Intraoperative imaging motion artifact suppression and quality enhancement method and apparatus

Abstract

The application relates to an intraoperative imaging motion artifact suppression and quality enhancement method and device, and relates to the technical field of intraoperative imaging processing. An intraoperative time sequence image sequence output by an intraoperative imaging system is acquired and input into a double-branch space-time adversarial network; the double-branch space-time adversarial network comprises a physically driven motion estimation branch and a medical prior enhancement branch; the physically driven motion estimation branch is based on optical flow field estimation and motion field decomposition; the medical prior enhancement branch performs quality assessment and prior constraint according to medical field knowledge of anatomical structure consistency and tissue texture distribution; and a high-quality image sequence after artifact suppression is generated through an image restoration network based on the motion estimation result and the medical prior constraint. The application suppresses motion artifacts and retains tissue texture and diagnostic characteristics, and improves the stability and diagnostic usability of intraoperative imaging.

Description

Technical Field

[0001] This invention relates to the field of intraoperative imaging signal processing technology, and in particular to a method and apparatus for suppressing motion artifacts and enhancing the quality of intraoperative imaging. Background Technology

[0002] In minimally invasive surgery, intraoperative imaging is the primary means for surgeons to obtain information about the surgical field. Endoscopic systems such as laparoscopes, thoracoscopes, and arthroscopes enter the body cavity through slender instruments, transmitting real-time images of the surgical area to a display device, supporting surgeons in performing delicate operations such as tissue identification, lesion resection, and suturing. However, the patient's continuous breathing, heartbeat, and the physical contact and traction of surgical instruments on tissues during surgery inevitably cause motion artifacts in the intraoperative imaging, manifesting as blurring, ghosting, jitter, and geometric distortion. Specifically, breathing causes overall jitter and periodic displacement, heartbeats transmit through tissues producing localized micro-tremors, and surgical instrument manipulation leads to high-frequency, non-periodic local deformations. The superposition of multiple motion sources severely affects the surgeon's accuracy in identifying tissue structures and the precision of surgical procedures. Various types of surgical video quality degradation (such as smoke removal, deblurring, highlight removal, and flicker removal) have become a research hotspot in the field of surgical video processing.

[0003] Existing methods for motion artifact processing can be broadly categorized into three types. One approach is hardware compensation, which uses motion sensors (such as inertial measurement units, electromagnetic trackers, etc.) attached to the imaging probe or patient surface to sense the motion state of the probe or tissue in real time and performs motion compensation at the image acquisition end. However, this method requires additional hardware, increasing surgical costs and operational complexity. Furthermore, the sensor measurement accuracy is easily affected by the electromagnetic environment of the operating room. More importantly, hardware compensation can only correct the overall rigid motion of the probe or body surface and cannot handle internal soft tissue deformations caused by respiration and heartbeat. Another approach is software post-processing, which performs offline processing after image acquisition. Typical methods include inter-frame interpolation based on optical flow, motion filtering based on frequency domain analysis, and traditional image deblurring algorithms. For example, motion amplification techniques based on time-domain bandpass filtering can separate periodic motion components such as respiration and heartbeat. However, such explicit frequency domain filtering methods assume strict periodicity and linear additive nature of motion, have poor adaptability to non-periodic local deformations caused by instrument manipulation, and have high overall computational complexity, making it difficult to meet the requirements of real-time intraoperative processing. While suppressing artifacts, they often damage effective diagnostic information in the image. One approach is general deep learning methods. In recent years, deep learning methods have made significant progress in the field of image restoration, such as video restoration networks based on deformable convolution and attention mechanisms, video restoration networks based on recurrent structures, and image enhancement networks based on generative adversarial networks. However, general image / video restoration models model degradation patterns of natural images and lack specific modeling for the special characteristics of intraoperative motion artifacts: superposition of multiple motion sources, respiratory and heartbeat movements with physiological cycles, and tissue deformations related to surgical operations. This results in limited restoration effects. Some enhancement methods for endoscopic videos only target a single type of degradation (such as equipment shake or smoke) and do not explicitly distinguish between multiple motion sources. The restoration process may incorrectly alter the real anatomical motion information.

[0004] Therefore, there is an urgent need for an intelligent method specifically designed to suppress motion artifacts in intraoperative imaging: one that can distinguish and process various motion sources such as breathing, heartbeat, and instrument manipulation, while maintaining the integrity of anatomical structures and diagnostic information and meeting the requirements for real-time intraoperative processing. Summary of the Invention

[0005] To solve the above-mentioned technical problems, or at least partially solve them, the present invention provides a method and apparatus for suppressing motion artifacts and enhancing the quality of intraoperative imaging.

[0006] In a first aspect, the present invention provides a method for suppressing motion artifacts and enhancing the quality of intraoperative imaging, comprising: acquiring an intraoperative temporal image sequence output by an intraoperative imaging system; The intraoperative temporal image sequence is input into a two-branch spatiotemporal adversarial network for processing. The two-branch spatiotemporal adversarial network includes a physics-driven motion estimation branch and a medical prior enhancement branch. The physics-driven motion estimation branch uses the motion estimation network to estimate the optical flow field and performs motion field decomposition on the optical flow field for fine-tuning according to interpretability, identifying and separating global respiratory motion, local cardiac pulsation motion, and surgical instrument motion in the intraoperative temporal image sequence. The medical prior enhancement branch performs quality assessment and prior constraints on the separated image frames based on medical domain knowledge of anatomical structure consistency and tissue texture distribution. Based on the motion optical flow field and medical prior constraints, a high-quality image sequence with artifact suppression is generated through an image restoration network.

[0007] Furthermore, the physical-driven motion estimation branch utilizes a motion estimation network to estimate the optical flow field. The motion estimation network combines deformable convolution and Transformer attention mechanisms, taking three consecutive frames as input and outputting the corresponding forward and backward optical flow fields. During training, the motion estimation network incorporates the physiological cycle frequencies of breathing and heartbeat as prior constraints into the network training process, improving the stability and interpretability of the periodic motion component estimation. The physical-driven motion estimation branch uses the physiological cycle frequency estimation unit to perform short-time spectral analysis on the forward optical flow field and the backward optical flow field, and estimates the instantaneous frequency of the global respiratory motion and the local heartbeat motion in real time. The physical-driven motion estimation branch uses the periodic / aperiodic dual-channel decoupling unit to decompose the motion field corresponding to the optical flow location into periodic motion channels and aperiodic motion channels according to the instantaneous frequency. The aperiodic motion channel receives the motion residuals not covered by the periodic motion channel. During training, a physiological cycle consistency loss is applied to constrain the frequency domain energy of the output optical flow of the periodic motion channel to be concentrated in the frequency band corresponding to the instantaneous frequency, and the bandwidth of the frequency band is adaptively adjusted with the instantaneous frequency. The physical-driven motion estimation branch uses the motion field decomposition module to decompose the motion field, and explicitly separates the estimated mixed motion field according to the motion source, into respiratory motion component, heartbeat motion component and surgical instrument motion component.

[0008] Furthermore, the embedding methods of physiological cycle frequency prior include: adding the physiological cycle frequency as a frequency domain constraint term to the training loss function, using the frequency band corresponding to the physiological cycle frequency as a frequency domain regularization constraint of the intermediate features of the motion estimation network, or injecting the physiological cycle frequency information as an additional input channel into the feature extraction process of the motion estimation network.

[0009] Furthermore, the motion field decomposition module employs a gated recurrent unit (GRU) to model the temporal evolution of each motion component. Three parallel GRUs model the temporal evolution of respiratory motion, cardiac pulsation, and surgical instrument motion, respectively. The gating mechanism dynamically allocates weights based on the differences in characteristics of each motion component in the frequency and spatial domains, decomposing the hybrid optical flow into three independent components. The training and supervision signals for the motion field decomposition module are generated using either synchronous annotation of physiological signals or synthetic data. The synchronous annotation of physiological signals involves simultaneously acquiring respiratory zone signals and electrocardiogram (ECG) signals while acquiring intraoperative image sequences, using these signals as temporal annotations for the motion components to supervise the learning of the motion field decomposition module. The synthetic data generation method involves generating training samples with known ground truth motion components using a synthetic motion model with known motion parameters to supervise the learning of the motion field decomposition module.

[0010] Furthermore, the medical prior enhancement branch includes an anatomical structure perception module, a texture fidelity module, and a discriminator based on a medical image diagnostic feature extraction network. The anatomical structure perception module extracts anatomical edges and organ contour features from the image using a pre-trained multi-scale U-Net network and applies anatomical structure consistency loss constraints to these features during the restoration process. The texture fidelity module constrains the noise power spectrum of the restored image based on a noise distribution model of medical images, ensuring that the restored image retains a noise distribution that conforms to the physical characteristics of medical imaging. The medical image diagnostic feature extraction network provides a diagnostic feature preservation loss function for adversarial training of the image restoration network.

[0011] Furthermore, the medical prior enhancement branch employs a diagnostic feature preservation loss function for adversarial training. This diagnostic feature preservation loss function includes a constraint on artifacts lacking diagnostic information and a constraint on the invariance of diagnostic features in the restored image. The constraint on artifacts lacking diagnostic information causes the response of the separated motion artifact components in the feature space of the pre-trained medical image diagnostic feature extraction network to approach zero. The constraint on the invariance of diagnostic features in the restored image ensures that the distribution of diagnostic features in the feature space is consistent between the restored image generated based on the motion separation results and the original clear image. During adversarial training, a pre-trained medical image diagnostic feature extraction network is used as the discriminator. The motion artifact component is used as a negative sample and the original clear image is used as a positive sample for adversarial training to ensure the integrity of the diagnostic information of the recovered image.

[0012] Furthermore, based on the motion separation results and medical prior constraints, the image restoration network performs inverse motion compensation on the image using the separated motion components, and further optimizes the image quality through adversarial training of the generative adversarial network. The reverse motion compensation can be implemented in the following ways: pixel-level resampling compensation based on optical flow field, deformation field compensation based on motion components, inter-frame interpolation compensation driven by motion components, or a combination of the above methods. The image restoration network adopts a generative adversarial network architecture, including a generator and an adversarial discriminator; The loss function of the generator is composed of a weighted average of image content reconstruction loss, perceptual loss, image quality perceptual loss, structural similarity loss, and medical prior loss. The adversarial discriminator is used to distinguish between the image generated by the image restoration network and the original clear image. The discriminator in the medical prior enhancement branch, which is composed of a pre-trained medical image diagnostic feature extraction network, is used to distinguish between the motion artifact component and the original clear image. The two discriminators constrain the image restoration network from the dimensions of image authenticity and diagnostic information integrity, respectively.

[0013] Furthermore, the weights of each loss component are adaptively adjusted according to the intraoperative imaging modality, including but not limited to: modality-preset weight combination, weight calibration based on validation set grid search, automatic weight learning based on uncertainty learning, transfer weight based on task similarity, or a combination of the above methods.

[0014] Furthermore, the temporal statistical features of each motion component output by the motion field decomposition module are used to construct an individualized motion feature descriptor. The individualized motion feature descriptor includes respiratory rate estimates, heart rate estimates, energy proportions of each motion component, and energy fluctuation amplitudes. When the image quality assessment index is lower than a preset threshold, and the deviation of the individualized motion feature descriptor from the motion feature distribution of the pre-training dataset exceeds a preset range, lightweight online fine-tuning of the physical-driven motion estimation branch and the motion field decomposition module is triggered. The lightweight online fine-tuning only updates the parameters of the last two layers of the motion estimation network, the parameters of the gated recurrent unit of the motion field decomposition module, and the parameters of the physiological cycle frequency estimation unit. A constraint term on the pre-training parameters is added to the fine-tuning loss to prevent catastrophic forgetting, enabling the bi-branch spatiotemporal adversarial network to adapt to specific surgical scenarios and the individualized motion characteristics of patients.

[0015] Secondly, the present invention provides an intraoperative imaging motion artifact suppression and quality enhancement device, comprising: at least one processing unit, the processing unit being connected to a storage unit via a bus unit, the storage unit storing a computer program executable by the processing unit, and the computer program being executed by the processing unit to implement the intraoperative imaging motion artifact suppression and quality enhancement method.

[0016] The technical solutions provided in the embodiments of the present invention have the following advantages compared with the prior art: This application uses a physics-driven motion estimation branch to explicitly identify and separate three motion sources: respiratory motion, heartbeat motion, and surgical instrument motion. This enables interpretable differentiation between motion artifacts and real tissue motion, overcoming the problem that traditional methods struggle to distinguish between motion artifacts and real tissue motion. This provides accurate motion source separation results for subsequent inverse motion compensation. This application introduces anatomical structure consistency constraints, noise power spectrum constraints, and diagnostic feature preservation constraints through a medical prior enhancement branch. This ensures the integrity of medical diagnostic information in the restored image while suppressing motion artifacts, thereby avoiding the problem of overcorrection in intraoperative scenarios by general image restoration methods. This application achieves a dynamic balance between pixel fidelity, perceptual quality, structural integrity, and medical constraints by weighting a combination of image content reconstruction loss, perceptual loss, image quality perceptual loss, structural similarity loss, and medical prior loss, with the weight of each loss component adaptively adjusted according to the intraoperative imaging modality. This balances the strength of artifact suppression with the preservation of tissue texture and diagnostic details. This application utilizes an online adaptive fine-tuning mechanism to enable the bi-branch spatiotemporal adversarial network to adapt to different surgical types and individualized motion characteristics of patients, continuously improving artifact suppression as the number of surgical cases accumulates. Attached Figure Description

[0017] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, those skilled in the art can obtain other drawings based on these drawings without creative effort.

[0019] Figure 1 A flowchart illustrating an intraoperative imaging motion artifact suppression and quality enhancement method provided in this embodiment of the invention; Figure 2 A schematic diagram of the overall framework of the intraoperative imaging motion artifact suppression and quality enhancement method provided in the embodiments of the present invention; Figure 3 A schematic diagram of a sports field decomposition module provided in an embodiment of the present invention; Figure 4 An architecture diagram of a medical prior enhancement branch provided for an embodiment of the present invention; Figure 5 This is a schematic diagram of image enhancement for the intraoperative imaging motion artifact suppression and quality enhancement method provided in an embodiment of the present invention; Figure 6This is a schematic diagram of an intraoperative imaging motion artifact suppression and quality enhancement device provided in an embodiment of the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, not all, of the embodiments of the present invention. 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, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0022] Example 1 like Figure 1 and Figure 5 As shown, the present invention provides a method for suppressing motion artifacts and enhancing the quality of intraoperative imaging, comprising: S100: Acquire the intraoperative time-series image sequence output by the intraoperative imaging system.

[0023] For example, in laparoscopic cholecystectomy, because the patient still has regular breathing movements under general anesthesia, the endoscopic image continuously shows image jitter and blurring in sync with the breathing rhythm.

[0024] S200. Input the intraoperative temporal image sequence into a dual-branch spatiotemporal adversarial network for processing.

[0025] The bi-branch spatiotemporal adversarial network comprises a physics-driven motion estimation branch and a medical prior enhancement branch. The physics-driven motion estimation branch, based on optical flow field estimation and motion field decomposition, identifies and separates global respiratory motion, local cardiac pulsation motion, and surgical instrument motion in intraoperative temporal image sequences. The medical prior enhancement branch, based on medical domain knowledge of anatomical structure consistency and tissue texture distribution, performs quality assessment and prior constraints on the separated image frames.

[0026] In the specific implementation process, such as Figure 2 As shown, the working process of a dual-branch spatiotemporal adversarial network includes: The physical-driven motion estimation branch uses a motion estimation network to estimate the optical flow field.

[0027] Specifically, the physics-driven motion estimation branch employs a motion estimation network combining deformable convolution and Transformer attention mechanisms, based on three consecutive frames of images. , , As input, multi-scale motion features are extracted through deformable convolutional layers, and then long-range spatiotemporal dependencies are modeled through Transformer self-attention layers to output the corresponding forward optical flow field. and backward optical flow field .

[0028] The combined structure of deformable convolution and Transformer attention balances the adaptability of local deformation of motion features with the ability to model long-range spatiotemporal dependencies: On the one hand, deformable convolution learns an offset for each sampling position of the convolution kernel, enabling the receptive field to adapt to the deformation of the image content, thereby adapting to the non-rigid local deformation of tissues caused by breathing, heartbeat and instrument traction during surgery, and preserving fine motion boundary information at the feature level; on the other hand, the Transformer self-attention layer performs global modeling of the dependency relationship between any two positions in the feature map, which can capture the spatiotemporal correlation of slow motion and periodic motion across multiple frames and large areas of the image. Those skilled in the art will understand that the motion estimation network is not limited to the above-described combined structure, and other network structures capable of motion field estimation can also be used, such as iteratively refined optical flow networks based on cost volumes, optical flow networks based on feature pyramids and raster pooling, or recursive optical flow networks based on convolutional gated recurrent units, as long as they can output a motion field representing inter-frame motion; correspondingly, the number of input frames of the motion estimation network is not limited to three frames, but can be two consecutive frames, multiple consecutive frames, or a frame sequence containing a temporal context window. The input multi-frame images can be original image frames or image features after preprocessing, downsampling, or feature extraction.

[0029] The motion estimation network incorporates the physiological cycle frequencies of respiration and heartbeat as prior constraints during training. Intraoperative motion artifacts are composed of multiple physical sources superimposed. Respiratory and heartbeat movements exhibit quasi-periodic characteristics consistent with human physiological rhythms, while surgical instrument manipulation movements manifest as aperiodic, sudden localized deformations. Introducing the physiological cycle frequency prior during the motion estimation network training guides the network to prioritize learning motion components that conform to physiological cycle patterns, suppressing overfitting to aperiodic noise, thereby improving the stability and interpretability of periodic motion component estimation. The embedding methods for the physiological cycle frequency prior include: adding the physiological cycle frequency as a frequency domain constraint to the training loss function; using the frequency band corresponding to the physiological cycle frequency as a frequency domain regularization constraint for intermediate features of the motion estimation network; or injecting physiological cycle frequency information as an additional input channel into the feature extraction process of the motion estimation network. As an example, the prior constraints are embedded in the training loss in the form of frequency domain constraints. The respiratory cycle frequency (typically resting value is about 0.2-0.3Hz, frequency band range 0.15-0.35Hz) is embedded as a frequency domain prior constraint, so that the network prioritizes learning motion components that conform to the physiological cycle.

[0030] The physical-driven motion estimation branch also includes a physiological cycle frequency estimation unit and a periodic / aperiodic dual-channel decoupling unit, performing periodic / aperiodic dual-channel decoupling based on instantaneous frequency. Specifically: The physiological cycle frequency estimation unit estimates the forward optical flow field. and the aforementioned backward optical flow field Short-time spectral analysis is performed to estimate the instantaneous frequencies of global respiratory motion and local cardiac pulsation in real time. The frequencies of respiratory and cardiac pulsation are not fixed but drift with factors such as anesthesia depth, body position, drug effects, and individual patient differences. Therefore, for the motion estimation network, relying solely on the fixed prior frequencies during the training phase is insufficient to adapt to the dynamically changing physiological state during surgery. This embodiment uses the physiological cycle frequency estimation unit to perform time-frequency analysis on the estimated optical flow field, estimating the instantaneous frequencies of each cycle's motion in real time, thereby adaptively tracking changes in physiological state during the inference phase. The spectral analysis can be implemented using any of the following methods: short-time Fourier transform, continuous wavelet transform, parameter spectrum estimation based on an autoregressive model, or time-frequency feature mapping based on a deep neural network. The spectral analysis can perform global analysis of the entire optical flow field or partitioned analysis of local regions of the optical flow field, such as a local window corresponding to the cardiac pulsation conduction region. As an example, the short-time spectrum analysis uses a 128-frame sliding window and a Hanning window to update the instantaneous frequency estimates of respiratory and cardiac pulsations frame by frame by tracking peak frequencies, thereby adaptively tracking instantaneous frequencies under physiological fluctuations such as changes in anesthesia depth and heart rate drift.

[0031] The periodic / aperiodic dual-channel decoupling unit decomposes the motion field corresponding to the optical flow location into periodic motion channels and aperiodic motion channels based on the instantaneous frequency. The aperiodic motion channel receives motion residuals not covered by the periodic motion channel. The various motion sources in the intraoperative motion field are distinguishable in the frequency domain: respiratory and cardiac pulsation motions are concentrated in the low-frequency quasi-periodic frequency band corresponding to physiological rhythms, while surgical instrument motion exhibits broadband, aperiodic characteristics. In this embodiment, the periodic / aperiodic dual-channel decoupling unit divides the motion field into periodic and aperiodic motion channels based on the real-time estimated instantaneous frequency, enabling subsequent motion field decomposition modules to perform more refined component separation on the decoupled channels. Simultaneously, the aperiodic motion channel receives motion residuals not covered by the periodic channel, preventing aperiodic instrument motion from being mistakenly classified into the periodic motion channel. As an example, the periodic motion channel receives respiratory and cardiac pulsation motions within the 0.15-1.2Hz frequency band, while the aperiodic motion channel receives high-frequency, aperiodic surgical instrument motion residuals.

[0032] During training, a physiological cycle consistency loss is applied to constrain the frequency domain energy of the optical flow output from the periodic motion channel to be concentrated in the frequency band corresponding to the instantaneous frequency. The bandwidth of this frequency band adaptively adjusts with the instantaneous frequency, thereby guiding motion estimation using physiological cycle patterns without relying on strict periodic assumptions, while also accommodating non-periodic surgical instrument movements. The purpose of the physiological cycle consistency loss is to constrain the optical flow output from the periodic motion channel to be concentrated in the frequency band corresponding to the instantaneous frequency during the training phase, ensuring that the periodic motion channel learned by the network truly represents physiological cycle motion. Simultaneously, the bandwidth adaptively adjusts with the instantaneous frequency, expanding or contracting as the physiological frequency shifts, thus maintaining the periodic constraint while avoiding dependence on strict periodic assumptions. The physiological cycle consistency loss can be implemented in various forms, including: a frequency band energy ratio form, an out-of-band energy penalty form, a cross-correlation form between periodic motion and a reference physiological signal, or a combination of the above forms. As an example, the physiological cycle consistency loss adopts the frequency band energy ratio form, i.e.: ; in, This is due to the loss of consistency in the menstrual cycle. The instantaneous frequency; The output optical flow of the periodic motion channel at the instantaneous frequency Frequency domain energy within the corresponding frequency band; The total frequency domain energy of the output optical flow of the periodic motion channel. As an example, the bandwidth is taken as ±30% of the instantaneous frequency.

[0033] The physical-driven motion estimation branch also includes a motion field decomposition module, which performs hybrid optical flow field decomposition. The estimated hybrid optical flow field is explicitly separated according to the motion source, that is, the hybrid optical flow is decomposed into respiratory motion components, heartbeat motion components, and surgical instrument motion components. This enables the subsequent medical prior enhancement branch and image restoration network to compensate and enhance for different motion sources, thereby solving the problem of difficulty in distinguishing motion artifacts from real tissue motion.

[0034] like Figure 3 As shown, the motion field decomposition module receives the mixed optical flow field and models the temporal evolution of respiratory motion, cardiac pulsation motion, and surgical instrument motion through three parallel gated recurrent units (GRUs). The gating mechanism dynamically allocates weights based on the differences in characteristics of each motion component in the frequency and spatial domains, decomposing the mixed optical flow into three independent components. Each motion source has distinguishable characteristics in the frequency and spatial domains: the respiratory motion component is concentrated in the low-frequency quasi-periodic frequency band, such as the overall organ displacement caused by diaphragmatic movement, with a large spatial range and mainly vertical displacement; the cardiac pulsation component is concentrated in the relatively high-frequency quasi-periodic frequency band, such as the local micro-vibrations transmitted from the heartbeat to adjacent tissues, with a small spatial range and periodic changes in direction; the surgical instrument motion component exhibits high-frequency, non-periodic local deformation. The gated recurrent units achieve dynamic separation of the motion components by modeling the temporal evolution of each component and combining the above-mentioned differences in frequency and spatial domain characteristics. The number of motion components is not limited to three and can be dynamically configured according to the actual number of motion sources. For example, a single component can be configured when only respiratory artifacts need to be suppressed, while a three-component configuration is used when respiratory, heartbeat, and instrument motion need to be suppressed simultaneously. Taking the respiratory motion component as an example, its frequency domain energy is concentrated in the range of 0.15-0.35Hz, and its spatial domain manifestation is the overall vertical displacement of the organ caused by diaphragmatic movement. The heartbeat component has a frequency domain energy concentrated in the range of 0.8-1.2Hz, manifesting as local micro-tremors transmitted from the heartbeat to adjacent tissues. The surgical instrument motion component manifests as high-frequency, non-periodic local deformation. As an example, the training supervision signal of the motion field decomposition module is generated using physiological signal synchronous annotation or synthetic data. The physiological signal synchronous annotation method is as follows: while acquiring intraoperative image sequences, respiratory zone signals and electrocardiogram signals are acquired simultaneously, and the respiratory zone signals and electrocardiogram signals are used as temporal annotations of motion components to supervise the learning of the motion field decomposition module; the synthetic data generation method is as follows: using a synthetic motion model with known motion parameters, such as rigid body displacement superimposed with periodic deformation superimposed with local non-periodic deformation, training samples with known true value motion components are generated to supervise the learning of the motion field decomposition module.

[0035] The medical prior enhancement branch performs quality assessment and prior constraints on the separated image frames based on medical domain knowledge of anatomical structure consistency and tissue texture distribution.

[0036] Specifically, such as Figure 4 As shown, the medical prior enhancement branch includes an anatomical structure perception module, a texture fidelity module, and a discriminator based on a medical image diagnostic feature extraction network.

[0037] The anatomical structure perception module extracts anatomical edge and organ contour features in the current frame using a pre-trained multi-scale U-Net network, and applies anatomical structure consistency loss constraints to the features during the restoration process to prevent the restoration process from incorrectly modifying the anatomical structure.

[0038] The anatomical structure perception module leverages the stable prior spatial distribution of anatomical structures in medical images, such as the relative positions of organ contours, blood vessel orientations, and tissue boundaries under normal physiological conditions. This constrains the image restoration network to maintain the geometrical accuracy of anatomical structures while suppressing motion artifacts, preventing erroneous modifications to anatomical edges due to over-smoothing or over-sharpening during restoration, thus avoiding misleading pseudo-anatomical structures. The anatomical structure consistency loss can apply to either intermediate network features (feature-level constraints) or the spatial mapping between the restored image and the reference anatomical structure (image-level constraints). As an example, the anatomical structure perception module uses a pre-trained multi-scale U-Net network to extract anatomical edge and organ contour features such as the gallbladder edge and common bile duct contour in the current frame. During restoration, it applies anatomical structure consistency loss constraints to these features, preventing erroneous modifications to anatomical structures. The pre-trained multi-scale U-Net network is trained on natural or medical image datasets, and its extracted anatomical features include, but are not limited to, geometrically stable structural features such as organ edges, tissue boundaries, blood vessel orientations, and organ contours.

[0039] The texture fidelity module, based on a noise distribution model of medical images, constrains the noise power spectrum of the restored image to ensure that the restored image retains a noise distribution that conforms to the physical characteristics of medical imaging.

[0040] Specifically, the texture fidelity module constrains the noise power spectrum of the reconstructed image based on a noise distribution model specific to medical images, ensuring that the reconstructed image retains a noise distribution consistent with the physical characteristics of medical imaging. As an example, in a laparoscopic scenario, the noise distribution model employs a Poisson-Gaussian mixture noise model. Laparoscopic imaging systems are limited by lighting conditions, and the photon counting process follows a Poisson distribution, making the signal dependent on the noise. In contrast, the thermal noise generated by electronic circuits follows a Gaussian distribution, making the signal independent of the noise. Simple pixel-domain constraints often lead to overly smoothed images, losing high-frequency texture details. By constraining the noise power spectrum of the reconstructed image using a noise distribution model, it ensures that the reconstructed image retains realistic noise consistent with the physical characteristics of the laparoscopic optical system, avoiding visual distortion and diagnostic misrepresentation caused by overly clean images.

[0041] In laparoscopic imaging, assuming the observed degraded image is y and the potential sharp image is x. Since the photon counting process follows a Poisson distribution, and the Gaussian noise introduced by the electronic circuitry follows a Gaussian distribution, the imaging process from the potential sharp image x to the degraded image y is modeled as follows: ,in, This represents a Poisson noise process. This indicates that the mean is 0 and the variance is 0. Additive Gaussian noise.

[0042] For the restored image, calculate its noise power spectrum. Specifically, the noise component of the recovered image is extracted using high-pass filtering. The squared modulus of the two-dimensional discrete Fourier transform of the image formed by the noise components yields the noise power spectrum of the reconstructed image. : ; in, For frequency domain coordinates, and For the width and height of the image, This represents the Fourier transform operator.

[0043] To constrain the noise distribution of the recovered image to conform to physical properties, a texture fidelity loss function is constructed. The constraint is achieved by minimizing the difference between the noise power spectrum of the recovered image and the noise power spectrum of the theoretical target: , in: This represents the actual noise power spectrum calculated from the recovered image; This represents the theoretical noise power spectrum derived based on the Poisson-Gaussian mixture model. express The norm is used to measure the distance between two spectral distributions.

[0044] By using texture fidelity loss function By incorporating the overall optimization objective, the network will not only focus on denoising effects during training, but will also be forced to learn to preserve noise texture structures that conform to physical laws.

[0045] The medical prior enhancement branch of the medical image diagnostic feature extraction network provides a diagnostic feature preservation loss function for adversarial training of the image restoration network.

[0046] Specifically, the diagnostic feature preservation loss function includes an artifact-free diagnostic information constraint and a restored image diagnostic feature invariance constraint. The artifact-free diagnostic information constraint ensures that the separated motion artifact component has a near-zero response in the feature space of the pre-trained medical image diagnostic feature extraction network. That is, motion artifacts, as non-diagnostic information superimposed during the imaging process, do not generate a response in the diagnostic feature space, thus preventing the restoration process from misclassifying diagnostic information as artifacts and deleting it. The motion artifact component consists of imaging artifacts caused by respiratory motion, heartbeat motion, and surgical instrument motion, i.e., the artifact portion in the motion separation result. The restored image diagnostic feature invariance constraint ensures that the diagnostic feature distribution of the restored image generated based on the motion separation result is consistent with that of the original clear image in the feature space.

[0047] During adversarial training, a pre-trained medical image diagnostic feature extraction network acts as a discriminator. Motion artifact components are used as negative samples, and the original clear image is used as a positive sample for adversarial training. The training signal directly targets the integrity of diagnostic information, thereby ensuring the integrity of the diagnostic information in the recovered image. In one embodiment, the motion artifact components are mapped to the discriminator input space via a shallow projection network before being input to the discriminator.

[0048] The mechanism of adversarial training for maintaining diagnostic features lies in the fact that diagnostic information in intraoperative images, such as lesion morphology, tissue boundaries, blood vessel distribution, and organ contours, is distinguishable from motion artifacts in the feature space. Motion artifacts are caused by physical motion during the imaging process and are non-diagnostic information superimposed on the real tissue structure. They should not produce a significant response in the feature space of the medical image diagnostic feature extraction network. Through adversarial training, the response of the motion artifact component in the diagnostic feature space is made close to zero, while constraining the distribution of diagnostic features in the restored image to be consistent with that in the original clear image. This ensures that diagnostic information is not deleted or tampered with while suppressing artifacts.

[0049] The implementation methods of the medical image diagnostic feature extraction network include, but are not limited to: feature extraction layers of a pre-trained medical image classification network, feature extraction networks pre-trained based on self-supervised contrastive learning, feature extraction networks based on anatomical structure recognition, or diagnostic feature networks based on multi-scale feature fusion. The implementation methods of adversarial training include: discriminative adversarial training based on generative adversarial networks, feature space adversarial training based on feature matching, domain adversarial training based on gradient inversion, or a combination of the above methods.

[0050] S300, based on motion optical flow field and medical prior constraints, generates high-quality image sequences with artifact suppression through image restoration network.

[0051] Specifically, based on the motion separation results and medical prior constraints, the image restoration network performs inverse motion compensation on the image using the separated motion components, and further optimizes the image quality through adversarial training with a generative adversarial network. On the one hand, inverse motion compensation is performed on the image based on the separated motion components, physically eliminating image misalignment and blurring caused by motion; on the other hand, adversarial training with a generative adversarial network seeks a balance between pixel-level fidelity and visual perception quality. Given the trade-off between artifact suppression and detail preservation, a single loss function cannot simultaneously consider pixel-level fidelity, perceptual quality, structural integrity, and medical constraints. Therefore, the image restoration network described in this embodiment adopts a generative adversarial network architecture, achieving the synergistic effect of the above objectives through joint optimization of multiple losses.

[0052] The reverse motion compensation can be implemented in the following ways: pixel-level resampling compensation based on optical flow field, deformation field compensation based on motion components, inter-frame interpolation compensation driven by motion components, or a combination of the above methods.

[0053] The image restoration network adopts a generative adversarial network architecture, including a generator and an adversarial discriminator; the implementation methods of the image restoration network include: a two-stage restoration architecture based on generative adversarial networks (GANs), an iterative restoration architecture based on a diffusion model, a restoration architecture based on a multi-scale encoder-decoder, and a global context restoration architecture based on Transformer.

[0054] The generator's loss function is a weighted sum of image content reconstruction loss, perceptual loss, image quality perceptual loss, structural similarity loss, and medical prior loss. The generator's total loss satisfies the following: ; in, This represents the total loss of the generator; Image content reconstruction loss; To perceive loss; Image quality perceived loss; For structural similarity loss; The medical prior loss is a weighted average of the loss of consistency of the anatomical structure and the loss of preservation of the diagnostic features. 、 、 、 、 These are the weights for the corresponding loss components, and the weights of each loss component are adaptively adjusted according to the intraoperative imaging modality. Image content reconstruction loss ensures pixel-level fidelity, perceptual loss ensures high-level feature consistency, image quality perception loss provides a perceptual evaluation of the overall quality of the restored image, structural similarity loss ensures image structural integrity, and medical prior loss ensures anatomical edge consistency. Different intraoperative imaging modalities have different degradation characteristics. For example, laparoscopic images are significantly affected by illumination and sensor noise, while thoracic images are significantly affected by the superposition of multiple motion sources. Therefore, the weights of each loss component are adaptively adjusted according to the imaging modality and artifact characteristics, so that the joint optimization objective is adapted to the specific application scenario. The adaptive adjustment of the weights can be achieved through: pre-set weight combination according to modality, weight calibration based on validation set grid search, automatic weight learning based on uncertainty learning, transfer weights based on task similarity, or a combination of the above methods. As an example of pre-set weight combination according to modality, laparoscopic... 、 、 、 、 The values ​​were set to 0.3, 0.12, 0.14, 0.20, and 0.24 respectively for the thoracoscopy. 、 、 、 、 The values ​​were set to 0.26, 0.18, 0.20, 0.16, and 0.20 respectively for arthroscopy. 、 、 、 、 The values ​​were set to 0.28, 0.12, 0.16, 0.22, and 0.22 respectively for open surgery. 、 、 、 、 The values ​​were set to 0.32, 0.16, 0.14, 0.20, and 0.18, respectively. Laparoscopy and arthroscopy increased structural similarity and medical prior weights to suppress pseudostructures under endoscopic high-brightness / low-texture conditions; thoracoscopy increased perception and image quality weights to adapt to more significant multi-source motion and visual field changes; and open surgery increased content reconstruction weights to constrain pixel and color shifts in a wider field of view.

[0055] The adversarial discriminator is used to distinguish between the image generated by the image restoration network and the original clear image. The discriminator in the medical prior enhancement branch, which is composed of a pre-trained medical image diagnostic feature extraction network, is used to distinguish between the motion artifact component and the original clear image. The two discriminators constrain the image restoration network from the dimensions of image authenticity and diagnostic information integrity, respectively.

[0056] This application supports online fine-tuning. In the initial deployment phase, a pre-trained model is used for artifact suppression. After accumulating data from several surgical cases, lightweight online fine-tuning of the network is performed using the accumulated intraoperative image data. Specifically, based on the temporal statistical characteristics of each motion component output by the motion field decomposition module, an individualized motion feature descriptor is constructed. This individualized motion feature descriptor includes respiratory rate estimates, heart rate estimates, the energy percentage of each motion component, and the energy fluctuation amplitude. As an example, the individualized motion feature descriptor constructed intraoperatively for a patient is as follows: respiratory rate estimate 0.2Hz, heart rate estimate 1.1Hz, respiratory motion component energy percentage 60%, and energy fluctuation amplitude within a preset fluctuation range. When the image quality assessment index is lower than a preset threshold, and the deviation of the individualized motion feature descriptor from the motion feature distribution of the pre-trained dataset exceeds a preset range, lightweight online fine-tuning of the physical-driven motion estimation branch and the motion field decomposition module is triggered. The estimated respiratory rate, estimated heart rate, energy proportion of each motion component, and energy fluctuation amplitude in the individualized motion feature descriptor are all motion features. In the pre-training dataset, the mean vector of motion features is calculated according to the imaging modality. Covariance Matrix As a motion feature distribution, the deviation of the individualized motion feature descriptor z from the motion feature distribution is expressed by the regularized squared Mahalanobis distance: The motion feature descriptor has an element dimension of 8 and follows a multivariate normal distribution. The regularized Mahalanobis distance squared approximately follows an 8-DOF chi-square distribution. This implementation uses the 99th percentile (20.09) as the deviation threshold, with a normal range of 0-20.09; values ​​exceeding 20.09 are considered high deviations. Furthermore, the main sources of deviation are determined based on the standard deviation of each element. This dual-condition triggering mechanism ensures that fine-tuning occurs only when both quality degradation and feature deviation are simultaneously met, avoiding the introduction of noise through unconditional continuous fine-tuning.

[0057] The process of selecting the preset threshold for image quality assessment indicators involves collecting multiple acceptable and unacceptable segments according to the imaging modality, calculating the recall rate of unacceptable segments and the specificity of acceptable segments for each candidate threshold, and selecting the threshold with the largest Youden index from the candidate thresholds with an unacceptable segment recall rate of not less than 95%. The resulting threshold is not lower than the safety lower limit.

[0058] The lightweight online fine-tuning only updates the parameters of the last two layers of the motion estimation network, which are two fully connected layers close to the output. The parameters of the gated recurrent unit of the motion field decomposition module and the parameters of the physiological cycle frequency estimation unit are updated, while the weights of the remaining parameters of the core feature extraction network remain unchanged. A constraint term on the pre-trained parameters is added to the fine-tuning loss to prevent catastrophic forgetting, ensuring that the fine-tuning process is stable and efficient.

[0059] Online fine-tuning enables the bi-branch spatiotemporal adversarial network to adapt to the characteristics of specific surgical types and hospital environments, thereby adaptively improving artifact suppression.

[0060] Example 2 like Figure 6 As shown, this embodiment of the invention provides an intraoperative imaging motion artifact suppression and quality enhancement device, comprising: at least one processing unit, the processing unit being connected to a storage unit via a bus unit, the storage unit serving as a computer-readable storage medium, which can be used to store software programs, computer-executable programs, and modules, such as the software program, computer-executable program, and module corresponding to the intraoperative imaging motion artifact suppression and quality enhancement method in this embodiment of the invention. The processing unit implements the aforementioned intraoperative imaging motion artifact suppression and quality enhancement method by running the software program, computer-executable program, and module stored in the storage unit.

[0061] Of course, the computer program stored in the storage unit of the intraoperative imaging motion artifact suppression and quality enhancement device provided in the embodiments of the present invention is not limited to the method operation described above, but can also execute related operations in the intraoperative imaging motion artifact suppression and quality enhancement method provided in any embodiment of the present invention.

[0062] Example 3 This invention provides a computer-readable storage medium storing a computer program that, when executed, implements the intraoperative imaging motion artifact suppression and quality enhancement method.

[0063] The computer-readable storage medium provided in the embodiments of the present invention stores a computer program that is not limited to the method operation described above, but can also execute related operations in the intraoperative imaging motion artifact suppression and quality enhancement method provided in any embodiment of the present invention.

[0064] In the embodiments provided by this invention, it should be understood that the disclosed structures and methods can be implemented in other ways. For example, the structural embodiments described above are merely illustrative. For instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be an indirect coupling or communication connection through some interfaces, structures, or units, and may be electrical, mechanical, or other forms.

[0065] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0066] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0067] The above description is merely a specific embodiment of the present invention, enabling those skilled in the art to understand or implement the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the present invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features claimed herein.

Claims

1. A method for suppressing motion artifacts and enhancing the quality of intraoperative imaging, characterized in that, include: Acquire the intraoperative time-series image sequence output by the intraoperative imaging system; The intraoperative temporal image sequence is input into a two-branch spatiotemporal adversarial network for processing. The two-branch spatiotemporal adversarial network includes a physics-driven motion estimation branch and a medical prior enhancement branch. The physics-driven motion estimation branch uses the motion estimation network to estimate the optical flow field and performs motion field decomposition on the optical flow field for fine-tuning according to interpretability, identifying and separating global respiratory motion, local cardiac pulsation motion, and surgical instrument motion in the intraoperative temporal image sequence. The medical prior enhancement branch performs quality assessment and prior constraints on the separated image frames based on medical domain knowledge of anatomical structure consistency and tissue texture distribution. Based on the motion optical flow field and medical prior constraints, a high-quality image sequence with artifact suppression is generated through an image restoration network.

2. The method for suppressing motion artifacts and enhancing the quality of intraoperative imaging according to claim 1, characterized in that, The physics-driven motion estimation branch uses a motion estimation network to estimate the optical flow field. The motion estimation network combines deformable convolution and Transformer attention mechanisms, taking three consecutive frames as input and outputting the corresponding forward and backward optical flow fields. During training, the motion estimation network incorporates the physiological cycle frequencies of breathing and heartbeat as prior constraints into the network training process to improve the stability and interpretability of the periodic motion component estimation. The physical-driven motion estimation branch uses the physiological cycle frequency estimation unit to perform short-time spectral analysis on the forward optical flow field and the backward optical flow field, and estimates the instantaneous frequency of the global respiratory motion and the local heartbeat motion in real time. The physical-driven motion estimation branch uses the periodic / aperiodic dual-channel decoupling unit to decompose the motion field corresponding to the optical flow location into periodic motion channels and aperiodic motion channels according to the instantaneous frequency. The aperiodic motion channel receives the motion residuals not covered by the periodic motion channel. During training, a physiological cycle consistency loss is applied to constrain the frequency domain energy of the optical flow output by the periodic motion channel to be concentrated in the frequency band corresponding to the instantaneous frequency, and the bandwidth of the frequency band is adaptively adjusted with the instantaneous frequency. The physical-driven motion estimation branch uses the motion field decomposition module to decompose the motion field, and explicitly separates the estimated mixed motion field according to the motion source, into respiratory motion component, heartbeat motion component and surgical instrument motion component.

3. The method for suppressing motion artifacts and enhancing the quality of intraoperative imaging according to claim 2, characterized in that, Embedding methods for physiological cycle frequency priors include: adding the physiological cycle frequency as a frequency domain constraint term to the training loss function, using the frequency band corresponding to the physiological cycle frequency as a frequency domain regularization constraint for intermediate features of the motion estimation network, or injecting physiological cycle frequency information as an additional input channel into the feature extraction process of the motion estimation network.

4. The method for suppressing motion artifacts and enhancing the quality of intraoperative imaging according to claim 2, characterized in that, The motion field decomposition module employs a gated recurrent unit (GRU) to model the temporal evolution of each motion component. Three parallel GRUs model the temporal evolution of respiratory motion, cardiac pulsation, and surgical instrument motion, respectively. The gating mechanism dynamically assigns weights based on the differences in characteristics of each motion component in the frequency and spatial domains, decomposing the hybrid optical flow into three independent components. The training and supervision signals for the motion field decomposition module are generated using either synchronous annotation of physiological signals or synthetic data. The synchronous annotation of physiological signals involves simultaneously acquiring respiratory zone signals and electrocardiogram (ECG) signals while acquiring intraoperative image sequences, using these signals as temporal annotations for the motion components to supervise the learning of the motion field decomposition module. The synthetic data generation method involves generating training samples with known ground truth motion components using a synthetic motion model with known motion parameters to supervise the learning of the motion field decomposition module.

5. The method for suppressing motion artifacts and enhancing the quality of intraoperative imaging according to claim 1, characterized in that, The medical prior enhancement branch includes an anatomical structure perception module, a texture fidelity module, and a discriminator based on a medical image diagnostic feature extraction network. The anatomical structure perception module extracts anatomical edges and organ contour features from the image using a pre-trained multi-scale U-Net network and applies anatomical structure consistency loss constraints to these features during the restoration process. The texture fidelity module constrains the noise power spectrum of the restored image based on a noise distribution model of the medical image, ensuring that the restored image retains a noise distribution that conforms to the physical characteristics of medical imaging. The medical image diagnostic feature extraction network provides a diagnostic feature preservation loss function for adversarial training of the image restoration network.

6. The method for suppressing motion artifacts and enhancing the quality of intraoperative imaging according to claim 5, characterized in that, The medical prior enhancement branch employs a diagnostic feature preservation loss function for adversarial training. This function includes a constraint on artifacts lacking diagnostic information and a constraint on the invariance of diagnostic features in the restored image. The constraint on artifacts lacking diagnostic information causes the separated motion artifact components to have a response close to zero in the feature space of the pre-trained medical image diagnostic feature extraction network. The constraint on the invariance of diagnostic features in the restored image ensures that the distribution of diagnostic features in the restored image generated based on the motion separation results is consistent with that of the original clear image in the feature space. During adversarial training, a pre-trained medical image diagnostic feature extraction network is used as the discriminator. The motion artifact component is used as a negative sample and the original clear image is used as a positive sample for adversarial training to ensure the integrity of the diagnostic information of the recovered image.

7. The method for suppressing motion artifacts and enhancing the quality of intraoperative imaging according to claim 1, characterized in that, Based on motion separation results and medical prior constraints, the image restoration network performs inverse motion compensation on the image using the separated motion components, and further optimizes the image quality through adversarial training of the generative adversarial network. The reverse motion compensation can be implemented in the following ways: pixel-level resampling compensation based on optical flow field, deformation field compensation based on motion components, inter-frame interpolation compensation driven by motion components, or a combination of the above methods. The image restoration network adopts a generative adversarial network architecture, including a generator and an adversarial discriminator; The loss function of the generator is composed of a weighted average of image content reconstruction loss, perceptual loss, image quality perceptual loss, structural similarity loss, and medical prior loss. The adversarial discriminator is used to distinguish between the image generated by the image restoration network and the original clear image. The discriminator in the medical prior enhancement branch, which is composed of a pre-trained medical image diagnostic feature extraction network, is used to distinguish between the motion artifact component and the original clear image. The two discriminators constrain the image restoration network from the dimensions of image authenticity and diagnostic information integrity, respectively.

8. The method for suppressing motion artifacts and enhancing the quality of intraoperative imaging according to claim 7, characterized in that, The weights of each loss component are adaptively adjusted according to the intraoperative imaging modality, including but not limited to: modality-preset weight combination, weight calibration based on validation set grid search, automatic weight learning based on uncertainty learning, transfer weight based on task similarity, or a combination of the above methods.

9. The method for suppressing motion artifacts and enhancing the quality of intraoperative imaging according to claim 2, characterized in that, The temporal statistical features of each motion component output by the motion field decomposition module are used to construct an individualized motion feature descriptor. The individualized motion feature descriptor includes respiratory rate estimates, heart rate estimates, energy proportions of each motion component, and energy fluctuation amplitudes. When the image quality assessment index is lower than a preset threshold, and the deviation of the individualized motion feature descriptor from the motion feature distribution of the pre-training dataset exceeds a preset range, lightweight online fine-tuning of the physical-driven motion estimation branch and the motion field decomposition module is triggered. The lightweight online fine-tuning only updates the parameters of the last two layers of the motion estimation network, the gated recurrent unit parameters of the motion field decomposition module, and the parameters of the physiological cycle frequency estimation unit. A constraint term on the pre-training parameters is added to the fine-tuning loss to prevent catastrophic forgetting, enabling the bi-branch spatiotemporal adversarial network to adapt to specific surgical scenarios and the individualized motion characteristics of patients.

10. A device for suppressing motion artifacts and enhancing the quality of intraoperative imaging, comprising: At least one processing unit, the processing unit being connected to a storage unit via a bus unit, characterized in that the storage unit stores a computer program executable by the processing unit, and when the computer program is executed by the processing unit, it implements the intraoperative imaging motion artifact suppression and quality enhancement method according to any one of claims 1 to 9.

Images