field multi-scale feature map and the intraoperative low-field multi-scale feature map, the feature fusion module is used to perform spatial modulation and channel modulation on the preoperative high-field multi-scale feature map, and the modulated features are injected into the low-field features to obtain a fused enhanced feature map. Step S54: Upsample and reconstruct the fused enhanced feature map using the decoder to obtain the target image.

[0014] Preferably, step S53 includes: Step S531: Based on the preoperative high-field multi-scale feature map and the intraoperative low-field multi-scale feature map, a spatial attention mask is generated through the spatial attention mechanism, and the spatial attention mask is used to spatially modulate the preoperative high-field multi-scale feature map. Step S532: Based on the preoperative high-field multi-scale feature map and the intraoperative low-field multi-scale feature map, a channel weight vector is generated through the channel attention mechanism, and the channel weight vector is used to perform channel modulation on the preoperative high-field multi-scale feature map. Step S533: The features after spatial modulation and channel modulation are fused to obtain the modulated features; Step S534: The modulated features are injected into the intraoperative low-field multi-scale feature map to obtain the fused enhanced feature map.

[0015] Preferably, the training process of the image processing model is divided into two stages, including: In the first training phase, the registration network in the image processing model is trained separately. In the second training phase, the registration network is fixed and the reconstruction network in the image processing model is trained.

[0016] Preferably, the reconstruction network is trained using a hybrid loss function, the specific formula of which is expressed as: L_total=λ3*L_intensity+λ4*L_perceptual+λ5*L_anatomical_fidelity; Where L_total is the mixed loss function, L_intensity is the image intensity loss, L_perceptual is the perceptual loss, L_anatomical_fidelity is the anatomical fidelity loss, and λ3, λ4, and λ5 are all weighting coefficients that balance the various losses.

[0017] A medical registration-guided intraoperative magnetic resonance imaging (MRI) image reconstruction system is used to implement the medical registration-guided intraoperative MRI image reconstruction method described above, comprising: The image acquisition module is used to acquire preoperative high-field magnetic resonance images and intraoperative low-field magnetic resonance images of the same patient. The model loading module, connected to the image acquisition module, is used to input the preoperative high-field magnetic resonance image and the intraoperative low-field magnetic resonance image into the trained image processing model. The deformation field generation module, connected to the model loading module, is used to extract and transform multimodal features from the preoperative high-field magnetic resonance image and the intraoperative low-field magnetic resonance image using the registration network in the image processing model, thereby generating a deformation field from the preoperative image space to the intraoperative image space. The image registration module, connected to the deformation field generation module, is used to perform spatial transformation on the preoperative high-field magnetic resonance image based on the deformation field to obtain a registered preoperative high-field image that is spatially aligned with the intraoperative low-field magnetic resonance image. The image reconstruction module, connected to the image registration module, is used to perform feature fusion and reconstruction of the registered preoperative high-field image and the intraoperative low-field magnetic resonance image using the reconstruction network in the image processing model to obtain the target image.

[0018] Beneficial effects: This invention introduces preoperative high-field magnetic resonance images of the same patient as individualized prior information, and uses a trained registration network to generate a deformation field from preoperative to intraoperative, effectively correcting the brain structure displacement caused by surgery. Subsequently, a reconstruction network is used to fuse the spatially aligned high-field image prior with the low-quality intraoperative image. Under the premise of ensuring the authenticity of individual anatomy and pathological specificity, the clarity and detail resolution of intraoperative images can be significantly improved, thereby providing surgeons with more accurate and reliable real-time image navigation and decision support. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the method steps of the present invention; Figure 2 This is an overall flowchart of the method of the present invention; Figure 3 This is a flowchart of the image reconstruction process of the present invention; Figure 4 This is a system block diagram of the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other.

[0022] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, but this is not intended to limit the scope of the invention.

[0023] Reference Figure 1 This invention provides a method for intraoperative magnetic resonance imaging reconstruction based on medical registration guidance, comprising: Step S1: Acquire preoperative high-field magnetic resonance imaging (MRI) images and intraoperative low-field MRI images of the same patient. Step S2: Input the preoperative high-field magnetic resonance image and the intraoperative low-field magnetic resonance image into the trained image processing model; Step S3: Using the registration network in the image processing model, the preoperative high-field magnetic resonance image is used as the moving image and the intraoperative low-field magnetic resonance image is used as the fixed image for registration, thereby generating a deformation field from the preoperative image space to the intraoperative image space. Step S4: Based on the deformation field, perform spatial transformation on the preoperative high-field magnetic resonance image to obtain a registered preoperative high-field image that is spatially aligned with the intraoperative low-field magnetic resonance image. Step S5: Using the reconstruction network in the image processing model, feature fusion and reconstruction are performed on the registered preoperative high-field image and the intraoperative low-field magnetic resonance image to obtain the target image.

[0024] Specifically, in this embodiment of the invention, to address the problems of false textures, blurred boundaries, and mismatch with pathological structures caused by the lack of individualized, high-resolution prior information in existing intraoperative image enhancement technologies, an end-to-end registration-reconstruction two-stage model is designed. First, a deep learning registration network is used to accurately capture and correct the nonlinear deformation of anatomical structures caused by "brain drift," and high-quality preoperative images are individually mapped to the intraoperative space. This avoids the systematic bias and insufficient information caused by directly using general templates or simply relying on intraoperative images. Under the premise of strictly following the patient's individual real-time anatomical structure and pathological characteristics, rich and accurate high-resolution prior knowledge is provided for the subsequent reconstruction network, thereby coordinating with intraoperative low-field images to reconstruct high-quality target images with clear details and realistic anatomy.

[0025] In a preferred embodiment of the present invention, the registration network is an unsupervised learning network, which includes a feature encoder, a dual attention module, and a deformation field prediction layer. The dual attention module includes a spatial attention mechanism and a channel attention mechanism.

[0026] Specifically, to effectively address the significant contrast differences and structural shifts between high- and low-field magnetic resonance (MR) images caused by varying field strengths, and to achieve accurate and robust unsupervised registration, traditional registration networks often struggle to simultaneously capture the complex semantic correspondences and subtle local deformations between cross-modal images. Based on this, referring to... Figure 2In this embodiment of the invention, a deep learning architecture combining spatial and channel attention mechanisms is employed. This deep learning architecture includes a feature encoder with shared weights, a cascaded Spatial-Channel Attention Block (SCAB), and a deformation field prediction layer. By inputting preoperative high-field MR images (moving images, I_pre) and intraoperative low-field MR images (fixed images, I_intra) in parallel into the shared encoder, multi-scale deep feature maps are extracted. These feature maps from the two modalities are then fed into the SCAB for interaction and refinement. Consequently, the deformation field prediction layer decodes the dense deformation field Φ from I_pre to I_intra from the refined high-level semantic features.

[0027] Thus, by performing a spatial transformation on I_pre based on this deformation field Φ, a preoperative high-field image I_pre_reg can be obtained that is aligned with I_intra in anatomical space but still maintains the contrast and noise characteristics of the high-field image.

[0028] In a preferred embodiment of the present invention, step S3 includes: Step S31: Use the feature encoder to perform joint feature encoding on the preoperative high-field magnetic resonance image and the intraoperative low-field magnetic resonance image to generate a multi-scale joint feature map; Step S32: Use the spatial attention mechanism to perform feature processing and weighting operations on the multi-scale joint feature map to obtain the first weighted feature map; Step S33: Use the channel attention mechanism to perform feature processing and weighting operations on the multi-scale feature map to obtain a second weighted feature map; Step S34: Fuse the first weighted feature map and the second weighted feature map to obtain a weighted fused feature map; Step S35: Process the weighted fusion feature map using the deformation field prediction layer to generate the deformation field.

[0029] Specifically, in order to accurately capture the complex correspondence between cross-modal images under unsupervised conditions and to address the challenges of local deformation caused by field strength differences and intraoperative "brain drift," the dual attention module in this embodiment of the invention adopts an efficient parallel computing structure based on a pure convolutional neural network (CNN), which has both spatial focusing and channel filtering capabilities to achieve feature refinement and enhancement.

[0030] Specifically, firstly, the feature encoder (such as the U-Net encoder part) downsamples the input I_pre and I_intra, extracts and fuses their multi-scale joint feature maps, denoted as F, with dimensions C×H×W, where C is the number of channels, and H and W are the spatial height and width. These feature maps encode cross-modal information ranging from fine textures to macroscopic structures at different scales.

[0031] Next, the spatial attention mechanism processes the multi-scale joint feature map F: a spatial weight map Ms of size 1×H×W is generated through multiple parallel convolutional layers (Conv+ReLU). This spatial weight map Ms can learn autonomously and highlight spatial regions that are significantly deformed or that contribute critically to registration (such as tumor boundaries and brain tissue edges that have been displaced due to traction). Subsequently, Ms is normalized to the range [0, 1] using the Sigmoid activation function, and then the original features are spatially weighted by element-wise multiplication to obtain the first weighted feature Fs. The specific formula is: Fs = F Ms.

[0032] Secondly, the channel attention mechanism operates on the same feature map F in parallel: First, global average pooling (GAP) is performed on F to compress spatial information and obtain a channel statistical vector v of size 1×C; then, this vector is processed through two fully connected layers (containing a ReLU activation function in between) to generate a channel weight vector wc, which is used to evaluate and strengthen feature channels that carry stable cross-modal correspondences (such as shared anatomical edges and structural contours), while suppressing redundant or noisy channels; wc is also normalized using a Sigmoid activation function; finally, by scaling the channel dimensions, the second weighted feature Fc is obtained, specifically expressed as: Fc = F wc.

[0033] Then, the spatially weighted feature Fs and the channel-weighted feature Fc are directly added together to obtain the enhanced feature map F_att, which integrates spatial and channel attention information. The specific formula is: F_att = Fs + Fc. This fusion method is simple and efficient, ensuring the complementarity of the two attention mechanisms.

[0034] Finally, the weighted fused feature map F_att is input into the deformation field prediction layer (usually composed of a series of convolutional layers and upsampling layers). This layer decodes the refined depth features and predicts the dense displacement vector field from the moving image I_pre to the fixed image I_intra, i.e., the final deformation field Φ, pixel by pixel.

[0035] Through this dual-attention parallel computation and feature fusion mechanism based on pure CNN, the network can adaptively focus on spatial regions and semantic channels that are crucial for registration without additional annotations, significantly improving its ability to model cross-modal differences and complex local deformations. At the same time, this design avoids the computational overhead of complex structures such as Transformers, resulting in higher computational efficiency and better meeting the stringent speed and stability requirements of real-time intraoperative image processing.

[0036] In a preferred embodiment of the present invention, the deformation field is constrained by an adaptive multi-scale regularization loss function, the specific formula of which is expressed as: L_reg=λ1 * L_smooth+λ2 * L_Jacobian L_smooth=∑ p∈Ω ( || Φ(p)||2) 2 L_Jacobian=∑ p∈Ω φ( |J Φ (p)| ) / |Ω| Where L_reg is the regularization loss function, L_smooth is the global smoothing loss, L_Jacobian is the local Jacobian determinant penalty, λ1 and λ2 are adjustable weight coefficients, p is a spatial coordinate in the image spatial domain Ω, and φ is an adaptive penalty function based on the Jacobian determinant value. Let J be the Laplace operator, Φ be the deformation field, and J be the deformation field. Φ Let be the determinant of the Jacobian matrix of the deformation field; The adaptive penalty function φ satisfies the following condition: When |J Φ When (p)| ≥ 0, φ( |J Φ (p)| )= exp(-(|J Φ (p)| - 1)²); When |J Φ When (p)| < 0, φ( |J Φ (p)| )= exp( |J Φ (p)| ) + α; Where α is a hyperparameter used to strongly penalize folding (negative Jacobian determinant).

[0037] Specifically, surgical procedures (such as tissue resection, traction, and cerebrospinal fluid loss) can cause significant, non-uniform displacement of local brain tissue (i.e., "brain drift"), while surrounding unaffected tissue remains relatively stable. This requires the deformation field Φ generated by the registration network to remain smooth and continuous on a macroscopic scale (consistent with the physical characteristics of biological tissues), while also flexibly capturing complex, large-scale nonlinear deformations in the locally affected areas. Traditional single regularization terms (such as first-order or second-order gradient smoothing) often excessively suppress necessary local deformation degrees of freedom while forcing global smoothness, leading to insufficient registration of "brain drift" regions. Based on this, in this embodiment of the invention, an adaptive multi-scale regularization loss function L_reg is designed. This function consists of a global smoothing loss L_smooth and a local Jacobian determinant penalty L_Jacobian, which are balanced by adjustable coefficients λ1 and λ2.

[0038] Specifically, the global smoothing loss L_smooth applies the Laplacian operator (second-order gradient) to the deformation field Φ and calculates the square of its L2 norm, penalizing high-frequency oscillations and non-physical drastic changes in the deformation field across the entire image domain Ω, thereby ensuring the overall smoothness and topological rationality of the deformation field. This is crucial for matching stable anatomical structures that have not been affected by surgery.

[0039] The local Jacobian determinant penalty L_Jacobian is specifically designed for finer control of the local characteristics of the deformation field. Its core lies in a carefully designed adaptive penalty function φ, which applies a penalty to the Jacobian determinant J of the deformation field. Φ The value will be used to generate a differentiated response, specifically including: When |J Φ When (p)| ≥ 0, φ( |J Φ (p)| )= exp(-(|J Φ (p)| - 1) 2 This function is in |J Φ The penalty is minimized when (p)|=1 (i.e., local volume is preserved), and as |J Φ (p)| Deviation from 1 (volume expansion or contraction) results in a gradual increase in penalty, which allows for reasonable volume changes in areas affected by surgery, such as around tumor resection cavities.

[0040] When |J Φ When (p)| < 0, φ( |J Φ (p)| )= exp( |J Φ(p)| ) + α, where α is a large hyperparameter, for example, set to 10. When the Jacobian determinant is negative, it means that the deformation field has folded locally, violating physical reversibility and topology preservation. This design causes the penalty value to increase sharply (mainly dominated by the α term) once a folding tendency is detected, thereby strongly suppressing the generation of this non-physical deformation and ensuring the anatomy of the registration result is reasonable.

[0041] Through this combined design, the regularization loss function achieves multi-scale constraints: L_smooth maintains the smoothness and continuity of the deformation field at the global scale, while L_Jacobian intelligently manages volume changes and strictly prohibits folding at the local pixel scale. This allows the network to grant sufficient degrees of freedom for deformation around the surgical area during training, while maintaining the coherence of the overall anatomical structure, to accurately compensate for complex local displacements caused by traction, resection, etc. Ultimately, this regularization scheme helps generate a physically reliable deformation field that is both smooth and flexible, capable of aligning stable tissues and accurately capturing "brain drift," thereby significantly improving registration accuracy and robustness for intraoperative scenarios.

[0042] In a preferred embodiment of the present invention, the reconstruction network is a deep learning network based on a dual encoder-single decoder architecture, and the reconstruction network includes a high-field encoder, a low-field encoder, a feature fusion module, and a decoder. The feature fusion module is a multi-level feature modulation network that includes spatial attention mechanism and channel attention mechanism.

[0043] Specifically, considering that the reconstruction task requires the full integration of the preoperative high-field image I_pre_reg, which retains high tissue contrast and signal-to-noise ratio after spatial alignment, and the intraoperative low-field magnetic resonance image I_intra, which, although lower in resolution and contrast, contains the true anatomical location and potential pathological changes of the current surgical scene, directly performing simple feature stitching or addition makes it difficult to achieve adaptive selection and deep fusion of information, thus limiting the improvement of the final reconstructed image quality. Based on this, referring to Figure 3 In this embodiment of the invention, an attention-guided feature fusion strategy is adopted, and a corresponding dual encoder-single decoder architecture is designed.

[0044] Specifically, the high-field encoder is dedicated to extracting clear and reliable prior anatomical features from the registered preoperative high-field image I_pre_reg, while the low-field encoder is responsible for capturing contextual features reflecting the real-time surgical status (including resections, displacements, bleeding, etc.) from the intraoperative low-field magnetic resonance image I_intra. These two encoders typically have a symmetrical structure and output multi-level feature maps in parallel at different scales.

[0045] The feature fusion module is embedded between the encoder and decoder, operating at each feature scale. At its core is a multi-level feature modulation network integrating spatial and channel attention mechanisms. The channel attention mechanism first analyzes corresponding feature channels from the two encoders. By calculating global dependencies between channels, it generates a weight vector to evaluate and strengthen feature channels that are stable in both image modalities and crucial for reconstructing high-quality anatomical details (e.g., clear sulci and gyri boundaries, white matter fiber tract orientation), while suppressing irrelevant channels that may be introduced by single-image noise or artifacts. The spatial attention mechanism operates on the channel-weighted features, learning to generate a spatial weight map that dynamically identifies and focuses on the spatial regions most critical to the current reconstruction task. For example, in tumor resection areas, the network may rely more on real-time information from I_intra to reflect the current morphology of the resection cavity; while in normal brain regions unaffected by surgery, it can rely more on the high-quality structural priors provided by the registered preoperative high-field image I_pre_reg.

[0046] This dual-attention modulation enables intelligent weighting and deep fusion of information sources. These refined multi-scale fused features are then fed into the decoder. Through a series of upsampling and convolution operations, the decoder gradually combines the fused deep semantic features with the shallow detail features retained during the encoding stage, ultimately reconstructing a high-quality target image that possesses both the clarity of high-field images and tissue contrast, while faithfully reflecting real-time intraoperative anatomical structures and pathological changes.

[0047] In a preferred embodiment of the present invention, step S5 includes: Step S51: Use the high-field encoder to perform multi-level feature encoding on the registered preoperative high-field image to obtain a preoperative high-field multi-scale feature map. Step S52: Use the low-field encoder to perform multi-level feature encoding on the intraoperative low-field magnetic resonance image to obtain an intraoperative low-field multi-scale feature map. Step S53: Based on the intraoperative low-field multi-scale feature map and the intraoperative low-field multi-scale feature map, the feature fusion module is used to perform spatial modulation and channel modulation on the preoperative high-field multi-scale feature map, and the modulated features are injected into the low-field features to obtain a fused enhanced feature map. Step S54: Upsample and reconstruct the fused enhanced feature map using the decoder to obtain the target image.

[0048] Specifically, in order to fully utilize the clear anatomical prior provided by the registered preoperative high-field image I_pre_reg, and to intelligently and efficiently fuse it with the real-time intraoperative low-field magnetic resonance image I_intra containing intraoperative scene information (such as newly emerging pathological features and signal intensity distribution), thereby synthesizing a high-definition intraoperative image with both high visual quality and real-time anatomical fidelity, refer to... Figure 3 In this embodiment of the invention, a dual encoder-single decoder reconstruction network architecture based on attention-guided modulation and feature injection is adopted.

[0049] Specifically, firstly, the high-field encoder Encoder_high takes the registered I_pre_reg as input and extracts the preoperative high-field multi-scale feature map F_pre_reg through multi-layer downsampling convolution. These feature maps encode high signal-to-noise ratio tissue contrast, fine structural edges and stable anatomical topology information at different resolutions.

[0050] Next, the low-field encoder Encoder_low takes the original intraoperative low-field magnetic resonance image I_intra as input and extracts its intraoperative low-field multi-scale feature map F_intra in parallel. These features faithfully reflect the image contrast, signal intensity, and local changes introduced by surgical interventions (such as resection, traction, and bleeding) at the current moment during the operation.

[0051] Secondly, at each corresponding feature scale, the feature fusion module operates as a precise "information modulation". Using the low-field multi-scale feature map F_intra as the basic context, it selectively modulates the preoperative high-field multi-scale feature map F_pre_reg under dual attention guidance. Then, the modulated features rich in effective information (such as clear structural details) are selectively injected into the basic low-field features to generate a fused enhanced feature map.

[0052] Finally, the decoder receives the fused and enhanced features from various scales, gradually restores the spatial resolution through a series of upsampling and convolutional layers, and integrates shallow detail information to ultimately reconstruct the target image I_recon. This target image I_recon is strictly aligned with the input intraoperative low-field MRI image I_intra in terms of spatial anatomical structure, ensuring its realism in reflecting the real-time scene. At the same time, its visual quality (such as signal-to-noise ratio, tissue contrast, and structural clarity) is significantly improved, approaching or even partially surpassing the original preoperative high-field image.

[0053] Through this "modulation-then-injection" fusion paradigm and encoding / decoding architecture, the reconstruction network can dynamically and adaptively determine where (space) and with what intensity (channel) to integrate high-field prior knowledge into low-field real-time images. This effectively suppresses noise in low-field images, enhances contrast, and restores clear anatomical details while preserving real intraoperative changes, providing surgeons with higher-quality and more reliable decision support images.

[0054] In a preferred embodiment of the present invention, step S53 includes: Step S531: Based on the preoperative high-field multi-scale feature map and the intraoperative low-field multi-scale feature map, a spatial attention mask is generated through the spatial attention mechanism, and the spatial attention mask is used to spatially modulate the preoperative high-field multi-scale feature map. Step S532: Based on the preoperative high-field multi-scale feature map and the intraoperative low-field multi-scale feature map, a channel weight vector is generated through the channel attention mechanism, and the channel weight vector is used to perform channel modulation on the preoperative high-field multi-scale feature map. Step S533: The features after spatial modulation and channel modulation are fused to obtain the modulated features; Step S534: The modulated features are injected into the intraoperative low-field multi-scale feature map to obtain the fused enhanced feature map.

[0055] Specifically, due to the complexity of the intraoperative scenario, the reliability and value of the registered preoperative high-field image I_pre_reg and intraoperative low-field MRI image I_intra differ across different spatial regions and semantic feature channels. For example, in deep brain regions not directly affected by surgery, the structural information from the high-field prior is highly reliable; while in areas directly manipulated by surgery (such as around the tumor resection cavity), the true signal from the low-field image may be more critical. Simple fusion can lead to information conflicts or redundancy. Based on this, refer to... Figure 3 In the embodiments of the invention, a parallel spatial-channel dual attention modulation mechanism is used as a feature fusion module to achieve refined, adaptive selection and enhancement of high-field prior features.

[0056] Specifically, firstly, the spatial attention mechanism takes the intraoperative low-field multi-scale feature map F_intra and the preoperative high-field multi-scale feature map F_pre_reg from the previous layer as input, and calculates a spatial attention mask of size 1×H×W through a lightweight convolutional sub-network (such as one containing 1x1 convolution and a sigmoid activation function). The value range of the mask is between [0, 1]. Its high-value regions identify spatial locations that are structurally blurred or have low contrast in the low-field image, but whose corresponding structural information is clear and reliable in the high-field prior (such as stable anatomical structures far from the surgical field).

[0057] Next, the mask is used to perform element-wise multiplication weighting on the preoperative high-field multiscale feature map F_pre_reg to achieve spatial modulation. This operation effectively enhances beneficial spatial locations in the high-field features while weakening unreliable regional information that may be affected by registration residual errors or surgical changes in local anatomy.

[0058] Correspondingly, the channel attention mechanism also takes the intraoperative low-field multi-scale feature map F_intra and the preoperative high-field multi-scale feature map F_pre_reg as inputs. After compressing spatial information through global average pooling, the relationship between channels is analyzed by a multilayer perceptron (MLP) to generate a channel weight vector of size Cx1x1, where C is the number of channels. This channel weight vector evaluates and labels the importance of different feature channels, tending to strengthen those channels that convey cross-modal stable anatomical characteristics (such as texture patterns and boundary features of specific tissues) and suppressing channels that may contain modality-specific noise or artifacts.

[0059] Secondly, the channel weight vector is used to scale the channel dimension of the preoperative high-field multi-scale feature map F_pre_reg to achieve channel modulation.

[0060] Furthermore, the two feature versions after the above spatial and channel modulation are fused, for example by element-wise addition or convolution after splicing, to obtain the comprehensive modulated high-field feature, which has been refined and enhanced based on the spatial location reliability and channel semantic importance.

[0061] Finally, a feature injection operation is performed: the modulated high-field features are combined with the original intraoperative low-field multi-scale feature map F_intra through residual connections or specific fusion (such as via a learnable gating or convolutional layer) to generate the final fused enhanced feature map. This operation ensures that the real-time contextual information provided by the low-field image is preserved as a foundation, while the refined high-field prior knowledge is intelligently added as an "enhancing patch".

[0062] Through this fine-grained dual attention modulation and injection mechanism, the network can achieve highly adaptive information fusion, ensuring that the reconstructed high-quality images not only make full use of the high-field prior's clarity advantage, but also strictly adhere to the real-time anatomical state and pathological details defined by the low-field images, ultimately outputting enhanced intraoperative images with high clinical credibility.

[0063] In a preferred embodiment of the present invention, the training process of the image processing model is divided into two stages, including: In the first training phase, the registration network in the image processing model is trained separately. In the second training phase, the registration network is fixed and the reconstruction network in the image processing model is trained. The reconstruction network is trained using a hybrid loss function, the specific formula of which is as follows: L_total=λ3*L_intensity+λ4*L_perceptual+λ5*L_anatomical_fidelity; Where L_total is the mixed loss function, L_intensity is the image intensity loss, L_perceptual is the perceptual loss, L_anatomical_fidelity is the anatomical fidelity loss, and λ3, λ4, and λ5 are all weighting coefficients that balance the various losses.

[0064] Specifically, considering that the paired data (preoperative high-field and intraoperative low-field MRI images of the same patient) required for training the image processing model is clinically expensive and limited in quantity, and that registration and reconstruction tasks are related but have different objective functions, direct end-to-end joint training could lead to a large number of parameters and complex task coupling, potentially causing the model to fail to converge or fall into suboptimal solutions. In this embodiment of the invention, a two-stage independent training strategy is employed to reduce training difficulty, improve data utilization efficiency, and ensure the stability of the performance of each sub-network.

[0065] Specifically, in the first training phase, the registration network is trained separately. The goal of this phase is to enable the registration network to accurately align the preoperative high-field MRI image I_pre to the intraoperative low-field MRI image I_intra. During training, an unsupervised loss function is used, which typically consists of a similarity metric (such as a mixture similarity metric) and the aforementioned adaptive multi-scale regularization loss L_reg. By minimizing this loss function, the network learns to predict the deformation field Φ that can accurately compensate for anatomical displacement without relying on any manually labeled anatomical landmarks. This phase focuses on optimizing the accuracy of spatial alignment and the physical plausibility of the deformation field.

[0066] In the second training phase, the parameters of the pre-trained registration network are fixed, and the reconstruction network is trained separately. In this phase, the pre-trained registration network parameters are fixed and used as a preprocessing module to generate a pre-registered high-field image I_pre_reg for each pair of training samples (I_pre, I_intra). The reconstruction network then takes (I_pre_reg, I_intra) as input and learns how to fuse the information from both to reconstruct a high-quality target image I_recon. The optimization objective of the second training phase is strictly defined by the aforementioned hybrid loss function L_total, aiming to guide the network to generate a reconstructed image I_recon that possesses both high visual quality and strict fidelity to the actual intraoperative anatomical state. Specifically: The image intensity loss L_intensity uses a weighted combination of normalized mutual information (NMI) and L1 loss to maintain consistency between I_recon and I_intra. The use of normalized mutual information (NMI) between I_recon and I_intra preserves the consistency of contrast across multiple modalities, while the L1 loss maintains the consistency of grayscale information. Generally, the weight of the L1 loss should be minimized to reduce the impact of noise.

[0067] The perceptual loss (L_perceptual) utilizes a deep network (such as VGG-19) pre-trained on a large natural image dataset (such as ImageNet) as a fixed feature extractor. The mean squared error between the reconstructed image I_recon and the registered image I_pre_reg (used as a high-quality reference) at selected network layers (set L) is calculated, and its specific formula is expressed as: L_perceptual=∑ l∈L ∥ (I_recon)− (I_pre_reg)∥2) 2 in, Let L represent the feature map of the l-th layer, and L be the selected set of feature layers. This loss forces I_recon to move closer to clear high-field priors at the high-level semantic feature level (such as texture and structural patterns), thereby significantly improving the visual clarity and naturalness of the output image.

[0068] The anatomical fidelity loss, L_anatomical_fidelity, introduces a lightweight U-Net segmentation module jointly trained with the reconstruction network. This module takes I_intra and I_recon as inputs and predicts segmentation maps of key anatomical structures (such as gray matter, white matter, and tumors) in real time. It calculates the Dice coefficient loss of I_recon and I_intra based on their predicted segmentation maps. This loss acts as an "anatomical supervisor," ensuring that the reconstruction process only enhances quality (e.g., denoising and sharpening) and never generates or erases real anatomical structures out of thin air, thus guaranteeing the clinical reliability and safety of the reconstruction results.

[0069] The weighting coefficients λ3, λ4, and λ5 are determined through grid search or validation set performance. Typically, to prioritize anatomical realism, L_anatomical_fidelity is given a higher weight (λ5) to prevent the introduction of false structures; λ4 (perceptual loss weight) is secondary to drive image quality improvement; and λ3 (intensity loss weight) is relatively low to balance noise suppression and detail preservation.

[0070] By employing this multi-objective, multi-level loss function design, combined with a two-stage training strategy, a high-performance image processing model can be stably and efficiently trained under limited data conditions, ultimately achieving reliable and high-quality reconstruction of intraoperative low-field magnetic resonance images. Furthermore, in order to fully explore the potential of the model with limited paired clinical data and further improve the reconstruction performance and robustness of the overall system, this embodiment of the invention introduces a third stage—end-to-end joint fine-tuning—after the two independent training stages are completed.

[0071] At this stage, all parameters of the registration network and the reconstruction network are released and jointly optimized. (Refer to...) Figure 2 This stage backpropagates the gradients of the image intensity loss L_intensity and the perceptual loss L_perceptual in the hybrid loss function L_total of the reconstruction network to the upstream registration network, thereby guiding the registration network to learn to generate an image that not only meets the spatial alignment geometric accuracy but also improves the estimation of the deformation field Φ.

[0072] Through this global optimization, the registration network can adjust its registration strategy. For example, it may prefer to maintain the morphological integrity or boundary sharpness of key anatomical structures during the alignment process. This makes it easier for the reconstruction network to generate I_recon with better visual quality and higher anatomical fidelity when the generated I_pre_reg is used as prior input.

[0073] It is evident that this end-to-end joint fine-tuning effectively breaks down the information barriers between the two sub-tasks of "registration" and "reconstruction," achieving task-driven adaptive optimization and thus achieving end-to-end image reconstruction results that are superior to those of the staged independent training model.

[0074] Reference Figure 4 The present invention also provides a medical registration-guided intraoperative magnetic resonance image reconstruction system for implementing the medical registration-guided intraoperative magnetic resonance image reconstruction method described above, comprising: Image acquisition module 110 is used to acquire preoperative high-field magnetic resonance images and intraoperative low-field magnetic resonance images of the same patient; The model loading module 120 is connected to the image acquisition module 110 and is used to input the preoperative high-field magnetic resonance image and the intraoperative low-field magnetic resonance image into the trained image processing model. The deformation field generation module 130 is connected to the model loading module 120 and is used to extract and transform multimodal features from the preoperative high-field magnetic resonance image and the intraoperative low-field magnetic resonance image using the registration network in the image processing model, so as to generate a deformation field from the preoperative image space to the intraoperative image space. The image registration module 140 is connected to the deformation field generation module 130 and is used to perform spatial transformation on the preoperative high-field magnetic resonance image based on the deformation field to obtain a registered preoperative high-field image that is spatially aligned with the intraoperative low-field magnetic resonance image. The image reconstruction module 150 is connected to the image registration module 140 and is used to perform feature fusion and reconstruction of the registered preoperative high-field image and the intraoperative low-field magnetic resonance image using the reconstruction network in the image processing model to obtain the target image.

[0075] In summary, this invention innovatively integrates high-quality anatomical priors from preoperative high-field images with real-time state information from intraoperative low-field images by designing an image reconstruction framework based on dual-attention unsupervised registration and conditional feature modulation fusion. Furthermore, it employs a multi-task-oriented hybrid loss function for joint optimization, thereby achieving high-quality, high-fidelity reconstruction of low-field magnetic resonance images in complex intraoperative environments.

[0076] Compared with the prior art, the present invention has the following significant advantages and advancements: The contradiction between "brain drift" and the utilization of prior information is resolved: by introducing an unsupervised registration network based on dual attention, it is possible to accurately and adaptively correct brain tissue displacement ("brain drift") caused by craniotomy, traction, resection and other operations, thereby ensuring that high-quality preoperative prior information is precisely aligned with the real-time anatomical structure during surgery, providing a reliable and positionally accurate anatomical reference for subsequent reconstruction, and fundamentally avoiding reconstruction misleading or information loss caused by spatial mismatch.

[0077] Intelligent fusion of cross-modal information was achieved: the reconstruction network adopted a dual encoder-single decoder architecture and innovatively integrated a conditional feature modulation fusion module (including spatial and channel dual attention). This module can dynamically and selectively modulate and inject clear and reliable anatomical detail information from the registered high-field images into the low-field real-time features, rather than simply splicing or adding them together. This "soft" fusion mechanism can significantly improve the signal-to-noise ratio, contrast, and detail clarity of the reconstructed images, while strictly preserving the real intraoperative pathological state reflected in the low-field images (such as tumor residue, edema, hemorrhage, etc.), effectively avoiding the generation of false structures or the erasure of real information.

[0078] A multi-level, multi-objective supervision system was constructed: a hybrid loss function was designed, incorporating image intensity loss, perceptual loss, and anatomical fidelity loss. This function simultaneously constrains the reconstruction process from three levels: pixel intensity consistency, high-level semantic feature similarity, and key anatomical structure fidelity. This ensures that the reconstruction results not only approximate the high-field image in visual quality and maintain consistency with the low-field image in grayscale distribution, but more importantly, remain absolutely faithful to the patient's actual intraoperative state in anatomical structure, greatly improving the clinical reliability and safety of the reconstruction results.

[0079] The training strategy and model robustness were optimized: a training strategy of independent training in stages followed by end-to-end joint fine-tuning was adopted. This strategy effectively reduced the model's dependence on a large amount of precisely paired data, achieving stable training with limited data. The joint fine-tuning stage further streamlined the optimization link between registration and reconstruction tasks, enabling the registration network to generate deformation fields that are more conducive to high-quality reconstruction, thereby improving the overall system performance and task synergy.

[0080] The above description is merely a preferred embodiment of the present invention and does not limit the implementation and protection scope of the present invention. Those skilled in the art should realize that any equivalent substitutions and obvious changes made based on the description and illustrations of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for intraoperative magnetic resonance imaging reconstruction based on medical registration guidance, characterized in that, include: Step S1: Acquire preoperative high-field magnetic resonance imaging (MRI) images and intraoperative low-field MRI images of the same patient. Step S2: Input the preoperative high-field magnetic resonance image and the intraoperative low-field magnetic resonance image into the trained image processing model; Step S3: Using the registration network in the image processing model, the preoperative high-field magnetic resonance image is used as the moving image and the intraoperative low-field magnetic resonance image is used as the fixed image for registration, thereby generating a deformation field from the preoperative image space to the intraoperative image space. Step S4: Based on the deformation field, perform spatial transformation on the preoperative high-field magnetic resonance image to obtain a registered preoperative high-field image that is spatially aligned with the intraoperative low-field magnetic resonance image. Step S5: Using the reconstruction network in the image processing model, feature fusion and reconstruction are performed on the registered preoperative high-field image and the intraoperative low-field magnetic resonance image to obtain the target image.

2. The method for intraoperative magnetic resonance image reconstruction based on medical registration guidance according to claim 1, characterized in that, The registration network is an unsupervised learning network, which includes a feature encoder, a dual attention module, and a deformation field prediction layer. The dual attention module includes a spatial attention mechanism and a channel attention mechanism.

3. The method for intraoperative magnetic resonance image reconstruction based on medical registration guidance according to claim 2, characterized in that, Step S3 includes: Step S31: Use the feature encoder to perform joint feature encoding on the preoperative high-field magnetic resonance image and the intraoperative low-field magnetic resonance image to generate a multi-scale joint feature map; Step S32: Use the spatial attention mechanism to perform feature processing and weighting operations on the multi-scale joint feature map to obtain the first weighted feature map; Step S33: Use the channel attention mechanism to perform feature processing and weighting operations on the multi-scale feature map to obtain a second weighted feature map; Step S34: Fuse the first weighted feature map and the second weighted feature map to obtain a weighted fused feature map; Step S35: Process the weighted fusion feature map using the deformation field prediction layer to generate the deformation field.

4. The method for intraoperative magnetic resonance image reconstruction based on medical registration guidance according to claim 1, characterized in that, The deformation field is constrained by an adaptive multi-scale regularization loss function, the specific formula of which is expressed as follows: L_reg=λ1 * L_smooth+λ2 * L_Jacobian L_smooth=∑ p∈Ω ( || Φ(p)||2) 2 L_Jacobian=∑ p∈Ω φ( |J Φ (p)| ) / |Ω| Where L_reg is the regularization loss function, L_smooth is the global smoothing loss, L_Jacobian is the local Jacobian determinant penalty, λ1 and λ2 are adjustable weight coefficients, p is a spatial coordinate in the image spatial domain Ω, and φ is an adaptive penalty function based on the Jacobian determinant value. Let J be the Laplace operator, Φ be the deformation field, and J be the deformation field. Φ Let be the determinant of the Jacobian matrix of the deformation field.

5. The method for intraoperative magnetic resonance image reconstruction based on medical registration guidance according to claim 1, characterized in that, The reconstruction network is a deep learning network based on a dual encoder-single decoder architecture, which includes a high-field encoder, a low-field encoder, a feature fusion module, and a decoder. The feature fusion module is a multi-level feature modulation network that includes spatial attention mechanism and channel attention mechanism.

6. The method for intraoperative magnetic resonance image reconstruction based on medical registration guidance according to claim 5, characterized in that, Step S5 includes: Step S51: Use the high-field encoder to perform multi-level feature encoding on the registered preoperative high-field image to obtain a preoperative high-field multi-scale feature map. Step S52: Use the low-field encoder to perform multi-level feature encoding on the intraoperative low-field magnetic resonance image to obtain an intraoperative low-field multi-scale feature map. Step S53: Based on the intraoperative low-field multi-scale feature map and the intraoperative low-field multi-scale feature map, the feature fusion module is used to perform spatial modulation and channel modulation on the preoperative high-field multi-scale feature map, and the modulated features are injected into the low-field features to obtain a fused enhanced feature map. Step S54: Upsample and reconstruct the fused enhanced feature map using the decoder to obtain the target image.

7. The intraoperative magnetic resonance image reconstruction method based on medical registration guidance according to claim 6, characterized in that, Step S53 includes: Step S531: Based on the preoperative high-field multi-scale feature map and the intraoperative low-field multi-scale feature map, a spatial attention mask is generated through the spatial attention mechanism, and the spatial attention mask is used to spatially modulate the preoperative high-field multi-scale feature map. Step S532: Based on the preoperative high-field multi-scale feature map and the intraoperative low-field multi-scale feature map, a channel weight vector is generated through the channel attention mechanism, and the channel weight vector is used to perform channel modulation on the preoperative high-field multi-scale feature map. Step S533: The features after spatial modulation and channel modulation are fused to obtain the modulated features; Step S534: The modulated features are injected into the intraoperative low-field multi-scale feature map to obtain the fused enhanced feature map.

8. The method for intraoperative magnetic resonance image reconstruction based on medical registration guidance according to claim 1, characterized in that, The training process of the image processing model is divided into two stages, including: In the first training phase, the registration network in the image processing model is trained separately. In the second training phase, the registration network is fixed and the reconstruction network in the image processing model is trained.

9. The method for intraoperative magnetic resonance image reconstruction based on medical registration guidance according to claim 8, characterized in that, The reconstruction network is trained using a hybrid loss function, the specific formula of which is as follows: L_total=λ3*L_intensity+λ4*L_perceptual+λ5*L_anatomical_fidelity; Where L_total is the mixed loss function, L_intensity is the image intensity loss, L_perceptual is the perceptual loss, L_anatomical_fidelity is the anatomical fidelity loss, and λ3, λ4, and λ5 are all weighting coefficients that balance the various losses.

10. A medical registration-guided intraoperative magnetic resonance imaging reconstruction system, characterized in that, For implementing the intraoperative magnetic resonance image reconstruction method based on medical registration as described in any one of claims 1-9, the system comprises: The image acquisition module is used to acquire preoperative high-field magnetic resonance images and intraoperative low-field magnetic resonance images of the same patient. The model loading module, connected to the image acquisition module, is used to input the preoperative high-field magnetic resonance image and the intraoperative low-field magnetic resonance image into the trained image processing model. The deformation field generation module, connected to the model loading module, is used to extract and transform multimodal features from the preoperative high-field magnetic resonance image and the intraoperative low-field magnetic resonance image using the registration network in the image processing model, thereby generating a deformation field from the preoperative image space to the intraoperative image space. The image registration module, connected to the deformation field generation module, is used to perform spatial transformation on the preoperative high-field magnetic resonance image based on the deformation field to obtain a registered preoperative high-field image that is spatially aligned with the intraoperative low-field magnetic resonance image. The image reconstruction module, connected to the image registration module, is used to perform feature fusion and reconstruction of the registered preoperative high-field image and the intraoperative low-field magnetic resonance image using the reconstruction network in the image processing model to obtain the target image.

Images