An intraoperative low-field magnetic resonance image processing method

Abstract

The application provides an intraoperative low-field magnetic resonance image processing method, and belongs to the technical field of intraoperative magnetic resonance, and comprises the following steps: after preoperative information is preprocessed, the preoperative information is input into a parameter recommendation model, and a scanning parameter recommendation combination is output; sequence acquisition magnetic resonance imaging signals are constructed to obtain intraoperative images; a spliced image of the preoperative image and a first frame of the intraoperative image is input into an anatomical change prediction model to output a significant displacement area; the preoperative image and the intraoperative image are input into a feature mapping model to output a matching feature vector pair; the significant displacement area is used as a constraint to initialize registration parameters; the registration model is optimized; parameters are optimized; and the significant displacement area and other areas are given differentiated weights to complete registration, and an intraoperative image after registration is output. The method has the beneficial effects that: the scanning efficiency and scene adaptability are improved by recommending scanning parameters based on preoperative information, and the image quality is guaranteed; and prior information is provided for dynamic registration by predicting anatomical changes, and the adaptability of registration to intraoperative anatomical changes is improved.

Description

Technical Field

[0001] This invention relates to the field of intraoperative magnetic resonance imaging (MRI) technology, and more particularly to an intraoperative low-field MRI image processing method. Background Technology

[0002] Intraoperative magnetic resonance imaging (MRI) is an important tool for improving surgical precision and safety. In clinical settings, low-field MRI equipment is typically used. Current research on this type of equipment is primarily focused on two independent directions: rapid imaging and multimodal image analysis.

[0003] However, existing low-field magnetic resonance imaging (MRI) techniques have many problems that urgently need to be solved. For example, current low-field MRI often uses scanning sequences with fixed parameters. Different diseases require different scanning sequences, and the optimal scanning parameters also vary depending on the stage of the same disease. The scanning parameters are not adapted to different surgical types and individual patient conditions, failing to meet diverse clinical needs. Some rapid scanning protocols use a fixed k-space filling strategy, failing to establish a precise matching relationship between preoperative information and parameters, resulting in unsatisfactory imaging results. To address magnetic field disturbances, simple filtering is often used, or the high signal-to-noise ratio of high-field equipment is relied upon for anti-interference, lacking effective adaptive correction methods in low-field scenarios, which seriously affects the quality and accuracy of imaging.

[0004] For example, existing technologies mostly use static registration algorithms for preoperative and intraoperative images. These algorithms are based on the assumption that there are no changes in the anatomical structure during surgery and do not consider dynamic changes such as brain tissue displacement caused by surgical operations, resulting in discrepancies between the registration results and the actual situation. During the registration process, only single image features are extracted for matching. For low-quality images acquired by rapid intraoperative scanning in low field, the registration accuracy is difficult to guarantee due to their low signal-to-noise ratio and blurred details. At the same time, existing intraoperative magnetic resonance imaging lacks real-time quantitative analysis tools that coordinate with the scanning and registration process. Key decision indicators such as residual tumor volume and resection rate need to be measured manually or analyzed offline, which cannot provide support for real-time intraoperative decision-making and limits the further development and application of intraoperative magnetic resonance imaging diagnosis and treatment technology. Summary of the Invention

[0005] To address the above technical problems, this invention provides an intraoperative low-field magnetic resonance imaging processing method.

[0006] The technical problem solved by this invention can be achieved by the following technical solutions: A method for intraoperative low-field magnetic resonance imaging processing includes: The preoperative information of the target patient is preprocessed, and the preprocessed preoperative information is input into a pre-trained parameter recommendation model to output a recommended combination of scanning parameters. A dual-gradient echo sequence is constructed based on the recommended combination of the scanning parameters. Magnetic resonance imaging signals are acquired based on the dual-gradient echo sequence to obtain intraoperative images from the magnetic resonance imaging signals. The stitched image of the first frame of the preoperative image and the intraoperative image of the target patient is input into the pre-trained anatomical change prediction model, and the significant displacement area of ​​the anatomical structure is output. The preoperative and intraoperative images are input into a pre-trained feature mapping model to output matching feature vector pairs. The registration parameters are initialized with the significant shift region as a constraint. The registration model optimizes the initialized registration parameters with the degree of matching of the feature vector pairs as a similarity measure, and assigns differential weights to the significant shift region and other regions other than the significant shift region to complete the registration and output the registered intraoperative image.

[0007] The intraoperative low-field magnetic resonance imaging processing method of the present invention includes the following preprocessing of the target patient's preoperative information: Obtain the preoperative information of the target patient, including surgical type labeling information, patient basic information, preoperative imaging information, and preoperative diagnostic report; The preoperative image information is preprocessed, and preoperative image features are extracted; Natural language processing is performed on the preoperative diagnostic report to generate diagnostic structured tags; The pre-processed preoperative information includes preoperative image features, structured surgical information tags, and patient basic information. The structured surgical information tags include surgical type labeling information and diagnostic structured tags.

[0008] The intraoperative low-field magnetic resonance imaging processing method of the present invention includes a pre-training method for the parameter recommendation model, comprising: Collect a first training dataset, which includes preoperative information of different patients, combinations of scanning parameters for intraoperative images, and imaging quality labels and parameter fit labels. A parameter recommendation model based on the fusion of random forest and convolutional neural network is constructed. It is trained on the first training dataset and optimized using a multi-task loss function. The model is converged through iterative optimization to obtain the trained parameter recommendation model. The multi-task loss function includes a parameter fit loss function and an imaging quality prediction loss function.

[0009] The intraoperative low-field magnetic resonance imaging processing method of the present invention further includes, after the recommended combination of output scanning parameters: In response to a parameter correction command input by the user through the interactive interface, the parameters in the recommended combination of scanning parameters are adjusted accordingly based on the parameter correction command. The parameters in the recommended combination of scanning parameters include at least one of sequence type, repetition time, echo time, k-space fill rate, and number of excitations.

[0010] The intraoperative low-field magnetic resonance imaging processing method of the present invention further includes, during the process of acquiring magnetic resonance imaging signals based on the dual-gradient echo sequence: Extract phase disturbance signals from navigation echoes; After adaptive filtering of the phase perturbation signal, reverse compensation is performed on the imaging echo to output a phase-corrected magnetic resonance imaging signal. The intraoperative images are obtained from phase-corrected magnetic resonance imaging signals.

[0011] The intraoperative low-field magnetic resonance imaging processing method of the present invention includes a pre-training method for the anatomical change prediction model, comprising: Construct a second training dataset, which includes preoperative images, images of anatomical structure displacement areas labeled at each stage of the operation, and corresponding surgical operation records; An improved U-Net network is constructed, comprising an input layer, an encoder, a decoder, and an output layer. The input layer is a stitched image of the first frame of the preoperative image and the intraoperative image. The encoder uses multi-layer concatenated convolutional blocks to extract multi-scale anatomical features through progressive downsampling. The decoder uses multi-layer deconvolutional blocks for feature upsampling and uses an attention layer to assign weights to the surgical region. The output layer is the displacement prediction result of the anatomical structure, and regions where the displacement prediction result exceeds a preset displacement threshold are identified as significant displacement regions. Using the second training dataset as training samples, the improved U-Net network is trained, and the model is converged through iterative optimization to obtain a trained anatomical change prediction model.

[0012] The intraoperative low-field magnetic resonance imaging processing method of the present invention, wherein inputting the preoperative image and the intraoperative image into a pre-trained feature mapping model to output a matching feature vector pair includes: Multi-dimensional feature extraction and fusion of preoperative images are performed to generate the first fused feature vector; Multi-dimensional feature extraction and fusion of intraoperative images are performed to generate a second fused feature vector; The feature mapping model is obtained by training a Siamese neural network. The input layer of the Siamese neural network consists of a preoperative image with a first quality and an intraoperative image with a second quality, wherein the second quality is lower than the first quality. The output layer of the Siamese neural network is provided with a feature adaptation layer. The feature mapping model uses a nearest neighbor retrieval algorithm to retrieve matching features from a pre-constructed feature matching dictionary based on the first fused feature vector and the second fused feature vector, and outputs the corresponding matching feature vector pairs of the preoperative image and the intraoperative image.

[0013] The intraoperative low-field magnetic resonance imaging processing method of the present invention includes the following registration steps: The mean pixel offset of the significantly displaced region in the preoperative and intraoperative images was statistically analyzed. An initial affine transformation matrix is ​​constructed based on the mean pixel offset, and the initial affine transformation matrix is ​​used as the initial registration parameter of the registration model; The registration model uses the degree of matching of feature vector pairs as a similarity measure and iteratively optimizes the initial registration parameters using the gradient descent method. During the parameter optimization process, a first weight is assigned to the significant shift region, and a second weight is assigned to other regions except the significant shift region, wherein the second weight is less than the first weight. Calculate the structural similarity index for the intraoperative images after parameter optimization and registration. If the structural similarity index exceeds the preset similarity threshold, output the registered intraoperative images; otherwise, return to the parameter optimization step to iterate again until the structural similarity index exceeds the preset similarity threshold.

[0014] The intraoperative low-field magnetic resonance imaging processing method of the present invention further includes, after outputting the registered intraoperative image: The intraoperative images are input into a pre-trained feature enhancement model, and the enhanced intraoperative images are output. The enhanced intraoperative images are input into a pre-trained lesion recognition model, and a binary mask image containing the residual lesion area is output. Based on the binary mask image containing the residual lesion area, the residual lesion volume and the lesion volume resection rate are calculated.

[0015] The intraoperative low-field magnetic resonance imaging processing method of the present invention further includes, after outputting the registered intraoperative image: Based on a pre-constructed multi-dimensional quality assessment index system, the comprehensive quality score of the intraoperative images is determined; wherein, the quality assessment index in the multi-dimensional quality assessment index system includes at least one of lesion boundary clarity, lesion residual volume calculation error, and image signal-to-noise ratio; When the overall quality score is lower than the preset quality threshold, the recommended combination of scanning parameters is fine-tuned.

[0016] The advantages or beneficial effects of the technical solution of this invention are as follows: This invention recommends scanning parameters based on preoperative information using a parameter recommendation model, improving scanning efficiency and scene adaptability while ensuring image quality and meeting the needs of rapid intraoperative imaging. It also uses adaptive phase correction based on dual-gradient echo sequences to suppress magnetic field disturbances in low-field, unshielded environments, enhancing anti-interference capabilities. Furthermore, it provides prior information for dynamic registration through anatomical change prediction, improving the adaptability of registration to intraoperative anatomical changes, especially achieving good registration results for low-quality intraoperative images acquired through rapid low-field scanning. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the intraoperative low-field magnetic resonance imaging processing method in a preferred embodiment of the present invention. Figure 2 This is a schematic diagram of the preoperative information preprocessing process in a preferred embodiment of the present invention; Figure 3 This is a schematic diagram of the parameter recommendation model pre-training process in a preferred embodiment of the present invention. Figure 4 This is a schematic diagram of the adaptive phase correction process based on a dual-gradient echo sequence in a preferred embodiment of the present invention. Figure 5 This is a schematic diagram of the pre-training process of the anatomical change prediction model in a preferred embodiment of the present invention. Figure 6 This is a schematic diagram of the feature mapping process in a preferred embodiment of the present invention; Figure 7 This is a schematic diagram of the registration process in a preferred embodiment of the present invention; Figure 8 This is a flowchart illustrating the quantitative analysis of residual lesion volume and lesion volume resection rate in a preferred embodiment of the present invention. Figure 9 This is a schematic diagram of the process for quality assessment and fine-tuning of scanning parameters in a preferred embodiment of the present invention. Detailed Implementation

[0018] This invention aims to address the technical problems existing in current low-field unshielded intraoperative MRI technology, such as insufficient scanning efficiency and scene adaptability, low multimodal registration accuracy, lack of real-time quantitative analysis and closed-loop optimization capabilities, and insufficient anti-interference capabilities. It provides an intraoperative low-field magnetic resonance imaging processing method that achieves precise adaptation of scanning parameters through artificial intelligence-assisted parameter recommendation based on preoperative information, and ensures imaging quality by combining magnetic field disturbance adaptive correction. It realizes the full-process collaboration of intraoperative dynamic scene adaptive rapid scanning, multimodal precise registration and fusion, real-time quantitative analysis, and parameter optimization, while ensuring intraoperative imaging efficiency and decision accuracy, and improving the level of intraoperative diagnosis and treatment integration.

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

[0020] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other.

[0021] 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.

[0022] In a preferred embodiment of the present invention, based on the above-mentioned problems existing in the prior art, a method for intraoperative low-field magnetic resonance imaging processing is provided, such as... Figure 1 As shown, it includes: S1, preprocess the preoperative information of the target patient, input the preprocessed preoperative information into the pre-trained parameter recommendation model, and output the recommended combination of scanning parameters; S2, construct a dual gradient echo (GRE) sequence based on the recommended combination of scanning parameters, acquire magnetic resonance imaging signals based on the dual gradient echo sequence, and obtain intraoperative images based on the magnetic resonance imaging signals; S3, input the stitched image of the first frame of the preoperative and intraoperative images of the target patient into the pre-trained anatomical change prediction model, and output the significant displacement area of ​​the anatomical structure; S4. Input the preoperative and intraoperative images into the pre-trained feature mapping model to output matching feature vector pairs. Initialize the registration parameters with the significant shift region as a constraint. The registration model optimizes the initialized registration parameters with the degree of matching of the feature vector pairs as a similarity measure. Differentiated weights are assigned to the significant shift region and other regions except the significant shift region to complete the registration and output the registered intraoperative image.

[0023] Specifically, to address the issues of fixed scanning parameters and poor adaptability in existing low-field magnetic resonance imaging (MRI) technology, this invention uses artificial intelligence-assisted personalized recommendation and adaptation of scanning parameters based on preoperative information to shorten scanning time while ensuring the imaging quality of low-field MRI during surgery. This meets the need for rapid intraoperative imaging and reduces surgical procedure delays and patient anesthesia risks caused by excessively long scanning times.

[0024] To address the limitations of existing static registration methods in adapting to dynamic changes in intraoperative anatomical structures and the low registration accuracy of low-quality intraoperative images, this invention employs an anatomical change prediction model to predict intraoperative anatomical changes, providing prior information for subsequent dynamic registration. This model is suitable for low-quality intraoperative images acquired through rapid low-field scanning. Simultaneously, a feature mapping model enables precise multi-dimensional feature matching between preoperative and intraoperative images. Using the matching degree of feature vector pairs as a similarity metric, and combined with a registration parameter optimization strategy based on differential weights, this approach is suitable for low-quality intraoperative images acquired through rapid low-field scanning, improving registration accuracy and robustness, and providing a foundation for image fusion in subsequent quantitative analysis.

[0025] The intraoperative low-field magnetic resonance imaging processing method of the present invention, such as Figure 2 As shown, preprocessing of the target patient's preoperative information includes: S101, Obtain the preoperative information of the target patient, including surgical type labeling information, patient basic information, preoperative imaging information, and preoperative diagnostic report; In this embodiment, preoperative structured surgical information can be obtained through a Standard Positioning System (SPS) or preoperative imaging data can be obtained through a Picture Archiving and Communication System (PACS). Specifically, the surgical type labeling includes, but is not limited to, neurosurgical craniotomy and laser interstitial thermotherapy (LITT); the patient's basic information includes, but is not limited to, age, weight, and skull thickness; the preoperative imaging information can be preoperative multimodal imaging data, such as gradient echo images, T1-weighted images, or T2-weighted images; the preoperative diagnostic report includes, but is not limited to, key information such as lesion location, size, and type.

[0026] S102, preprocess the preoperative image information and extract the preoperative image features; In this embodiment, preprocessing includes, but is not limited to, normalization, voxel resampling, and region segmentation. Specifically, the pixel values ​​of the preoperative images are normalized to map them to the [0,1] range, ensuring data consistency and comparability. The normalized preoperative images are then resampled to a preset voxel size, which in this embodiment is preferably 1mm×1mm×1mm to improve image resolution and accuracy. The resampled images are then subjected to region identification and segmentation to remove the skull region and eliminate unnecessary interference factors.

[0027] After preprocessing, preoperative image features are extracted, including but not limited to lesion grayscale features and brain tissue anatomical features.

[0028] S103, perform natural language processing on the preoperative diagnostic report to generate diagnostic structured tags; In this embodiment, a Natural Language Processing (NLP) algorithm is used to extract keywords from the unstructured information in the preoperative diagnostic report, transforming the unstructured text information into diagnostic structured tags. For example, taking tumor surgery as an example, diagnostic structured tags may include, but are not limited to, deep tumors, cystic tumors, and small-volume tumors.

[0029] The pre-processed preoperative information includes preoperative imaging features, structured surgical information tags, and patient basic information. The structured surgical information tags include surgical type labeling information and diagnostic structured tags.

[0030] Furthermore, this invention adopts a parameter recommendation model architecture based on feature fusion and multi-task learning. The input is preprocessed preoperative image features, structured surgical information labels, and patient basic information. The output is the recommended combination of the best scanning parameters for the current case. The scanning parameters in the combination include, but are not limited to, sequence type, repetition time (TR), echo time (TE), k-space fill rate, and number of excitations.

[0031] The intraoperative low-field magnetic resonance imaging processing method of the present invention, such as Figure 3 As shown, the pre-training methods for parameter recommendation models include: S111, Collect the first training dataset, which includes preoperative information of different patients, scanning parameter combinations of intraoperative images, and imaging quality labels and parameter fit labels. S112, Construct a parameter recommendation model based on the fusion of two branches of Random Forest and Convolutional Neural Network (CNN). Train it on the first training dataset and use a multi-task loss function for training optimization. Iterate and optimize the model until it converges to obtain the trained parameter recommendation model. The multi-task loss function includes a parameter fit loss function and an imaging quality prediction loss function.

[0032] Specifically, in this embodiment, during the training process of the parameter recommendation model, firstly, several complete preoperative-intraoperative matching case data are collected as the model training dataset. To solve the problem that a single model cannot simultaneously process images and structured features, and the model is redundant, this embodiment of the invention adopts a dual-branch hybrid model architecture that integrates random forest and lightweight 3D convolutional neural network. The lightweight 3D convolutional neural network is used to process preoperative image features, and the random forest is used to process structured surgical information labels and patient basic information. After attention fusion, two types of results are output: parameter fit and image quality prediction.

[0033] The recommended combination of scanning parameters output by the model can be the candidate parameter combination with the highest parameter fit score that has qualified image quality prediction in the candidate parameter library; or it can be multiple candidate parameter combinations with qualified image quality prediction and high parameter fit scores.

[0034] The parameter recommendation model is trained and optimized using a multi-task loss function, which is a weighted sum of parameter fit loss and image quality prediction loss. Specifically, the parameter fit loss function uses a weighted mean squared error (MSE) loss function, while the image quality prediction loss function uses a binary cross-entropy loss function specific to low-field magnetic resonance imaging (MRI) equipment. The multi-task loss function is as follows: L total =W1·L adapt +W2·L quality ; Among them, L adapt This represents the mean squared error loss with parameterized weights to ensure the accuracy of adaptation for core parameters such as K-space fill rate; L quality This refers to the binary cross-entropy loss for low-field magnetic resonance imaging, used to quickly determine whether the image quality is acceptable or unacceptable, to meet the needs of intraoperative assessment; L total This represents the total loss function.

[0035] The optimizer used is AdamW, with an initial learning rate of 1e-4, which decays by 0.5 every 10 epochs. The number of training iterations is set to 100 epochs, where one epoch represents the process of the model traversing the entire training dataset once. The batch size is 8 to ensure that the model can fully learn the information in the dataset.

[0036] After training, the model is quantized and optimized. For example, optimization can be achieved through Quantization-Aware Training (QAT) and operator fusion. Specifically, firstly, the trained model is simplified by removing redundant layers to reduce its complexity; then, a calibration set covering major clinical scenarios is constructed; next, 8-bit integer (INT8) symmetric quantization is performed using PyTorch, and convolution operators, batch normalization (BN) operators, and ReLU operators are fused to further improve the model's running speed; finally, the optimized model is deployed to the embedded engine of the MRI machine for validation. Model quantization optimization compresses the inference time to less than 20ms, ensuring rapid preoperative parameter recommendation.

[0037] The intraoperative low-field magnetic resonance imaging processing method of the present invention, after outputting the recommended combination of scanning parameters, further includes: In response to a parameter correction command input by the user through the interactive interface, the corresponding adjustment operation is performed on the parameters in the recommended combination of scanning parameters based on the parameter correction command. The parameters in the recommended combination of scanning parameters include at least one of sequence type, repetition time, echo time, k-space fill rate, and number of excitations.

[0038] Specifically, after the parameter recommendation model outputs the initial recommended combination of scanning parameters, it supports manual fine-tuning within the recommended parameter range of each scanning parameter and generates parameter correction instructions to obtain the final recommended combination of scanning parameters.

[0039] The method of the present invention can be applied to an intraoperative low-field magnetic resonance imaging system. The system establishes real-time communication with the scanning control module of the magnetic resonance imaging device through the SDK interface, and includes the recommended combination or fine-tuned combination of scanning parameters output by the model in the parameter command and sends it to the magnetic resonance imaging device to achieve accurate adaptation of preoperative parameters.

[0040] This invention uses an artificial intelligence model to deeply mine the matching relationship between preoperative information and scanning parameters, and realizes personalized optimal scanning parameter recommendations, thereby minimizing scanning time while ensuring image quality.

[0041] The intraoperative low-field magnetic resonance imaging processing method of the present invention, based on the process of acquiring magnetic resonance imaging signals using a dual-gradient echo sequence, such as... Figure 4As shown, it also includes: S201, extract phase disturbance signal based on navigation echo; S202 performs adaptive filtering on the phase perturbation signal, performs reverse compensation on the imaging echo, and outputs a phase-corrected magnetic resonance imaging signal. S203, intraoperative images are obtained based on phase-corrected magnetic resonance imaging signals.

[0042] Specifically, to further improve registration and analysis accuracy, this invention optimizes the fast scanning sequence and adopts a magnetic field disturbance adaptive phase correction algorithm based on a dual-gradient echo sequence. This enhances the resistance of fast scanning images to magnetic field disturbances in low-field unshielded environments, reduces artifacts caused by magnetic field disturbances, and provides high-quality image data for subsequent registration and analysis. It also simplifies the anti-interference algorithm architecture and reduces system computational overhead.

[0043] The steps of the adaptive phase correction algorithm for magnetic field disturbance are as follows: First, based on the dual-gradient echo sequence framework, signal acquisition is performed synchronously in two echo acquisition stages, namely the first echo time TE1 and the second echo time TE2. Then, the phase difference calculation model is used: Δφ=φ2-φ1 Where φ1 represents the phase of the acquired signal corresponding to the first echo time TE1; φ2 represents the phase of the acquired signal corresponding to the second echo time TE2; and Δφ represents the phase difference.

[0044] Next, the magnetic field disturbance offset ΔB is fitted using the least squares method: ΔB=Δφ / (γ·(TE2-TE1)) Where γ represents the proton gyromagnetic ratio, γ = 42.58 MHz / T.

[0045] Finally, based on the constructed adaptive phase correction filter, the acquired k-space data is phase corrected to eliminate phase distortion caused by magnetic field disturbance.

[0046] Current low-field magnetic resonance imaging (MRI) magnetic field correction uses fixed-parameter filtering, such as a single Gaussian filter or a fixed-order infinite impulse response (IIR) filter, without adaptive design incorporating the phase characteristics of the dual-gradient echo sequence. The filter of this invention, however, is designed for low-field, unshielded intraoperative scenarios and integrates phase unwinding and adaptive filtering strategies.

[0047] The working principle of this adaptive phase correction filter is as follows: Data separation: Separate imaging echo and navigation echo from the k-space data of the dual-gradient echo sequence; Phase extraction: Perform inverse Fourier transform on the navigation echo, calculate its phase angle and remove the mean to obtain the original phase offset caused by magnetic field disturbance, i.e. the original phase difference Δφ caused by magnetic field disturbance. Adaptive filtering: By default, a 4th-order Butterworth low-pass infinite impulse response filter is used for zero-phase filtering with a cutoff frequency of 0.5; or it can be switched to an 11th-order Gaussian filter with a Gaussian kernel standard deviation σ=2 to filter out phase noise. Phase correction: The filtered phase is unwound to generate a phase compensation factor, and phase weighting compensation is performed on the k-space data of the imaging echo dimension by dimension to eliminate phase distortion caused by magnetic field disturbance.

[0048] To ensure that the real-time performance of rapid scanning is not affected, this invention embeds the magnetic field disturbance adaptive phase correction algorithm into the signal acquisition link of the dual-gradient echo sequence. The phase disturbance is extracted through the navigation echo, and the imaging echo is compensated in reverse after adaptive filtering, so that the correction processing delay is controlled within 50ms.

[0049] The intraoperative low-field magnetic resonance imaging processing method of the present invention, such as Figure 5 As shown, the pre-training methods for the anatomical change prediction model include: S301, Construct the second training dataset, which includes preoperative images, images with labels of anatomical structure displacement areas at each stage of the operation, and corresponding surgical operation records. S302, construct an improved U-Net network. The improved U-Net network includes an input layer, an encoder, a decoder, and an output layer. The input layer is a stitched image of the first frame of the preoperative and intraoperative images. The encoder uses multi-layer concatenated convolutional blocks to extract multi-scale anatomical features through progressive downsampling. The decoder uses multi-layer deconvolutional blocks for feature upsampling and uses an attention layer to assign weights to the surgical area. The output layer is the displacement prediction result of the anatomical structure. Areas where the displacement prediction result exceeds a preset displacement threshold are identified as significant displacement areas. S303: Using the second training dataset as training samples, the improved U-Net network is trained, and the model is converged through iterative optimization to obtain a well-trained anatomical change prediction model.

[0050] Specifically, to overcome the limitations of existing static registration, this embodiment constructs an anatomical change prediction model for predicting intraoperative anatomical structure displacement, providing prior information for dynamic registration and improving the adaptability of registration to intraoperative anatomical changes.

[0051] The process of constructing the anatomical change prediction model is as follows: First, we collected several cases of continuous scanning data, including preoperative images, intraoperative MRI images at different stages, and corresponding surgical operation records. Then, the preoperative and intraoperative images are preprocessed. The preprocessing includes, but is not limited to: image normalization, mapping pixel values ​​to the [0,1] value range; resampling the normalized image to a preset voxel size, which in this embodiment is preferably 1mm×1mm×1mm; performing region identification and segmentation on the resampled image to remove the skull region; and marking the anatomical structure displacement region, such as brain tissue and tumor boundaries, as labels for the anatomical structure displacement region.

[0052] Next, an improved U-Net network is used to stitch preoperative images with the first intraoperative frame image to form an input image with 6 channels. The encoder uses 4 convolutional blocks, each containing two 3×3×3 convolutional layers, a batch normalization (BN) layer, and a ReLU activation function, and downsampling with a stride of 2 to extract multi-scale anatomical features. The decoder uses 4 deconvolutional blocks, each containing one 2×2×2 deconvolutional layer and a feature fusion layer, and employs a Convolutional Block Attention Module (CBAM) to enhance the weights of features in the surgical region (lesion region). The output layer outputs a 3-channel anatomical displacement prediction map, corresponding to the displacement along the x, y, and z axes. This improved U-Net uses multi-channel input, a CBAM attention module, and region-weighted loss to adapt to intraoperative anatomical displacement prediction.

[0053] Next, the improved U-Net network was trained and optimized using the second training dataset as training samples. Specifically, the mean squared error (MSE) loss function was adopted, and loss weights were set according to the importance of different regions; for example, the loss weight for lesion regions was 3, and the loss weight for normal brain tissue was 1. The optimizer AdamW was selected, with an initial learning rate of 1e-4, decreasing by 0.5 every 10 epochs, and the number of training iterations was set to 100 epochs with a batch size of 8. After training, model quantization (INT8) was used for optimization, compressing the prediction inference time to less than 30ms to meet the requirements of real-time intraoperative processing.

[0054] During the model application phase, preoperative images and the first intraoperative image are used as input to the model, and the model outputs the predicted displacement results of anatomical structures. To present these prediction results more intuitively and clearly, a graphical representation can be used to create a heatmap of the predicted displacement of anatomical structures. In the heatmap, different colors represent different degrees of anatomical structure displacement; darker colors indicate greater displacement, while lighter colors indicate smaller displacement.

[0055] Furthermore, the displacement prediction result is compared with a preset displacement threshold. If the displacement prediction result exceeds the preset displacement threshold, the region is identified as a significantly displaced region; if the displacement prediction result does not exceed the preset displacement threshold, the region is identified as another normal tissue region other than a significantly displaced region. For example, the preset displacement threshold can preferably be set to 2 mm.

[0056] The intraoperative low-field magnetic resonance imaging processing method of the present invention, such as Figure 6 As shown, inputting preoperative and intraoperative images into a pre-trained feature mapping model to output matching feature vector pairs includes: S401, perform multi-dimensional feature extraction and fusion on preoperative images to generate the first fused feature vector; Specifically, firstly, multi-dimensional feature extraction is performed on preoperative images to obtain the first structural feature, the first texture feature, and the first intensity feature. The structural feature uses a 3D Sobel operator with a 3×3×3 convolution kernel to extract edge information. The texture feature uses a gray-level co-occurrence matrix with a distance of 1 and angles of 0°, 45°, 90°, and 135°, and calculates three parameters: energy, entropy, and contrast. The intensity feature uses an image gray-level histogram with 64 bins. Then, L2 normalization is performed on each of the three feature classes, and principal component analysis (PCA) is used for dimensionality reduction to retain 95% of the information, constructing a first fusion feature vector with a dimension of 256. This first fusion feature vector contains the first structural feature, the first texture feature, and the first intensity feature.

[0057] S402, Multi-dimensional feature extraction and fusion of intraoperative images are performed to generate a second fused feature vector; Specifically, multi-dimensional feature extraction is performed on the intraoperative images to obtain second structural features, second texture features, and second intensity features. The feature extraction and fusion process is the same as that for preoperative images and will not be repeated here. In this embodiment, the second fused feature vector is a fused feature vector containing the second structural features, second texture features, and second intensity features.

[0058] S403 is a feature mapping model trained based on a Siamese neural network. The input layer of the Siamese neural network consists of a preoperative image with a first quality and an intraoperative image with a second quality, where the second quality is lower than the first quality. The output layer of the Siamese neural network has a feature adaptation layer. The feature mapping model uses a nearest neighbor retrieval algorithm to retrieve matching features from a pre-built feature matching dictionary based on the first and second fused feature vectors, and outputs the corresponding matching feature vector pairs between the preoperative and intraoperative images.

[0059] Specifically, a Siamese neural network is used to train the feature mapping model. The network input consists of high-quality preoperative images and low-quality intraoperative images, with further optimization of input voxel sizes to 16mm×16mm×16mm image blocks. The network output is the matched feature vector pairs. After training with a large number of image blocks, a feature matching dictionary is constructed, and a nearest neighbor retrieval algorithm, such as the KD-tree algorithm, is used to optimize the matching retrieval speed, compressing the retrieval time to ≤5ms / feature vector.

[0060] Compared to high-field / long-duration preoperative scans, which have a signal-to-noise ratio (SNR) ≥ 50 dB, intraoperative images acquired by unshielded rapid low-field scans, after correction for magnetic field disturbances, have an SNR ≤ 30 dB. Image artifacts are reduced, and image quality is relatively improved. However, image details are blurred, such as the loss of key information like lesion boundaries and brain tissue texture. Furthermore, there are slight artifacts caused by magnetic field disturbances. In other words, these intraoperative images are still considered low-quality images.

[0061] Existing Siamese neural networks are mostly used for general image matching, and their inputs are usually image patches of the same quality. To adapt to the registration requirements of the high-quality preoperative image patches and optimized low-quality intraoperative images of this invention, this invention adds a feature adaptation layer to the output layer of the Siamese neural network. This feature adaptation layer can solve the problem of feature matching misalignment between high-quality and low-quality images, ensuring feature matching between images of different quality.

[0062] The intraoperative low-field magnetic resonance imaging processing method of the present invention, such as Figure 7 As shown, registration includes: S411, Calculate the mean pixel offset of the significantly displaced region in preoperative and intraoperative images; S412, construct an initial affine transformation matrix based on the average pixel offset, and use the initial affine transformation matrix as the initial registration parameter of the registration model; S413, the registration model uses the degree of matching of feature vector pairs as a similarity measure, and iteratively optimizes the initial registration parameters using the gradient descent method; during the parameter optimization process, a first weight is assigned to the significantly shifted region, and a second weight is assigned to other regions except the significantly shifted region, and the second weight is less than the first weight; S414: Calculate the structural similarity index for the intraoperative images after parameter optimization and registration. If the structural similarity index exceeds the preset similarity threshold, output the registered intraoperative images; otherwise, return to the parameter optimization step to iterate again until the structural similarity index exceeds the preset similarity threshold.

[0063] Specifically, to address the issue of low registration accuracy in low-quality, rapidly scanned images, and to improve registration accuracy while ensuring registration efficiency to meet real-time intraoperative requirements, this embodiment employs elastic registration. The registration parameters are initialized using an affine transformation matrix, with the displacement prediction region output by the anatomical change prediction model as a constraint. The specific steps are as follows: Input the displacement prediction heatmap output by the anatomical change prediction model and extract the coordinate range of significant displacement regions; for significant displacement regions, calculate the average pixel offset between preoperative and intraoperative images; construct an initial affine transformation matrix based on the average offset; use this initial affine transformation matrix as the initial parameter for B-spline elastic registration, initializing the registration grid only within significant displacement regions to reduce unnecessary calculations and improve registration efficiency.

[0064] The registration model of this invention adopts a B-spline elastic registration model, using the mutual information of multi-feature fusion as a similarity measure, with a window size of 9×9×9. The mutual information of multi-feature fusion refers to the fused feature vectors of structural, textural, and intensity features of preoperative and intraoperative image blocks as the calculation object, used to characterize the degree of information correlation between the two sets of fused feature vectors. This serves as the similarity measure for the B-spline elastic registration model, determining the feature matching degree of preoperative and intraoperative image blocks during the registration process.

[0065] When optimizing the registration parameters, gradient descent is used to optimize the initialized registration parameters with a learning rate of 5e-3 and 50 iterations. Simultaneously, an adaptive weight iteration strategy is introduced, assigning a higher first weight to significant displacement regions exceeding 2mm and a second weight to normal regions other than significant displacement regions, thereby improving the registration accuracy of critical surgical areas. In this embodiment, the first weight is preferably set to 1.5, and the second weight is preferably set to 1.0. However, this is not a limitation; in other embodiments, specific weight values ​​can be set as needed, and are not restricted here.

[0066] After registration is completed, the structural similarity index (SSIM) of the registered intraoperative image is calculated. The structural similarity index (SSIM) is compared with a preset similarity threshold. In this embodiment, the preset similarity threshold can be set to 0.85. That is, when SSIM≥0.85, the registration result is output; otherwise, the optimization is iterated again to ensure the registration accuracy.

[0067] The intraoperative low-field magnetic resonance imaging processing method of the present invention, such as Figure 8 As shown, after outputting the registered intraoperative images, the following steps are also included: S501 inputs intraoperative images into a pre-trained feature enhancement model and outputs enhanced intraoperative images; Specifically, the feature enhancement model employs a lightweight convolutional neural network, specifically a 3D improved version of the MobileNetV3 model. This model extends the 2D architecture of MobileNetV3 to 3D to adapt to the 3D characteristics of intraoperative low-field MRI images. Through a combination of architecture pruning and channel pruning, the model removes two redundant bottleneck blocks, one global average pooling layer, and one terminal fully connected layer from the original MobileNetV3, retaining only the core feature extraction module, thus achieving deep compression. Simultaneously, L1 regularization is used to calculate the weights of each convolutional channel, and redundant channels with weights less than a preset weight threshold (e.g., 0.01) are pruned. For example, the number of channels in some bottleneck blocks is reduced from 64 to 40, compressing the network depth to 60% of the original MobileNetV3.

[0068] A depthwise separable convolution with a 3×3×3 kernel is employed to reduce computational cost while maintaining effective feature extraction. The model input consists of low-quality intraoperative images acquired through rapid scanning. The images are processed through a four-layer feature extraction and enhancement module (including residual connections) to enhance key features such as lesion boundaries and blood vessel orientation, outputting enhanced intraoperative images that improve the signal-to-noise ratio and shorten inference time.

[0069] S502 inputs the enhanced intraoperative images into the pre-trained lesion recognition model and outputs a binary mask image containing the residual lesion area; Specifically, the lesion identification model automatically identifies lesion boundaries, determines residual lesion areas, and outputs the results visually in the form of a binary mask. The residual lesion area refers to the lesion tissue region within the intraoperative low-field magnetic resonance imaging scan range that was not removed after surgical manipulation and remains in the patient's brain tissue. This region is located within the preoperative lesion marking area and appears significantly different from normal brain tissue signal characteristics (grayscale, texture) in intraoperative images. In the binary mask output by the lesion identification model, the pixel region corresponding to the "1" value is the lesion tissue region, and the "0" value corresponds to normal tissue or the surgical cavity.

[0070] The lesion recognition model uses an improved Mask R-CNN network, with a backbone network of a deep residual network (ResNet), such as a lightweight version of ResNet50, which halves the number of channels. A detail enhancement branch for low-field images is added to the feature pyramid network (FPN) to reduce computational complexity while ensuring the extraction of lesion boundaries. Anchor boxes are initialized based on preoperative lesion annotation information. This invention sets anchor boxes of various sizes, with selectable sizes of 16×16×16, 32×32×32, and 64×64×64, to ensure accurate coverage of tumor regions of different sizes.

[0071] The lesion identification model employs the focal loss function, with a balance factor α=0.25 and a modulation factor γ=2, to address the class imbalance problem between the lesion (e.g., tumor) region and the background. The model outputs a binary mask image of the residual lesion region, with boundary localization accuracy ≤1mm.

[0072] S503, based on a binary mask image containing residual lesion areas, calculates the residual lesion volume and the lesion volume resection rate.

[0073] Specifically, based on the binary mask image of the residual lesion area, the voxel counting method is used to calculate the residual lesion volume: V 残留 = N 体素 × V 单一体素 in, V 残留 Indicates the residual volume of the lesion; N 体素 This represents the number of voxels in the residual lesion region of the binary mask image; V 单一体素 This indicates the volume of a single voxel, such as a voxel size of 1mm × 1mm × 1mm. V 单一体素 =1mm 3 .

[0074] Simultaneously, based on preoperative lesion volume data, the lesion resection rate was calculated: R 切除 =( V 术前 - V 残留 ) / V 术前 ×100% in, V 术前 This indicates the total volume of the lesion as marked preoperatively; R 切除 This indicates the lesion volume resection rate.

[0075] This invention enables real-time acquisition of key intraoperative decision indicators through real-time quantitative calculations, eliminating the need for doctors to manually measure or perform offline analysis, thus providing precise data support for surgical decisions.

[0076] The intraoperative low-field magnetic resonance imaging processing method of the present invention further includes, after outputting the registered intraoperative image: S601, based on a pre-constructed multi-dimensional quality assessment index system, determines the comprehensive quality score of intraoperative images; wherein, the quality assessment index in the multi-dimensional quality assessment index system includes at least one of lesion boundary clarity, lesion residual volume calculation error, and image signal-to-noise ratio; Specifically, a multi-dimensional quality assessment indicator system will be established, which will consist of: Lesion boundary clarity: This indicator is measured by the mean of the edge gradient amplitude; the threshold for this indicator is set to ≥20. When the mean of the edge gradient amplitude reaches or exceeds this threshold, the clarity of the lesion boundary is considered to meet the requirements.

[0077] Lesion residual volume calculation error: This indicator reflects the degree of deviation between the residual volume calculated by the algorithm and the actual residual volume. The lesion residual volume calculation error is calculated using the gold standard comparison method, and the specific formula is as follows:

[0078] in, E 体积 This indicates the error in calculating the residual volume of the lesion; V 1 indicates that the residual volume calculated by the algorithm is the same as the one obtained above. V 残留 ; V 2 indicates the residual volume obtained from manually annotated gold standards. The manual annotation of the gold standard involves two associate chief physicians or above manually outlining the residual area and calculating the volume, with the average value used as the benchmark. The threshold for this indicator is set to ≤5%, meaning the residual volume error calculated by the algorithm in this invention is required to be within acceptable limits. E 体积 Keep it within the range of ≤5%.

[0079] Image signal-to-noise ratio (SNR): Image SNR is an important indicator of image quality, reflecting the ratio of signal to noise in an image. The threshold for this indicator is set to ≥30dB to ensure that the processed image has sufficiently high quality.

[0080] To comprehensively consider the impact of each indicator on the overall quality, the analytic hierarchy process (AHP) was used to determine the weights of each indicator. As a specific, but not limiting, weight allocation is as follows: 40% for lesion boundary clarity, 35% for lesion residual volume calculation error, and 25% for image signal-to-noise ratio.

[0081] Based on the above indicators and weights, the score of each indicator is multiplied by its corresponding weight, and then the weighted scores of all indicators are summed to obtain the comprehensive quality score.

[0082] S602, when the overall quality score is lower than the preset quality threshold, performs parameter fine-tuning on the recommended combination of scanning parameters.

[0083] Specifically, when the overall quality score exceeds a preset quality threshold, such as 80 points or above, the image quality is considered acceptable. When the overall quality score falls below the preset quality threshold, a parameter fine-tuning strategy is triggered, making targeted adjustments based on the specific circumstances of different indicators, or manually triggering scan parameter optimization according to clinical needs.

[0084] The specific parameter fine-tuning strategy is as follows: When the clarity of the lesion boundary is not up to standard, increase the k-space fill rate by 10%-20%. Increasing the k-space fill rate can increase the high-frequency information of the image, making the lesion boundary clearer; when the error in calculating the residual volume of the lesion is not up to standard, adjust the echo time of the scanning sequence by ±2ms; when the image signal-to-noise ratio is not up to standard, enable the signal averaging frequency enhancement function of the sequence, increasing the signal averaging frequency from 1 to 2.

[0085] After completing parameter fine-tuning, the scan-registration-analysis process is re-executed for another quality assessment. If the overall quality score remains below the preset quality threshold after a preset number of fine-tuning attempts (e.g., 3 times), an alarm mechanism is triggered, sending an audible and visual alarm to the surgeon's workstation and providing an interface for manually adjusting parameters to ensure the reliability and safety of closed-loop optimization.

[0086] This invention can ensure a dynamic balance between intraoperative imaging quality and decision-making accuracy, avoiding decision-making errors caused by image quality issues.

[0087] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms to form an intraoperative low-field magnetic resonance imaging processing system. This system includes three main modules: a dynamic scene adaptive fast scanning module, a multimodal dynamic registration and fusion module, and a real-time quantitative analysis and closed-loop optimization module. The dynamic scene adaptive fast scanning module is used to execute steps S1-S2, the multimodal dynamic registration and fusion module is used to execute steps S3-S4, and the real-time quantitative analysis and closed-loop optimization module is used to execute steps S501-S503 and S601-S602. The real-time quantitative analysis and closed-loop optimization module also uses a message queue mechanism (MQTT protocol) to push the quantitative analysis results of residual lesion volume and lesion volume resection rate, along with the comprehensive quality score, to the dynamic scene adaptive fast scanning module in real time, with a push delay ≤10ms, achieving real-time feedback and parameter fine-tuning.

[0088] By adopting the above solution, this invention achieves optimal scanning parameter matching for different surgical types and magnetic field environments through dynamic scene adaptive scanning parameter adjustment, shortens scanning time, and ensures image quality, meeting the needs of rapid intraoperative imaging; it solves the problems of fixed scanning parameters and poor adaptability in existing technologies, improves scanning efficiency and scene adaptability, and reduces surgical procedure delays and patient anesthesia risks caused by excessively long scanning times.

[0089] This invention, through an anatomical change prediction model and a multi-feature fusion registration algorithm, can accurately adapt to low-quality images acquired by rapid low-field scanning. It solves the problems of existing static registration being unable to adapt to intraoperative anatomical changes and low registration accuracy of low-quality images, improves the accuracy of multimodal registration, and provides a precise image fusion foundation for subsequent quantitative analysis.

[0090] This invention achieves real-time quantitative analysis and closed-loop optimization. The quantitative analysis can output key indicators such as residual lesion volume and resection rate in real time without time delay. The closed-loop optimization mechanism can automatically adjust scanning parameters according to the analysis results to ensure that the imaging quality always meets the decision-making requirements. This solves the problems of offline quantitative analysis and inability to reverse optimize scanning parameters in existing technologies, improves the accuracy and timeliness of surgical decisions, and reduces the risk of tumor residue or over-resection due to decision-making errors.

[0091] This invention presents an adaptive correction algorithm for magnetic field disturbances based on dual-echo technology. This algorithm can effectively suppress magnetic field disturbances in low-field unshielded environments, enhance anti-interference capabilities, reduce image artifacts, and further improve registration and analysis accuracy.

[0092] This invention is based on software algorithms, which does not require major modifications to existing low-field magnetic resonance imaging hardware, thus reducing the cost of technology implementation; it can be widely adapted to various low-field unshielded intraoperative magnetic resonance imaging systems, improving the accessibility of the technology; by improving the level of intraoperative diagnosis and treatment integration, it can shorten operation time, reduce the incidence of postoperative complications, and reduce the risk of secondary surgery for patients.

[0093] The above are merely preferred embodiments of the present invention and are not intended to limit the implementation methods and protection scope of the present invention. Those skilled in the art should recognize that any equivalent substitutions and obvious changes made using the content of this specification and illustrations should be included within the protection scope of the present invention.

Claims

1. A method for intraoperative low-field magnetic resonance imaging processing, characterized in that, include: The preoperative information of the target patient is preprocessed, and the preprocessed preoperative information is input into a pre-trained parameter recommendation model to output a recommended combination of scanning parameters. A dual-gradient echo sequence is constructed based on the recommended combination of the scanning parameters. Magnetic resonance imaging signals are acquired based on the dual-gradient echo sequence to obtain intraoperative images from the magnetic resonance imaging signals. The stitched image of the first frame of the preoperative image and the intraoperative image of the target patient is input into the pre-trained anatomical change prediction model, and the significant displacement area of ​​the anatomical structure is output. The preoperative and intraoperative images are input into a pre-trained feature mapping model to output matching feature vector pairs. The registration parameters are initialized with the significant shift region as a constraint. The registration model optimizes the initialized registration parameters with the degree of matching of the feature vector pairs as a similarity measure, and assigns differential weights to the significant shift region and other regions other than the significant shift region to complete the registration and output the registered intraoperative image.

2. The intraoperative low-field magnetic resonance imaging processing method according to claim 1, characterized in that, The preprocessing of the target patient's preoperative information includes: Obtain the preoperative information of the target patient, including surgical type labeling information, patient basic information, preoperative imaging information, and preoperative diagnostic report; The preoperative image information is preprocessed, and preoperative image features are extracted; Natural language processing is performed on the preoperative diagnostic report to generate diagnostic structured tags; The pre-processed preoperative information includes preoperative image features, structured surgical information tags, and patient basic information. The structured surgical information tags include surgical type labeling information and diagnostic structured tags.

3. The intraoperative low-field magnetic resonance imaging processing method according to claim 1, characterized in that, The pre-training methods for the parameter recommendation model include: Collect a first training dataset, which includes preoperative information of different patients, combinations of scanning parameters for intraoperative images, and imaging quality labels and parameter fit labels. A parameter recommendation model based on the fusion of random forest and convolutional neural network is constructed. It is trained on the first training dataset and optimized using a multi-task loss function. The model is converged through iterative optimization to obtain the trained parameter recommendation model. The multi-task loss function includes a parameter fit loss function and an imaging quality prediction loss function.

4. The intraoperative low-field magnetic resonance imaging processing method according to claim 1, characterized in that, The recommended combination of output scan parameters also includes: In response to a parameter correction command input by the user through the interactive interface, the parameters in the recommended combination of scanning parameters are adjusted accordingly based on the parameter correction command. The parameters in the recommended combination of scanning parameters include at least one of sequence type, repetition time, echo time, k-space fill rate, and number of excitations.

5. The intraoperative low-field magnetic resonance imaging processing method according to claim 1, characterized in that, The process of acquiring magnetic resonance imaging signals based on the dual-gradient echo sequence also includes: Extract phase disturbance signals from navigation echoes; After adaptive filtering of the phase perturbation signal, reverse compensation is performed on the imaging echo to output a phase-corrected magnetic resonance imaging signal. The intraoperative images are obtained from phase-corrected magnetic resonance imaging signals.

6. The intraoperative low-field magnetic resonance imaging processing method according to claim 1, characterized in that, The pre-training method for the anatomical change prediction model includes: Construct a second training dataset, which includes preoperative images, images of anatomical structure displacement areas labeled at each stage of the operation, and corresponding surgical operation records; An improved U-Net network is constructed, comprising an input layer, an encoder, a decoder, and an output layer. The input layer is a stitched image of the first frame of the preoperative image and the intraoperative image. The encoder uses multi-layer concatenated convolutional blocks to extract multi-scale anatomical features through progressive downsampling. The decoder uses multi-layer deconvolutional blocks for feature upsampling and uses an attention layer to assign weights to the surgical region. The output layer is the displacement prediction result of the anatomical structure, and regions where the displacement prediction result exceeds a preset displacement threshold are identified as significant displacement regions. Using the second training dataset as training samples, the improved U-Net network is trained, and the model is converged through iterative optimization to obtain a trained anatomical change prediction model.

7. The intraoperative low-field magnetic resonance imaging processing method according to claim 1, characterized in that, The step of inputting the preoperative image and the intraoperative image into a pre-trained feature mapping model to output a matching feature vector pair includes: Multi-dimensional feature extraction and fusion of preoperative images are performed to generate the first fused feature vector; Multi-dimensional feature extraction and fusion of intraoperative images are performed to generate a second fused feature vector; The feature mapping model is obtained by training a Siamese neural network. The input layer of the Siamese neural network consists of a preoperative image with a first quality and an intraoperative image with a second quality, wherein the second quality is lower than the first quality. The output layer of the Siamese neural network is provided with a feature adaptation layer. The feature mapping model uses a nearest neighbor retrieval algorithm to retrieve matching features from a pre-constructed feature matching dictionary based on the first fused feature vector and the second fused feature vector, and outputs the corresponding matching feature vector pairs of the preoperative image and the intraoperative image.

8. The intraoperative low-field magnetic resonance imaging processing method according to claim 1, characterized in that, The registration includes: The mean pixel offset of the significantly displaced region in the preoperative and intraoperative images was statistically analyzed. An initial affine transformation matrix is ​​constructed based on the mean pixel offset, and the initial affine transformation matrix is ​​used as the initial registration parameter of the registration model; The registration model uses the degree of matching of feature vector pairs as a similarity measure and iteratively optimizes the initial registration parameters using the gradient descent method. During the parameter optimization process, a first weight is assigned to the significant shift region, and a second weight is assigned to other regions except the significant shift region, wherein the second weight is less than the first weight. Calculate the structural similarity index for the intraoperative images after parameter optimization and registration. If the structural similarity index exceeds the preset similarity threshold, output the registered intraoperative images; otherwise, return to the parameter optimization step to iterate again until the structural similarity index exceeds the preset similarity threshold.

9. The intraoperative low-field magnetic resonance imaging processing method according to claim 1, characterized in that, The output of the registered intraoperative images also includes: The intraoperative images are input into a pre-trained feature enhancement model, and the enhanced intraoperative images are output. The enhanced intraoperative images are input into a pre-trained lesion recognition model, and a binary mask image containing the residual lesion area is output. Based on the binary mask image containing the residual lesion area, the residual lesion volume and the lesion volume resection rate are calculated.

10. The intraoperative low-field magnetic resonance imaging processing method according to claim 1, characterized in that, The output of the registered intraoperative images also includes: Based on a pre-constructed multi-dimensional quality assessment index system, the comprehensive quality score of the intraoperative images is determined; wherein, the quality assessment index in the multi-dimensional quality assessment index system includes at least one of lesion boundary clarity, lesion residual volume calculation error, and image signal-to-noise ratio; When the overall quality score is lower than the preset quality threshold, the recommended combination of scanning parameters is fine-tuned.

Images