A method for recognizing the frosting state of the target area of ​​a cryoablation needle based on image processing

Abstract

This invention discloses a method for identifying the frost state of the target area of ​​a cryoablation needle based on image processing, relating to the field of cryoablation needle technology. The method involves constructing an original image sequence of the target area in a unified spatiotemporal coordinate system; outputting corresponding degradation stage information; obtaining a high-fidelity clear image sequence of the target area; aggregating signal subspace features through a channel gating mechanism to obtain a signal enhancement feature map set; generating a multi-scale semantic feature tensor of the target area; outputting an initial frost probability distribution map of the target area; performing morphological constraint segmentation optimization to obtain a morphologically consistent frost region mask; obtaining a temporally stable frost region mask sequence and constructing a dynamic evolution model of the frost state, displaying warning information in real time. This invention effectively suppresses the feature submersion problem caused by mixed noise and significantly improves the recognizability and stability of the underlying structural information of the target area in a low-temperature imaging environment.

Description

Technical Field

[0001] This invention relates to the field of cryoablation needle technology, and in particular to a method for recognizing the frosting state of the target area of ​​a cryoablation needle based on image processing. Background Technology

[0002] With the widespread application of minimally invasive cryoablation technology in tumor treatment and precision surgery, how to perceive the extent of frost formation and its dynamic evolution in the target area in real time and accurately during surgery has become a key technical issue to ensure the safety and effectiveness of the procedure. In actual cryoablation, the visual information of the target area acquired by the imaging equipment is often under extreme low temperature conditions. It is affected by a combination of factors such as tissue fluid coagulation, phase transition refraction, and specular scattering of ice crystals, resulting in images exhibiting highly complex and dynamically changing mixed degradation characteristics over time.

[0003] In existing technologies, methods for dealing with medical image degradation problems mostly rely on image enhancement or denoising algorithms targeting a single degradation type, such as dehazing and denoising methods based on histogram equalization, Retinex models, or convolutional neural networks. These methods typically assume that the degradation type is relatively stable in space or time, making it difficult to cope with the dynamic mixed degradation phenomenon caused by the coupling of multiple physical mechanisms during cryoablation. When the surgery enters different stages, the image degradation characteristics change significantly, causing the original model to fail to generalize effectively, resulting in the loss of feature information or recognition failure. This causes the critical frosting boundary information to be masked by complex noise, seriously affecting the accuracy of subsequent analysis.

[0004] In terms of the identification and segmentation of frosted areas, most existing mainstream methods are based on pixel-level semantic segmentation networks of deep learning. They mainly rely on local grayscale, texture or semantic features for classification and discrimination, and lack the ability to model the biological heat transfer and phase transition physical laws during the freeze-thaw process. When the contrast between the soft tissue in the target area and the frosted area is extremely low, it is easy to have discontinuous segmentation results, internal voids or false detection of isolated patches in non-frosted areas, that is, there are obvious topological errors and non-physical morphological artifacts. Summary of the Invention

[0005] One objective of this invention is to propose a method for identifying the frosting state of the target area of ​​a cryoablation needle based on image processing. This invention effectively suppresses the feature submersion problem caused by mixed noise and significantly improves the recognizability and stability of the underlying structural information of the target area in a low-temperature imaging environment.

[0006] A method for identifying the frosting state of a cryoablation needle target area based on image processing according to an embodiment of the present invention includes:

[0007] During cryoablation surgery, multimodal raw visual data of the target area were collected, and time synchronization registration and spatial geometric correction were performed to construct a raw image sequence of the target area under a unified spatiotemporal coordinate system.

[0008] Degradation stage labels are generated based on the original image sequence of the target area, and the corresponding degradation stage information is output.

[0009] The original image sequence of the target area and the degradation stage information are input into the stage adaptive DiffUIR model to obtain a high-fidelity clear image sequence of the target area.

[0010] Reversible residual hourglass mapping is performed on high-fidelity target area clear image sequences to decompose the latent feature space into signal subspace and noise subspace. The signal subspace features are then aggregated through a channel gating mechanism to obtain a signal enhancement feature map set.

[0011] Input the signal enhancement feature map set into the weak feature multi-scale coding network to generate a multi-scale semantic feature tensor of the target region.

[0012] Input the multi-scale semantic feature tensor of the target area into the iterative topologically consistent segmentation and decoding network, and output the initial frost probability distribution map of the target area;

[0013] A morphological prior scheduler is constructed, and the connectivity constraints, curvature smoothing constraints, and volume growth constraints output by the morphological prior scheduler are applied to the initial frosting probability distribution map of the target area. Morphological constraint segmentation optimization is performed to obtain a morphologically consistent frosting region mask.

[0014] Optical flow reverse mapping comparison and anomaly fine-tuning are performed on the frost region mask with the frost region mask with the same morphology in the previous frame to obtain a temporally stable frost region mask sequence and construct a dynamic evolution model of frost state. The dynamic evolution model of frost state is then used to interact with the real-time surgical monitoring system to display early warning information in real time.

[0015] Optionally, the multimodal raw visual data of the target area includes synchronized endoscopic images, ultrasound images, optical coherence tomography data, and corresponding temperature-time curves.

[0016] Optionally, generating degradation stage labels based on the original image sequence of the target region includes:

[0017] For each frame of the original image of the target area, a degradation characterization quantity is extracted, and a degradation feature vector consisting of fogging response quantity, refraction enhancement response quantity, and ice crystal high reflectivity response quantity is constructed.

[0018] The input sequence of the degradation stage detection network is constructed using the degradation feature vectors of consecutive frames. The input sequence is then fed into the degradation stage detection network, which outputs the degradation stage prediction vector corresponding to the original image of the target area, which is constrained by the monotonicity of the unidirectional thermodynamic phase transition.

[0019] Based on the degradation stage prediction vector, degradation stage labels corresponding to the original image of the target area are generated. Temporal smoothing is then performed on the degradation stage labels to obtain smoothed degradation stage labels.

[0020] Based on the smoothed degradation stage label and degradation stage prediction vector, the corresponding degradation stage information is output.

[0021] Optionally, inputting the original image sequence of the target region along with the degradation stage information into the stage-adaptive DiffUIR model includes:

[0022] Based on the smoothed degradation stage label corresponding to the original image of the target region in frame t, a diffusion coefficient is constructed for the original image of the target region in frame t.

[0023] A forward diffusion process is performed on the original image of the target region in frame t based on the diffusion coefficient to generate a diffusion state sequence;

[0024] Based on the degradation stage information, noise estimation is performed on each diffusion state in the diffusion state sequence to generate the corresponding noise estimation results and cumulative retention coefficients.

[0025] Based on the degradation stage information of frame t, a noise reduction intensity adjustment coefficient is constructed in the reverse diffusion process;

[0026] Based on the noise estimation results, cumulative preservation coefficients, and denoising intensity adjustment coefficients, a reverse diffusion process is performed to obtain the clear image reconstruction result corresponding to the original image of the target region in frame t.

[0027] The reconstruction results of each clear image are output frame by frame to construct a high-fidelity clear image sequence of the target area.

[0028] Optionally, the reversible residual hourglass mapping processing of the high-fidelity target region sharp image sequence includes:

[0029] The sharp image reconstruction result in the high-fidelity target region sharp image sequence is input into the reversible residual hourglass mapping processing unit to generate the corresponding initial sharp feature map;

[0030] Perform reversible residual hourglass mapping on the initial clear feature map to obtain the corresponding bottleneck potential features;

[0031] Based on the bottleneck potential features, the potential feature space is decomposed to obtain signal subspace features and noise subspace features;

[0032] Channel gating coefficients are constructed based on signal subspace features and noise subspace features. The channel gating coefficients are then used to enhance the signal subspace features while suppressing the noise subspace features, thus obtaining gated signal features.

[0033] Hourglass decoding and reconstruction are performed on the gated signal features of frame t to obtain the signal enhancement feature set of frame t. The signal enhancement feature sets of each frame are output frame by frame to construct the total signal enhancement feature set.

[0034] Optionally, inputting the signal enhancement feature map set into the weak feature multi-scale coding network includes:

[0035] The signal enhancement feature map set is input into the weak feature multi-scale coding network, and fractional gradient response is extracted for the signal enhancement feature map at each scale level to obtain the phase transition microtexture features at the corresponding scale.

[0036] A spatial offset field is generated based on phase transition microtexture features, and deformable convolutional coding is performed on the signal enhancement feature map using the spatial offset field to obtain morphology-aware features that adaptively fit the ice crystal growth morphology.

[0037] Cross-scale nonlocal attention interaction is performed on morphological perception features at various scales, and global topological context and local micro-texture are fused to generate multi-scale semantic feature tensors for the target region.

[0038] Optionally, the step of inputting the multi-scale semantic feature tensor of the target region into the iterative topologically consistent segmentation and decoding network includes:

[0039] The multi-scale semantic feature tensor of the target area is input into an iterative topologically consistent segmentation and decoding network to perform cross-scale feature concatenation decoding and feature resolution restoration, resulting in full-resolution decoded features consistent with the spatial scale of the original image of the target area.

[0040] Extract the structural topology attributes of the full-resolution decoded features, construct a boundary-aware gating mechanism, perform topology feature correction on the full-resolution decoded features, and obtain topology-consistent boundary enhancement features;

[0041] The topological consistency boundary enhancement feature is mapped to a single-channel probability space to generate an initial frost probability distribution map of the target area.

[0042] Optionally, applying the connectivity constraints, curvature smoothing constraints, and volume growth constraints output by the morphological prior scheduler to the initial frosting probability distribution map of the target region includes:

[0043] A morphological prior scheduler is constructed, and a corresponding physical morphological prior field, as well as connectivity constraint coefficient, curvature smoothing constraint coefficient, and volume growth constraint coefficient, are generated based on the initial frost probability distribution map of the target area.

[0044] Based on the connectivity constraint coefficient and the physical morphology prior field, the initial frost probability distribution map of the target area is optimized by connectivity constraint to obtain the connectivity constraint probability map.

[0045] Based on the curvature smoothing constraint coefficient, curvature smoothing constraint optimization is performed on the connectivity constraint probability graph to obtain the curvature smoothing probability graph.

[0046] Based on the volume growth constraint coefficient, volume growth constraint optimization is performed on the curvature smoothing probability map to obtain a frosting region mask with consistent shape.

[0047] Frame by frame, output the frost region mask with consistent shape in each frame to construct a sequence of frost region masks with consistent shape.

[0048] Optionally, the step of performing optical flow reverse mapping comparison and anomaly fine-tuning on the morphologically consistent frost region mask and the morphologically consistent frost region mask of the previous frame to obtain a temporally stable frost region mask sequence and constructing a dynamic evolution model of frost state includes:

[0049] Optical flow reverse mapping comparison and anomaly fine-tuning are performed on the frost region mask with the frost region mask with the same shape as the previous frame to obtain a temporally stable frost region mask sequence.

[0050] Based on a time-stable frosting region mask sequence, a three-dimensional frosting state model is constructed, and a set of frosting quantification indicators for the target area is calculated.

[0051] A dynamic evolution model of frost state is generated based on the three-dimensional frost state model and the frost quantification index set. The dynamic evolution model of frost state is then used to interact with the real-time surgical monitoring system to perform monitoring, intervention and early warning.

[0052] Optionally, the monitoring, intervention, and early warning system includes:

[0053] When the frost expansion rate exceeds the preset safe expansion rate threshold, or the minimum spatial spacing is less than the preset safe distance threshold and spatial collision and overlap occur, a surgical monitoring early warning signal containing a cooling and decompression command or an equipment shutdown command is generated. The surgical monitoring early warning signal is sent to the surgical real-time monitoring system to trigger the corresponding emergency highlight rendering and forced equipment flow interruption intervention.

[0054] When the frost expansion rate is less than or equal to the preset safe expansion rate threshold, and the minimum spatial spacing is greater than or equal to the preset safe distance threshold, a rendering monitoring signal is generated and sent to the surgical real-time monitoring system to display the frost range and frost speed in real time.

[0055] The beneficial effects of this invention are:

[0056] This invention constructs a DiffUIR selective hourglass distribution mapping diffusion model based on degradation stage perception, achieving unified modeling and adaptive elimination of dynamic mixed optical degradation during cryoablation. Compared with traditional image enhancement methods that rely on a single degradation assumption, this invention uses a stage-adaptive diffusion coefficient to model differentiated noise for different degradation stages during forward diffusion, and dynamically controls the noise stripping intensity through a denoising intensity adjustment coefficient during reverse diffusion. This allows the model to continuously and adaptively switch between three degradation stages: fogging, refractive enhancement, and high reflectivity of ice crystals, avoiding the failure or blinding phenomenon that occurs when traditional algorithms switch stages. At the same time, through the selective hourglass distribution mapping mechanism, different degradation distributions are uniformly converged to a shared clear potential distribution space, effectively suppressing the feature submersion problem caused by mixed noise, and significantly improving the recognizability and stability of the underlying structural information of the target area in low-temperature imaging environments.

[0057] This invention achieves highly sensitive capture of early microscopic features of the water-ice phase transition through a reversible residual hourglass mapping and a fractional gradient-driven multi-scale encoding mechanism for weak features. In the latent feature space, signal and noise subspaces are decomposed, and channel gating coefficients are used to enhance the effective structural response and suppress residual noise. Anisotropic nonlinear microtexture amplification is applied to the signal enhancement feature map, significantly enhancing the weak grayscale perturbations and high-frequency structures caused by ice crystal nucleation in the feature space. Simultaneously, the gradient-guided spatial offset field dynamically adjusts the deformable convolution sampling position, ensuring the convolution receptive field aligns with the ice crystal growth direction. Even in low-contrast soft tissue backgrounds, it can accurately extract fine-grained structural features of the phase transition boundary, improving the detection sensitivity and feature representation capability of early frost regions.

[0058] This invention optimizes the physical consistency of segmentation results by constructing a morphological prior scheduler and introducing connectivity constraints, curvature smoothing constraints, and volume growth constraints. At the segmentation backend, a cryoablation biothermal transfer model and a priori geometric growth of ice balls are explicitly introduced. Spatial consistency constraints on the frost probability distribution are imposed through a physical morphological prior field. Main connected regions are selected based on connectivity scores to eliminate isolated false detection regions. Curvature smoothing responses are used to adaptively smooth boundaries to suppress jagged edges and sharp abrupt changes. Simultaneously, volume growth constraints limit the expansion of frost regions between adjacent time frames, ensuring that the segmentation results conform to actual thermodynamic laws in both spatial topology and temporal evolution, thereby improving the structural integrity and clinical reliability of frost region identification. Attached Figure Description

[0059] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0060] Figure 1 This is a flowchart of a method for recognizing the frosting state of a cryoablation needle target area based on image processing, as proposed in this invention.

[0061] Figure 2 This is a schematic diagram of the stage-adaptive DiffUIR model structure in the image processing-based cryoablation needle target area frost state recognition method proposed in this invention. Detailed Implementation

[0062] Example 1: Reference Figures 1-2 A method for recognizing the frosting state of the target area of ​​a cryoablation needle based on image processing, comprising:

[0063] During cryoablation surgery, multimodal raw visual data of the target area were collected, and time synchronization registration and spatial geometric correction were performed to construct a raw image sequence of the target area under a unified spatiotemporal coordinate system.

[0064] In this embodiment, the multimodal raw visual data of the target area includes synchronized endoscopic images, ultrasound images, optical coherence tomography data, and corresponding temperature-time curves.

[0065] Degradation stage labels are generated based on the original image sequence of the target area, and the corresponding degradation stage information is output.

[0066] In this embodiment, degradation stage labels are generated based on the original image sequence of the target region, including:

[0067] For each frame of the original image of the target area, a degradation characterization quantity is extracted, and a degradation feature vector consisting of fogging response quantity, refraction enhancement response quantity, and ice crystal high reflectivity response quantity is constructed.

[0068] In Example 1, the gradient magnitude of each pixel position is calculated in the pixel domain of the target area, the maximum gradient magnitude of the original image of the target area in the corresponding frame in the pixel domain of the target area is obtained, the gradient magnitude of all pixels in the pixel domain of the target area is divided by the maximum gradient magnitude and then summed, the summation result is divided by the total number of pixels in the target area and then subtracted by 1 to obtain the fogging response amount.

[0069] The absolute difference between the normalized gray value at each pixel location in the original image of the target region and the average normalized gray value of the corresponding local neighborhood is calculated. The absolute difference is divided by the average normalized gray value of the local neighborhood. The results of all pixels in the pixel domain of the target region are summed and divided by the total number of pixels in the target region to obtain the refractive enhancement response, which is used to represent the degree of brightness disturbance caused by local phase transition refraction during the cryoablation process.

[0070] A highlight indicator function is constructed in the original image of the target area. The function is set to 1 when the normalized gray value of the pixel position is greater than or equal to the ice crystal high reflectivity discrimination threshold, and 0 otherwise. At the same time, the Laplacian response value of each pixel position is calculated and the maximum Laplacian response value of the original image of the target area in the corresponding frame within the pixel domain of the target area is obtained. For the pixel position that satisfies the highlight indicator function of 1, its Laplacian response value is divided by the maximum Laplacian response value, and then weighted and summed. This summation is then divided by the total number of pixels in the target area to obtain the ice crystal high reflectivity response.

[0071] The input sequence of the degradation stage detection network is constructed using the degradation feature vectors of consecutive frames. The input sequence is then fed into the degradation stage detection network, which outputs the degradation stage prediction vector corresponding to the original image of the target area, which is constrained by the monotonicity of the unidirectional thermodynamic phase transition.

[0072] In Example 1, the degradation stage detection network includes a degradation feature embedding layer, a temporal correlation coding layer, a thermodynamic transfer matrix penalty layer, and a restricted classification output layer.

[0073] The extracted fogging response, refraction enhancement response, and ice crystal high reflectivity response are constructed into an input sequence according to the time series. After high-dimensional mapping and feature denoising through the degradation feature embedding layer, a continuous frame degradation representation sequence is obtained. The continuous frame degradation representation sequence is input into the temporal correlation coding layer, and the context dependency model is performed using the cyclic gating mechanism to output the unconstrained initial state distribution of the t-th frame.

[0074] Based on the irreversible law of unidirectional cooling in freeze-drying, a Markov transition matrix with non-zero elements on the main diagonal and adjacent upper elements, and a lower triangular penalty value approaching zero, is constructed in the thermodynamic transition matrix penalty layer as a unidirectional phase transition constraint matrix. This matrix is ​​then multiplied with the final state prediction vector of frame t-1 to generate a physical prior evolution distribution representing the evolution boundary at the current moment. The unconstrained initial state distribution and the physical prior evolution distribution are input into a restricted classification output layer for gated dot-multiplication and weighting. Temporal jump noise that violates the thermodynamic unidirectional cooling law is filtered out to obtain a weighted fusion feature. After normalization, the prediction probability of each degradation stage is generated, and the predicted vector of the degradation stage of frame t, constrained by unidirectional thermodynamic phase transition, is spliced ​​to output the result.

[0075] Based on the degradation stage prediction vector, degradation stage labels corresponding to the original image of the target area are generated. Temporal smoothing is then performed on the degradation stage labels to obtain smoothed degradation stage labels.

[0076] In Example 1, the degradation stage label is obtained by selecting the category with the largest value among the predicted probability of fogging stage, the predicted probability of refraction enhancement stage, and the predicted probability of ice crystal high reflectivity stage. When the selected category is fogging stage, the corresponding label value is 1; when the selected category is refraction enhancement stage, the corresponding label value is 2; and when the selected category is ice crystal high reflectivity stage, the corresponding label value is 3.

[0077] Based on the smoothed degradation stage label and degradation stage prediction vector, the corresponding degradation stage information is output.

[0078] In Example 1, the degradation stage information consists of a smoothed degradation stage label and a degradation stage confidence score. The degradation stage confidence score is obtained by selecting the maximum value among the prediction probabilities of the fogging stage, the refraction enhancement stage, and the ice crystal high reflectivity stage. The degradation stage information is used to represent the degradation state of the original image of the target area in the t-th frame in the current cryoablation stage and its discrimination reliability.

[0079] The original image sequence of the target area and the degradation stage information are input into the stage adaptive DiffUIR model to obtain a high-fidelity clear image sequence of the target area.

[0080] In this embodiment, the original image sequence of the target region and the degradation stage information are input together into the stage-adaptive DiffUIR model, including:

[0081] Based on the smoothed degradation stage label corresponding to the original image of the target region in frame t, a diffusion coefficient is constructed for the original image of the target region in frame t.

[0082] In Example 1, the stage-adaptive DiffUIR model includes a conditional input receiving unit, a forward diffusion processing unit, a selective hourglass distribution mapping denoising unit, and a reverse diffusion reconstruction unit.

[0083] The conditional input receiving unit is used to receive and construct the joint input of the original image of the target region in frame t and the degradation stage information in frame t, which is the conditional input of the stage adaptive DiffUIR model.

[0084] In the forward diffusion processing unit, based on the smoothed degradation stage label corresponding to the original image of the target area in frame t, the corresponding diffusion coefficient sequence is selected from the pre-set diffusion coefficient sequence of the fogging stage, the diffusion coefficient sequence of the refraction enhancement stage, and the diffusion coefficient sequence of the ice crystal high reflectivity stage. The selected diffusion coefficient sequence is used as the diffusion coefficient of the original image of the target area in frame t throughout the diffusion process, which is used to control the noise injection intensity of the original image of the target area in frame t at each step of the diffusion process.

[0085] ;

[0086] in, This represents the diffusion coefficient of the original image of the target region in frame t at the diffusion state in step n, used to control the intensity of noise injected into the image during this diffusion step. This represents the diffusion coefficient corresponding to the fogging stage in the nth diffusion state, and is used to model the noise injection intensity when performing noise perturbation modeling on the original image of the target region in the tth frame under fogging degradation conditions. This represents the diffusion coefficient corresponding to the refractive enhancement stage in the nth diffusion state, and is used to model the noise injection intensity when performing noise perturbation modeling on the original image of the target region in the nth frame under refractive enhancement degradation conditions. This represents the diffusion coefficient corresponding to the high reflectivity stage of ice crystals in the nth diffusion step. It is used to model the noise injection intensity when performing noise perturbation modeling on the original image of the target region in the nth frame under the condition of high reflectivity degradation of ice crystals. This represents the indicative function, used to select the corresponding diffusion coefficient based on the degradation stage label. This represents the smoothed degradation stage label corresponding to the original image of the target region in the i-th frame. This represents the time frame index in the original image sequence of the target region. This represents the step index in the diffusion process, used to identify the diffusion state at step n.

[0087] A forward diffusion process is performed on the original image of the target region in frame t based on the diffusion coefficient to generate a diffusion state sequence;

[0088] In Example 1, the original image of the target region in frame t is used as the initial state. Random noise conforming to a Gaussian distribution is gradually superimposed onto the original image of the target region in frame t. The noise intensity is controlled according to the diffusion coefficient in each superposition process, so that the original image of the target region in frame t is gradually transformed into a diffusion state image under different noise levels, and a diffusion state sequence from the initial state to the maximum noise state is obtained.

[0089] Based on the degradation stage information, noise estimation is performed on each diffusion state in the diffusion state sequence to generate the corresponding noise estimation results and cumulative retention coefficients.

[0090] In Example 1, the selective hourglass distribution mapping denoising unit includes a shallow feature extraction module, a stage information embedding module, an hourglass encoding module, a bottleneck mapping module, an hourglass decoding module, and a noise prediction output module.

[0091] The shallow feature extraction module is used to perform convolution mapping on the nth diffusion state to obtain the shallow diffusion features corresponding to the nth diffusion state.

[0092] The stage information embedding module is used to jointly embed and encode the diffusion step number and the degradation stage information of frame t to obtain stage conditional embedding features. The degradation stage information of frame t is composed of the smoothed degradation stage label corresponding to the original image of the target region of frame t and the degradation stage confidence corresponding to the original image of the target region of frame t.

[0093] The hourglass coding module is used to perform stepwise downsampling and multi-scale feature aggregation on shallow diffusion features to obtain a multi-scale diffusion representation under compressed conditions.

[0094] The bottleneck mapping module is used to fuse multi-scale diffusion representation with stage-conditional embedding features in a compressed state. By sharing the distribution alignment term, it forces the feature tensors under different physical degradation stages to be mapped and constrained to a general latent prior distribution space that represents a clear organizational structure, thereby eliminating the spatial domain shift caused by dynamic hybrid degradation and obtaining stage-selected and distribution-aligned bottleneck mapping features.

[0095] The hourglass decoding module is used to upsample the bottleneck mapping features after stage selection and distribution alignment, and to perform skip connection fusion with the features of the corresponding scale in the hourglass encoding module to obtain multi-scale reconstruction features in the restored state.

[0096] The noise prediction output module is used to output the noise estimation result corresponding to the nth step diffusion state based on the multi-scale reconstruction features under the recovery state.

[0097] For the nth diffusion state in the diffusion state sequence, extract the diffusion coefficients corresponding to each step from the 1st diffusion state to the nth diffusion state. Subtract 1 from each diffusion coefficient and multiply them continuously in the order of diffusion steps to obtain the cumulative retention coefficient corresponding to the nth diffusion state.

[0098] Based on the degradation stage information of frame t, a noise reduction intensity adjustment coefficient is constructed in the reverse diffusion process;

[0099] In Example 1, based on the smoothed degradation stage label corresponding to the original image of the target region in frame t and the degradation stage confidence level corresponding to the original image of the target region in frame t, a corresponding denoising adjustment amount is selected from the preset denoising adjustment amounts for the fogging stage, the refraction enhancement stage, and the ice crystal high reflectivity stage. The selected denoising adjustment amount is then weighted and adjusted in conjunction with the degradation stage confidence level corresponding to the original image of the target region in frame t to obtain the denoising intensity adjustment coefficient for the corresponding diffusion state, which is used to control the denoising intensity of the original image of the target region in frame t during the reverse diffusion process.

[0100] ;

[0101] in, This represents the noise intensity adjustment coefficient. This represents the noise reduction intensity adjustment amount corresponding to the fogging stage in the nth diffusion state. This represents the amount of noise reduction intensity adjustment corresponding to the refraction enhancement stage in the nth diffusion state. This represents the noise reduction intensity adjustment amount corresponding to the high reflectivity stage of ice crystals in the nth diffusion state. This represents the confidence level of the degradation stage corresponding to the original image of the target region in the i-th frame.

[0102] Based on the noise estimation results, cumulative preservation coefficients, and denoising intensity adjustment coefficients, a reverse diffusion process is performed to obtain the clear image reconstruction result corresponding to the original image of the target region in frame t.

[0103] In Example 1, the reverse diffusion reconstruction unit starts from the state with the maximum noise in the diffusion state sequence and gradually removes the noise components in each diffusion state. In each step of the reverse denoising calculation, the denoising intensity adjustment coefficient is directly coupled as a scaling factor into the mean derivation of the reverse diffusion, so that the denoised component predicted by the noise estimation result is scaled proportionally, controlling the noise cancellation intensity for different physical degradation stages. At the same time, the denoising intensity adjustment coefficient is introduced into the variance calculation of the reverse diffusion, which is the mean of the original image of the target region in frame t when reverse diffusion is performed in the nth diffusion state. Represented as:

[0104] ;

[0105] in, This represents the mean of the inverse-diffusion Gaussian distribution, used to determine the desired center of the denoised image state. This represents the image representation of the diffusion state at step n. This represents the corresponding cumulative retention coefficient. This represents the noise estimation result output by the selective hourglass distribution mapping denoising unit.

[0106] Noise reduction intensity adjustment coefficient This constitutes an external posterior-guided gating mechanism independent of the network's internal feature embedding. Due to the cognitive uncertainty in feature extraction under extremely low temperature conditions, the stage information embedding within the network is used to predict the direction of degradation evolution, while the external denoising intensity adjustment coefficient... This is used to control the denoising step size in this direction based on the confidence level of the degradation stage, when the network outputs noise estimation results based on low-confidence features. At that time, the noise reduction intensity adjustment coefficient By using a conservative coefficient value approaching 1, non-physical topological illusions caused by network over-extension are forcibly suppressed at the sampling level, while at high confidence levels, the denoising intensity adjustment coefficient... It acts as an extrapolation amplification factor to accelerate the removal of specific physical degradation noise, thus achieving logical decoupling between internal direction prediction and external confidence step size control.

[0107] The variance of the original target region image in frame t when reverse diffusion is performed at the nth diffusion step. Represented as:

[0108] ;

[0109] in, This represents the variance of the backward-diffusion Gaussian distribution. This represents the cumulative retention coefficient in the (n-1)th diffusion state. The squared term representing the denoising intensity adjustment coefficient is used to scale the range of random perturbations in the generation process in accordance with the deterministic synchronization of the front-end classification.

[0110] Combined with mean With variance The image state at step n-1 is obtained by sampling from a Gaussian distribution. Through the By iterating backwards from the maximum diffusion step number down to 1, a clear image reconstruction result corresponding to the original image of the target region in frame t is obtained.

[0111] The reconstruction results of each clear image are output frame by frame to construct a high-fidelity clear image sequence of the target area.

[0112] The staged adaptive DiffUIR model proposed in this embodiment is structurally composed of a conditional input receiving unit, a forward diffusion processing unit, a selective hourglass distribution mapping denoising unit, and a reverse diffusion reconstruction unit cascaded sequentially. Compared with the existing standard DiffUIR model, the improvements and significant advantages of the staged adaptive DiffUIR model are reflected in three dimensions:

[0113] First, a physically conditional dynamic noise scheduling mechanism is introduced into the forward diffusion layer. By switching the diffusion coefficient in real time according to the specific physical degradation label, the stage-adaptive DiffUIR model can take an independent Markov evolution trajectory for different thermodynamic degradation states such as fogging, refraction, and high reflectivity, thus overcoming the limitations of fixed noise scheduling in conventional diffusion models.

[0114] Second, by innovatively embedding a shared distribution alignment term at the bottleneck layer, the multi-source hybrid degradation characteristics are forcibly constrained to a unified and clearly organized potential prior space, fundamentally eliminating the spatial domain shift caused by dynamic hybrid degradation at extreme low temperatures.

[0115] Third, a posterior-guided gating mechanism based on confidence decoupling was constructed in the inverse reconstruction layer. The denoising intensity adjustment coefficient was directly coupled into the mean and variance of the inverse sampling, realizing the mathematical decoupling between the internal direction prediction of the network and the external denoising step size control. This can accelerate the stripping of specific physical noise at high confidence and cause the model to degenerate into a conservative denoising state at low confidence, thus completely suppressing the non-physical topological illusion that pure data-driven models are prone to generate at the edges of soft tissues with extremely low contrast.

[0116] Reversible residual hourglass mapping is performed on high-fidelity target area clear image sequences to decompose the latent feature space into signal subspace and noise subspace. The signal subspace features are then aggregated through a channel gating mechanism to obtain a signal enhancement feature map set.

[0117] In this embodiment, reversible residual hourglass mapping processing is performed on the high-fidelity target region clear image sequence, including:

[0118] The sharp image reconstruction result in the high-fidelity target region sharp image sequence is input into the reversible residual hourglass mapping processing unit to generate the corresponding initial sharp feature map;

[0119] In Example 1, the reversible residual hourglass mapping processing unit includes an input feature mapping module, a reversible residual hourglass encoding module, a latent feature decomposition module, a channel gating aggregation module, and a signal enhancement output module.

[0120] The input feature mapping module maps the sharp image reconstruction result pixel by pixel into a high-dimensional feature space. By performing feature extraction and channel recombination on the response values ​​of all pixel positions, an initial sharp feature map corresponding one-to-one with the sharp image reconstruction result is obtained.

[0121] Perform reversible residual hourglass mapping on the initial clear feature map to obtain the corresponding bottleneck potential features;

[0122] In Example 1, the reversible residual hourglass coding module is used to perform stepwise downsampling and reversible residual mapping on the initial clear feature map. In each scale level, the features at the current scale are divided into first branch features and second branch features along the channel dimension. The second branch features are nonlinearly mapped using the first residual mapping function and superimposed on the first branch features to obtain the first branch updated features. The first branch updated features are nonlinearly mapped using the second residual mapping function and superimposed on the second branch features to obtain the second branch updated features. The first branch updated features and the second branch updated features are concatenated by channels. The concatenated features are spatially compressed using downsampling mapping to obtain the compressed state features of the next scale level. By sequentially performing reversible residual hourglass mapping on all scale levels, the compressed state features are obtained at the deepest scale level, which serve as the bottleneck potential features corresponding to the initial clear feature map.

[0123] Based on the bottleneck potential features, the potential feature space is decomposed to obtain signal subspace features and noise subspace features;

[0124] In Example 1, the latent feature decomposition module is used to generate a signal decomposition mask for the bottleneck latent features. By performing nonlinear mapping and normalization on the response of the bottleneck latent features at each channel position, a signal decomposition mask with a value range of 0 to 1 is obtained. By inverting the signal decomposition mask element by element, a noise decomposition mask is obtained, which is used to isolate the background components containing the low-frequency response of homogeneous soft tissue and residual thermodynamic disturbances in the feature space.

[0125] By multiplying the bottleneck potential features element-wise with the signal decomposition mask, the signal subspace features are obtained, thus forcibly stripping the weak phase transition background tensor from the complex organizational background. Similarly, by multiplying the bottleneck potential features element-wise with the noise decomposition mask, the noise subspace features are obtained.

[0126] Channel gating coefficients are constructed based on signal subspace features and noise subspace features. The channel gating coefficients are then used to enhance the signal subspace features while suppressing the noise subspace features, thus obtaining gated signal features.

[0127] In Example 1, the channel gating aggregation module calculates the mean global spatial response of the signal subspace features and noise subspace features in each channel, respectively, and obtains the statistical descriptor of the signal subspace channel and the statistical descriptor of the noise subspace channel. The two are then subjected to joint linear mapping and normalization by gating weights and biases to generate channel gating coefficients.

[0128] Hourglass decoding and reconstruction are performed on the gated signal features of frame t to obtain the signal enhancement feature set of frame t. The signal enhancement feature sets of each frame are output frame by frame to construct the total signal enhancement feature set.

[0129] In Example 1, the signal enhancement output module is used to perform upsampling processing on the gated signal features step by step to restore them to the current scale resolution. After channel splicing and fusion with the compressed state features at the corresponding scale level, nonlinear mapping is performed to obtain the signal enhancement feature map of the current scale level. By restoring from the deepest scale level to the shallowest scale level step by step, a set of signal enhancement feature maps covering all scale levels is obtained.

[0130] Input the signal enhancement feature map set into the weak feature multi-scale coding network to generate a multi-scale semantic feature tensor of the target region.

[0131] In this embodiment, the signal enhancement feature map set is input into the weak feature multi-scale coding network, including:

[0132] The signal enhancement feature map set is input into the weak feature multi-scale coding network, and fractional gradient response is extracted for the signal enhancement feature map at each scale level to obtain the phase transition microtexture features at the corresponding scale.

[0133] In Example 1, the weak feature multi-scale coding network includes a fractional gradient sensing module, a gradient-guided deformable coding module, and a multi-scale semantic aggregation module.

[0134] The fractional gradient sensing module is used to capture extremely subtle grayscale gradients and texture abrupt changes in the early stage of the water-ice phase transition within the cryoablation target area. Fractional partial differential operators are used to extract anisotropic textures from the signal enhancement feature map and calculate the phase transition micro-texture features corresponding to the signal enhancement feature map.

[0135] A spatial offset field is generated based on phase transition microtexture features, and deformable convolutional coding is performed on the signal enhancement feature map using the spatial offset field to obtain morphology-aware features that adaptively fit the ice crystal growth morphology.

[0136] In Example 1, the gradient-guided deformable coding module is used to overcome the defect that the fixed receptive field of conventional standard convolution cannot capture the irregular dendrite growth of ice crystals.

[0137] Phase transition microtexture features are input into offset prediction convolutional layers to learn and generate spatial offset fields at corresponding scales. These spatial offset fields are then applied to the sampling grid of standard convolutions, and deformable convolution operations are performed on the signal enhancement feature maps. The feature response values ​​of morphological perception features at specific spatial locations are calculated to mathematically bind the deformation of the receptive field to the micro-gradient direction of ice crystal growth, forcing the network to extract features along the cryothermodynamic penetration path, thereby achieving accurate morphological perception in the early stages of phase transition.

[0138] Cross-scale nonlocal attention interaction is performed on morphological perception features at various scales, and global topological context and local micro-texture are fused to generate multi-scale semantic feature tensors for the target region.

[0139] In Example 1, the multi-scale semantic aggregation module is used to perform channel dimensionality reduction and scale alignment on morphological perception features of different depths.

[0140] Linear feature mapping and spatial dimension flattening are performed on deep low-resolution features to obtain the query feature sequence; linear feature mapping and spatial dimension flattening are performed on shallow high-resolution features to obtain the key feature sequence and value feature sequence; the matrix inner product between the transpose of the query feature sequence and the key feature sequence is calculated to obtain the cross-scale feature inner product matrix; exponential normalization is performed on each spatial position dimension of the cross-scale feature inner product matrix to obtain the global similarity affinity matrix; the morphological perception features are weighted and aggregated using the global similarity affinity matrix to obtain the weighted aggregated features; the weighted aggregated features at each scale are concatenated to output the multi-scale semantic feature tensor of the target area, which is used to represent the tissue-ice crystal mixing state in different thermodynamic evolution stages in the target area at the current time.

[0141] From the morphological perception features at each scale, the morphological perception features corresponding to the scale level with the largest spatial resolution are selected as shallow high-resolution features, and the morphological perception features corresponding to the scale level with the smallest spatial resolution are selected as deep low-resolution features.

[0142] Input the multi-scale semantic feature tensor of the target area into the iterative topologically consistent segmentation and decoding network, and output the initial frost probability distribution map of the target area;

[0143] In this embodiment, the multi-scale semantic feature tensor of the target region is input into an iterative topologically consistent segmentation and decoding network, including:

[0144] The multi-scale semantic feature tensor of the target area is input into an iterative topologically consistent segmentation and decoding network to perform cross-scale feature concatenation decoding and feature resolution restoration, resulting in full-resolution decoded features consistent with the spatial scale of the original image of the target area.

[0145] In Example 1, the iterative topology consistent segmentation decoding network includes a multi-stage cascaded decoding module, a topology boundary awareness module, and a probability distribution mapping module.

[0146] The multi-stage cascaded decoding module splits the multi-scale semantic feature tensor of the target region along the scale dimension, extracts scale features of different spatial resolutions, and starts from the scale feature with the smallest spatial resolution. It performs bilinear interpolation upsampling on the current scale feature to double its spatial resolution. The upsampled feature is then concatenated with the adjacent high spatial resolution scale features in the multi-scale semantic feature tensor of the target region. The concatenated feature is then subjected to depthwise separable convolution operation for feature smoothing and dimensionality reduction. The output of the depthwise separable convolution operation is used as the input of the next scale level.

[0147] Repeatedly perform upsampling, channel concatenation, and depth-separable convolution operations until the spatial resolution of the features is restored to the maximum spatial resolution of the original image of the target region, to obtain the decoding feature representations at each level, and use the decoding feature representation corresponding to the maximum spatial resolution as the full-resolution decoding feature.

[0148] Extract the structural topology attributes of the full-resolution decoded features, construct a boundary-aware gating mechanism, perform topology feature correction on the full-resolution decoded features, and obtain topology-consistent boundary enhancement features;

[0149] In Example 1, the topology boundary perception module performs morphological dilation and morphological erosion operations on the full-resolution decoded features in the spatial dimension, and erodes the features. The dilated features and eroded features are subtracted element-wise to obtain the topology boundary prior features. The topology boundary prior features are channel-wise concatenated and fused with the full-resolution decoded features. The fused features are then processed sequentially by channel global average pooling, multilayer perceptron linear mapping, and logistic regression activation function to obtain topology correction gating weights. The full-resolution decoded features are multiplied element-wise with the topology correction gating weights to obtain topology consistency boundary enhancement features. These features are used to correct the jagged distortion of the edge of the frost feature in the target area and the internal isolated holes, ensuring the physical connectivity and consistency of the frost feature response in space.

[0150] The topological consistency boundary enhancement feature is mapped to a single-channel probability space to generate an initial frost probability distribution map of the target area.

[0151] In Example 1, the probability distribution mapping module performs pixel-by-pixel linear classification mapping on the topological consistency boundary enhancement features, compresses the multi-channel high-dimensional features into single-channel classification prediction features, performs non-linear normalization activation function processing on each spatial pixel position of the single-channel classification prediction features in the spatial dimension, strictly maps and constrains the feature response value of each spatial pixel position to a continuous probability interval of zero to one, obtains the frost prediction probability value of each spatial pixel position, and arranges the frost prediction probability values ​​of all spatial pixel positions according to the original spatial coordinate domain to generate the initial frost probability distribution map of the target area.

[0152] A morphological prior scheduler is constructed, and the connectivity constraints, curvature smoothing constraints, and volume growth constraints output by the morphological prior scheduler are applied to the initial frosting probability distribution map of the target area. Morphological constraint segmentation optimization is performed to obtain a morphologically consistent frosting region mask.

[0153] In this embodiment, the connectivity constraints, curvature smoothing constraints, and volume growth constraints output by the morphological prior scheduler are applied to the initial frosting probability distribution map of the target region, including:

[0154] A morphological prior scheduler is constructed, and a corresponding physical morphological prior field, as well as connectivity constraint coefficient, curvature smoothing constraint coefficient, and volume growth constraint coefficient, are generated based on the initial frost probability distribution map of the target area.

[0155] In Example 1, the morphological prior scheduler generates various constraint parameters by performing spatial statistical analysis, geometric structure fitting, and physical consistency evaluation on the initial frosting probability distribution map of the target area. Specifically, it extracts the real-time pose information of the cryoablation needle and projects its three-dimensional spatial coordinates onto the two-dimensional pixel space of the original image of the target area to obtain the pixel position of the ablation needle tip, which is used as the center of the physical target point.

[0156] Using the center of the physical target as the reference point for the diffusion of the physical thermodynamic source, and combining the axial and radial growth scales output by the cryoablation biological heat transfer physical model, the spatial distance from each pixel position in the target region to the center of the physical target is calculated. The spatial distance is then subjected to directional normalization attenuation processing based on the axial and radial growth scales to obtain the physical consistency response value of each pixel position. The physical consistency response values ​​of all pixel positions constitute the physical morphological prior field.

[0157] The cryoablation biological heat transfer physical model is used to calculate the temperature distribution of each pixel position in the target area during the freezing process. The specific calculation method is as follows: the temperature value of the synchronized temperature-time curve in the current time frame is extracted and set as the cold source boundary temperature of the central pixel where the ablation needle is located. By performing spatial diffusion iteration on the temperature values ​​of all pixel positions in the target area in the current time frame, the corresponding temperature field distribution is obtained. Based on the temperature field distribution, the temperature values ​​of all pixel positions are compared with the preset freezing point temperature threshold. Pixel positions with temperatures lower than or equal to the freezing point temperature threshold are marked as icing candidate areas to obtain the spatial distribution of the icing area. The spatial distribution of the icing area is then subjected to geometric morphology fitting. Using the spatial coordinates of all pixel positions in the icing area as input, the least squares ellipsoid fitting method is used to fit the overall morphology of the icing area to obtain the corresponding ellipsoidal structure parameters. Based on the ellipsoidal structure parameters, the maximum extension length along the main direction is extracted as the axial growth scale, and the average extension length in the direction perpendicular to the main direction is used as the radial growth scale. These are used as the axial growth scale and radial growth scale in the current time frame, respectively.

[0158] The frosting prediction probability value of each pixel location is coupled with the physical consistency response value of the corresponding pixel location on a pixel-by-pixel basis, and the coupling results of all pixel locations are normalized and statistically analyzed to obtain the connectivity constraint coefficient.

[0159] The second-order variation intensity of the frosting probability distribution in the spatial neighborhood is statistically analyzed, and the variation intensity is coupled with the physical consistency response value of the corresponding pixel position. The coupling result is normalized and statistically analyzed to obtain the curvature smoothing constraint coefficient.

[0160] The frosting prediction probability value of each pixel location is coupled with the physical consistency response value of the corresponding pixel location pixel by pixel. Combined with the frosting area volume information of the previous frame, the coupling result is normalized and statistically analyzed to obtain the volume growth constraint coefficient.

[0161] Based on the connectivity constraint coefficient and the physical morphology prior field, the initial frost probability distribution map of the target area is optimized by connectivity constraint to obtain the connectivity constraint probability map.

[0162] In Example 1, pixel positions greater than or equal to a preset threshold are extracted from the initial frost probability distribution map of the target area, an initial candidate frost region mask is constructed, connected component decomposition is performed on the initial candidate frost region mask, and the average product of the frost prediction probability value and the physical morphology prior response value in each connected region is calculated as the connectivity score.

[0163] The connected region with the highest connectivity score is selected as the main connected region. Pixel positions belonging to the main connected region are retained, while pixel positions not belonging to the main connected region are suppressed. For each pixel position, its frost prediction probability value is updated by weighting according to the connectivity constraint coefficient, so that the frost prediction probability value in the main connected region is enhanced, and the frost prediction probability value in other regions is reduced according to the connectivity constraint coefficient, thus obtaining the connectivity constraint probability map.

[0164] Based on the curvature smoothing constraint coefficient, curvature smoothing constraint optimization is performed on the connectivity constraint probability graph to obtain the curvature smoothing probability graph.

[0165] In Example 1, the predicted frosting probability value of each pixel position in the connectivity constraint probability map is compared with the frosting probability discrimination threshold to obtain the connectivity candidate frosting region mask.

[0166] For the connectivity candidate frosting region mask, the discrete Laplacian operator is calculated in the spatial neighborhood. For each pixel position, the absolute value of the second difference in its neighborhood is obtained to obtain the curvature response value of the corresponding pixel position. For the curvature response value of each pixel position, an exponential decay mapping is performed so that the position with the large curvature response value corresponds to the small smoothing weight, and the position with the small curvature response value corresponds to the large smoothing weight, thus obtaining the curvature smoothing response.

[0167] For each pixel location in the connectivity constraint probability map, the average of the frost prediction probability values ​​of all pixels in its local neighborhood is calculated to obtain the local mean of the pixel location. The difference between the frost prediction probability value of each pixel location and its local mean is weighted and adjusted according to the curvature smoothing response. The adjustment result is then globally adjusted based on the curvature smoothing constraint coefficient to make the frost prediction probability values ​​in the boundary region approach their local mean, thus achieving boundary smoothing and obtaining the curvature smoothed probability map.

[0168] ;

[0169] in, Indicates the first Frame pixel position Curvature smoothing probability plot at the location. Indicates the first Frame pixel position The probability value of connectivity constraints at a given location. Indicates the first Frame curvature smoothing constraint coefficients Indicates the first Frame pixel position Curvature smoothing response value at the point, Indicates the first Frame pixel position The average probability value of the connectivity constraint probability graph at a given location within its local neighborhood. Represents the spatial position coordinates within the pixel domain of the target region. This represents the time frame index in the original image sequence of the target region.

[0170] Based on the volume growth constraint coefficient, volume growth constraint optimization is performed on the curvature smoothing probability map to obtain a frosting region mask with consistent shape.

[0171] In Example 1, when t=1, the frost region mask with consistent shape in the previous frame is initialized to an all-zero mask. When t>1, the frost region mask with consistent shape in the (t-1)th frame is used as the frost region mask with consistent shape in the previous frame. The mask values ​​of all pixel positions in the frost region mask with consistent shape in the previous frame are summed to obtain the frost region volume of the previous frame. The positions of all pixels that meet the frost probability discrimination threshold in the curvature smoothing probability map of the current frame are statistically analyzed to obtain the estimated frost region volume of the current frame.

[0172] The maximum theoretical growth rate allowed by physics in the current frame is calculated. The maximum theoretical growth rate is obtained by multiplying the volume growth constraint coefficient by the global average response value of the physical morphology prior field. The initial nucleation reference volume is extracted. The initial nucleation reference volume is used to characterize the instantaneous icing physical scale caused by the rapid cooling of the needle tip in the initial stage of freeze-thaw. The initial nucleation reference volume is obtained by counting the pixel positions in the physical morphology prior field whose response value is greater than the preset physical effective threshold. It is determined whether the volume of the frost region in the previous frame is zero. When the volume of the frost region in the previous frame is zero, the initial nucleation reference volume is used as the maximum incremental volume. When the volume of the frost region in the previous frame is greater than zero, the volume of the frost region in the previous frame is multiplied by the maximum theoretical growth rate to obtain the maximum incremental volume. The volume of the frost region in the previous frame is added to the maximum incremental volume to obtain the upper limit of the allowed growth volume in the current frame. The estimated volume of the frost region in the current frame is compared with the upper limit of the allowed growth volume, and the smaller value between the two is selected as the target retained volume.

[0173] To maintain the physical topological connectivity of the frosting region while limiting the volume, the outer contour boundary pixels of the connected pixel regions that meet the frosting probability discrimination threshold in the curvature smoothing probability map are extracted. The volume constraint response value corresponding to each boundary pixel is calculated. The volume constraint response value is obtained by multiplying the frosting prediction probability value of the pixel with the physical shape prior response value. The total number of pixels in the current connected pixel region is compared with the target retained volume. When the total number of pixels is greater than the target retained volume, the state of the outer contour boundary pixels is reversed from the frosting region to the non-frosting region in order of increasing volume constraint response value. The outer contour boundary is dynamically updated until the total number of pixels in the connected pixel region equals the target retained volume. The region is then marked as the frosting region, and the remaining pixel positions are marked as non-frosting regions, resulting in a frosting region mask with consistent shape.

[0174] Frame by frame, output the frost region mask with consistent shape in each frame to construct a sequence of frost region masks with consistent shape.

[0175] Optical flow reverse mapping comparison and anomaly fine-tuning are performed on the frost region mask with the frost region mask with the same morphology in the previous frame to obtain a temporally stable frost region mask sequence and construct a dynamic evolution model of frost state. The dynamic evolution model of frost state is then used to interact with the real-time surgical monitoring system to display early warning information in real time.

[0176] In this embodiment, optical flow reverse mapping comparison and anomaly fine-tuning are performed on the morphologically consistent frost region mask and the morphologically consistent frost region mask of the previous frame to obtain a temporally stable frost region mask sequence and construct a dynamic evolution model of the frost state, including:

[0177] Optical flow reverse mapping comparison and anomaly fine-tuning are performed on the frost region mask with the frost region mask with the same shape as the previous frame to obtain a temporally stable frost region mask sequence.

[0178] In Example 1, the dense optical flow field between the morphologically consistent frosting region mask of the current time frame and the temporally stable frosting region mask of the previous time frame is calculated. The dense optical flow field is used to perform spatial distortion reverse mapping on the morphologically consistent frosting region mask of the current time frame to obtain the forward prediction mask.

[0179] The forward prediction mask is compared with the temporally stable frosting region mask of the previous time frame by performing a pixel-wise intersection-union (IoU) calculation to obtain the inter-frame topological consistency score. The inter-frame topological consistency score is compared with a preset consistency tolerance threshold. When the inter-frame topological consistency score is lower than the preset consistency tolerance threshold, the non-overlapping difference region between the morphologically consistent frosting region mask of the current time frame and the temporally stable frosting region mask of the previous time frame is extracted. The state of all pixel positions in the non-overlapping difference region is forcibly reversed to a non-frosting region, triggering a fine-tuning loop to complete the mask update. The updated mask is used as the temporally stable frosting region mask of the current time frame.

[0180] When the inter-frame topology consistency score is greater than or equal to the preset consistency tolerance threshold, the original pixel state of the morphologically consistent frosting region mask of the current time frame remains unchanged, and it is directly used as the temporally stable frosting region mask of the current time frame. The masks are then stitched together in chronological order to construct a temporally stable frosting region mask sequence.

[0181] Based on a time-stable frosting region mask sequence, a three-dimensional frosting state model is constructed, and a set of frosting quantification indicators for the target area is calculated.

[0182] In Example 1, the real-time three-dimensional pose information of the cryoablation needle and the spatial calibration parameters of the medical imaging probe are extracted. The spatial calibration parameters are used to inversely project the temporally stable frosting region mask of the current time frame from the two-dimensional pixel coordinate system to the three-dimensional physical coordinate system with the cryoablation needle as the origin. Combined with the ellipsoidal structure parameters in the physical morphology prior field, the inversely projected mask is reconstructed by three-dimensional rotation and sweep along the axis of the cryoablation needle to generate a three-dimensional frosting state model.

[0183] Spatial volume integration is performed on the three-dimensional voxels contained in the three-dimensional frosting state model to obtain the frosting volume. The shortest spatial Euclidean distance from the vertices of the three-dimensional surface mesh of the frosting state model to the surface of the cryoablation needle is calculated. The maximum distance and the distance distribution gradient are extracted to form the frosting thickness distribution. The spatial difference between the frosting volume of the current time frame and the frosting volume of the previous time frame is calculated. The spatial difference is divided by the time interval between adjacent time frames to obtain the frosting expansion rate. The frosting volume, frosting thickness distribution and frosting expansion rate are combined to obtain the frosting quantification index set.

[0184] A dynamic evolution model of frost state is generated based on the three-dimensional frost state model and the frost quantification index set. The dynamic evolution model of frost state is then used to interact with the real-time surgical monitoring system to perform monitoring, intervention and early warning.

[0185] In Example 1, the three-dimensional frosting state model and the frosting quantification index set are spatiotemporally aligned and spliced ​​to generate a dynamic evolution model of the frosting state.

[0186] The current time frame 3D frost state model in the dynamic evolution model of frost state is registered and aligned with the pre-input surgical clinical target planning model in 3D space. The frost expansion rate in the frost quantification index set is numerically compared with the preset safe expansion rate threshold. At the same time, the minimum spatial distance between the outermost surface mesh of the registered current time frame 3D frost state model and the preset three-dimensional safety boundary of the dangerous organ is calculated.

[0187] In this implementation method, monitoring, intervention, and early warning include:

[0188] When the frost expansion rate exceeds the preset safe expansion rate threshold, or the minimum spatial spacing is less than the preset safe distance threshold and spatial collision and overlap occur, a surgical monitoring early warning signal containing a cooling and decompression command or an equipment shutdown command is generated. The surgical monitoring early warning signal is sent to the surgical real-time monitoring system to trigger the corresponding emergency highlight rendering and forced equipment flow interruption intervention.

[0189] When the frost expansion rate is less than or equal to the preset safe expansion rate threshold, and the minimum spatial spacing is greater than or equal to the preset safe distance threshold, a rendering monitoring signal is generated and sent to the surgical real-time monitoring system to display the frost range and frost speed in real time.

[0190] Example 2: In a set of verification processes for identifying the frost formation in the target area of ​​a cryoablation needle, the implementer deployed an endoscopic imaging device, an ultrasound imaging device, and an optical coherence tomography (OCT) device for a minimally invasive cryoablation scenario of deep soft tissue lesions. The cryoablation needle temperature acquisition module and pose acquisition module were simultaneously connected. The goal of the entire identification task was to provide the boundary, connectivity, volume expansion, and spatial relationship between the frost formation in the target area and the surrounding danger zone in real time as the freezing process continued, thus verifying the identification stability of the present invention in a dynamic mixed degradation environment.

[0191] In this validation process, a total of 13,240 frames of continuous freezing process images were acquired, including 13,240 endoscopic images, 6,620 ultrasound images, 4,410 sets of optical coherence tomography (OCT) slices, 26,480 temperature-time curve sampling points, and 13,240 cryoablation needle pose sampling points. The original images were uniformly registered to a spatial resolution of 1024×1024 pixels, and the time sampling interval was uniformly 0.04 s. The implementers first performed time synchronization and spatial geometric correction on the three types of visual data and the temperature-time curve to obtain the original image sequence of the target area in a unified spatiotemporal coordinate system. Before correction, the average spatial deviation between the three modal imaging was 2.84 mm, and the average timestamp deviation was 63 ms; after correction, the spatial deviation was reduced to 0.41 mm, and the time deviation was reduced to 6 ms, meeting the requirements for subsequent joint recognition.

[0192] In the initial period after freezing began, the researchers observed significant phased degradation in the original image sequence of the target area. Over 820 consecutive frames, the average gradient magnitude decreased from 18.7 to 9.4, the local contrast decreased from 0.46 to 0.21, while the proportion of bright scattering pixels increased from 1.3% to 14.8%. This indicates that the image underwent a continuous degradation process, from fogging to refraction enhancement, and then to ice crystal high reflectivity. The researchers input each frame of the original target area image into the degradation phase detection network and extracted the fogging response, refraction enhancement response, and ice crystal high reflectivity response for each frame.

[0193] In Example 2, in three consecutive representative time slices, the fogging response of the original image of the target area in one frame is 0.71, the refractive enhancement response is 0.18, and the ice crystal high reflectivity response is 0.09, and the corresponding degradation stage label is determined to be the fogging stage; in another frame, the fogging response is 0.36, the refractive enhancement response is 0.58, and the ice crystal high reflectivity response is 0.22, and the corresponding degradation stage label is determined to be the refractive enhancement stage; in yet another frame, the fogging response is 0.12, the refractive enhancement response is 0.27, and the ice crystal high reflectivity response is 0.81, and the corresponding degradation stage label is determined to be the ice crystal high reflectivity stage. After analyzing the entire sequence, it was found that the number of frames in the fogging stage accounted for 24.6%, the number of frames in the refraction enhancement stage accounted for 31.8%, and the number of frames in the ice crystal high reflectivity stage accounted for 43.6%. After time-series smoothing, the number of misjudged transitions in the stages decreased from 214 to 37, indicating that the degradation stage information has good stability.

[0194] After obtaining the smoothed degradation stage labels and degradation stage confidence scores, the implementer inputs the original image sequence of the target region and the corresponding degradation stage information into the stage-adaptive DiffUIR model. The forward diffusion processing unit of the stage-adaptive DiffUIR model automatically selects the corresponding diffusion coefficient sequence according to the stage to which the current frame belongs. Taking a foggy stage image as an example, the average diffusion coefficient of the first 10 steps is 0.0082; taking a refraction enhancement stage image as an example, the average diffusion coefficient of the first 10 steps is 0.0126; taking a high-reflectivity ice crystal stage image as an example, the average diffusion coefficient of the first 10 steps is 0.0174. Since the high-reflectivity stage requires stronger noise modeling capabilities, the diffusion coefficient is generally higher than the first two stages. The implementer performs forward diffusion with a maximum diffusion step count of 100 to obtain a complete diffusion state sequence from the original image state to the high-noise state.

[0195] The selective hourglass distribution mapping denoising unit progressively performs reverse diffusion recovery on the diffusion state sequence. Before processing a frame of a highly reflective image, the area of ​​its bright saturated region accounts for 18.3% of the target area pixel domain, the image signal-to-noise ratio is 14.2dB, and the structural similarity is 0.61; after processing, the area of ​​its bright saturated region decreases to 4.7%, the image signal-to-noise ratio increases to 24.9dB, and the structural similarity increases to 0.86. After statistical analysis of 300 consecutive images, the average signal-to-noise ratio of the high-fidelity clear target area image sequence output by the method of this invention is 25.3dB, while the traditional histogram equalization method is only 17.8dB, the traditional dehazing network is 20.1dB, and the Retinex enhancement method is 21.4dB. Especially in the stage where refraction enhancement and high reflectivity occur together, the structural recovery stability of the method of this invention is better, with a standard deviation of structural similarity between consecutive frames of only 0.019, while that of the traditional dehazing network is 0.067.

[0196] After generating a high-fidelity target region sharp image sequence, the implementer inputs the sharp image reconstruction results into the reversible residual hourglass mapping processing unit. The system performs convolution mapping on the sharp image reconstruction results of each frame to obtain an initial sharp feature map. Taking a certain frame as an example, the input number of channels is 3, and after mapping, a 64-channel initial sharp feature map is generated. Subsequently, the reversible residual hourglass coding module performs four levels of downsampling and reversible residual mapping on this feature map, and the feature space resolution is successively reduced from 1024×1024 to 512×512, 256×256, 128×128, and 64×64. The deepest bottleneck latent feature obtained has 256 channels in the channel dimension.

[0197] The latent feature decomposition module then generates signal decomposition masks and noise decomposition masks for the bottleneck latent features. Statistical analysis of 100 representative frames revealed that the average activation intensity of the signal subspace features was 0.73, while the average activation intensity of the noise subspace features was 0.28, indicating that structural information and residual perturbation information were effectively separated after the reversible residual hourglass mapping. Further channel-weighted enhancement using the channel-gated aggregation module showed that 61 out of 256 channels had gating coefficients greater than 0.005, mainly concentrated in feature channels related to boundary response, high-frequency abrupt changes, and gray-level gradient variations. The weights of the remaining noise-dominant channels were significantly suppressed; the average value of the boundary response channel for a given frame before processing was 0.43, which increased to 0.71 after processing; the average value of the high-frequency noise-dominant channel decreased from 0.31 to 0.12. The output signal enhancement feature set retained higher confidence levels of the ice crystal structure response across multiple scales.

[0198] The implementer inputs the signal enhancement feature map set into a weak feature multi-scale coding network. A fractional-order gradient sensing module performs anisotropic texture extraction on the signal enhancement feature maps at each scale. Taking a shallow high-resolution feature map as an example, in the micro-texture region that is almost indistinguishable to the naked eye in the early stage of phase transition, the average response of the traditional integer-order Sobel gradient is only 0.084, while the average response of the phase transition micro-texture after adopting fractional-order gradient sensing in this invention is increased to 0.193. Statistical analysis of 600 manually labeled early frost-covered small regions shows that the fractional-order gradient sensing module achieves an average response gain of 2.11 times for weak boundaries.

[0199] After generating phase transition microtexture features, the gradient-guided deformable coding module continues to generate a spatial offset field based on the phase transition microtexture features, deforming the original sampling mesh. Taking a 3×3 convolution kernel as an example, the sampling position of a standard convolution remains fixed, while the horizontal offset generated by this invention near a set of typical frost boundaries ranges from -1.7 to 1.5 pixels, and the vertical offset ranges from -1.4 to 1.9 pixels, enabling the receptive field to adaptively fit along the direction of the ice crystal front. After this processing, the response integrity of local ice crystal dendrite morphology is improved from 68.2% in traditional convolution to 89.4%.

[0200] The multi-scale semantic aggregation module performs cross-scale nonlocal attention interaction between deep low-resolution features and shallow high-resolution features. The implementer observed that while shallow features offer rich edge details, overall shape judgment is unstable; in deep features, the hockey puck's macroscopic outline is clear, but the edges are diffuse. After cross-scale attention fusion, the system's output multi-scale semantic feature tensor for the target area simultaneously possesses macroscopic outline constraints and microscopic phase transition frontal detail representation. Measurements on the same set of samples showed that before fusion, the average response of shallow features to the boundary region was 0.62, and the average response of deep features to the overall region was 0.75; after fusion, the combined responses of the multi-scale semantic feature tensor to the boundary and the overall region reached 0.81 and 0.84, respectively.

[0201] The implementer inputs the multi-scale semantic feature tensor of the target region into an iterative topologically consistent segmentation and decoding network. A multi-stage cascaded decoding module progressively restores the spatial resolution, ultimately outputting full-resolution decoded features consistent with the spatial scale of the original image of the target region. A probability distribution mapping module maps these features to an initial frost probability distribution map of the target region. Within a typical high-reflectivity time slice, the initial frost region generated by the traditional U-Net exhibits five isolated high-probability patches, with the largest isolated patch area being 217 pixels, and two obvious holes appearing within the main frost region. In contrast, the method of this invention, under the same input, although also exhibiting local unstable probability regions, reduces the number of isolated patches to two, the largest isolated patch area to 61 pixels, and the area of ​​the internal hole region is significantly reduced, indicating that the front-end restoration and mid-level feature modeling have reduced the risk of missegmentation in the back-end.

[0202] Based on this, the morphological prior scheduler begins operation. The implementer first performs spatial statistical analysis on the initial frosting probability distribution map of the target area, calculates the probability centroid position, and then combines this with the cryoablation biological heat transfer physical model to obtain the temperature field distribution of the current frame. The axial and radial growth scales are then obtained through ellipsoidal structure fitting. Taking a specific frame as an example, the probability centroid is found to be located near pixel coordinates (512.6, 498.3), corresponding to an axial growth scale of 146 pixels and a radial growth scale of 109 pixels. After generating the physical morphological prior field based on these parameters, the system obtains a connectivity constraint coefficient of 0.83, a curvature smoothing constraint coefficient of 0.67, and a volume growth constraint coefficient of 0.58.

[0203] Connectivity constraint optimization was performed on the initial frosting probability distribution map of the target area. The system first thresholded to obtain the initial candidate frosting region mask, and then decomposed it into connected components. In one frame, a total of 7 connected regions were decomposed, of which the main connected region had an area of ​​28411 pixels, and the areas of the other 6 connected regions were all less than 420 pixels. After connectivity scoring, the score of the main connected region was 0.792, and the score of the largest out-of-film isolated region was only 0.118. After applying connectivity constraints, the average frosting probability of the main connected region increased from 0.76 to 0.81, and the average probability of the remaining isolated regions decreased from 0.43 to 0.09, and isolated artifacts were basically eliminated.

[0204] The system performs curvature smoothing constraint optimization on the connectivity constraint probability map. Statistical analysis of the discrete curvature surrogate quantity in the boundary region of a certain frame shows that the traditional method has a mean second-order change intensity of 1.84 at the boundary jagged edges. After introducing the curvature smoothing response, this invention brings the predicted frosting probability in this region closer to the local mean, reduces the boundary jagged edge length from 39 pixels to 11 pixels, and reduces the number of sharp abrupt changes in hole edges from 4 to 0. Statistical results over 400 consecutive frames show that after curvature smoothing constraints, the average perimeter fluctuation rate of the boundary decreases from 13.6% to 4.1%.

[0205] The system performs volume growth constraint optimization. For a given pair of consecutive frames, the volume of the frosting region in the previous frame is 29,874 pixels, while the estimated volume of the frosting region in the current frame without volume constraints is 33,690 pixels. The upper limit of the allowable volume growth calculated based on the physical morphology prior field and the volume growth constraint coefficient is 31,942 pixels. The system then multiplies the curvature smoothing probability map pixel by pixel with the physical morphology prior field, sorts them by response value, and retains only the top 31,942 pixels with the highest response values ​​as the final frosting region for the current frame. After this step, the sporadic outward expansion that originally appeared in the peripheral areas where the low temperature did not fully reach is completely suppressed. The main frosting region maintains continuous growth but does not exhibit any non-physical abrupt increases. The generated frosting region mask with consistent morphology is visually more consistent with the actual morphology of freeze-thaw cycles.

[0206] During the temporal stabilization phase, the implementer performs an optical flow inverse mapping comparison between the current frame's consistent frosting region mask and the previous frame's temporally stable frosting region mask. Within a certain continuous time period, due to slight local tissue displacement, the average displacement of the dense optical flow between the current and previous frames is 2.7 pixels. The system uses this dense optical flow field to perform a spatial distortion inverse mapping on the current frame mask to obtain a forward prediction mask and calculates the inter-frame topological consistency score. If the score is lower than a preset consistency tolerance threshold, fine-tuning is performed on non-overlapping difference regions. Actual recordings show that fine-tuning loops were triggered in 611 out of 13240 frames, accounting for 4.62%. These triggered frames mainly occurred during rapid boundary expansion and slight changes in imaging angle. After fine-tuning, the inter-frame Dice consistency improved from 0.887 to 0.946, indicating a significant enhancement in temporal stability.

[0207] After generating a time-stable frosting region mask sequence, the system combines the real-time 3D pose information of the cryoablation needle with spatial calibration parameters to perform inverse projection on the 2D mask. It then performs 3D rotational sweep reconstruction using ellipsoidal structure parameters from the physical morphology prior field to obtain a 3D frosting state model. Taking a single frame as an example, the reconstructed frosting volume is 4.28 cm³, the maximum frosting thickness is 12.6 mm, the average frosting thickness is 7.9 mm, and the frosting expansion rate between adjacent frames is 0.37 mm / s. After spatial registration of the 3D frosting state model with the surgical planning model, the minimum spatial distance between the outermost surface of the current frosting and a certain danger boundary is found to be 3.4 mm. When the frosting expansion rate rises to 0.82 mm / s and the minimum spatial distance decreases to 1.1 mm at a later time, the system generates a surgical monitoring warning signal. After the expansion rate falls back to 0.41 mm / s and the minimum distance recovers to 2.8 mm, the system switches back to the rendering monitoring signal, displaying only the frosting range and frosting speed in real time.

[0208] To verify the effectiveness of the method of this invention, the implementers constructed a unified training sample set and a test sample set. The training sample set contains 186 sets of cryoablation sequences, totaling 96,420 frames, including 22,840 frames of samples in the fogging stage, 30,170 frames of samples in the refraction enhancement stage, and 43,410 frames of samples in the ice crystal high reflectivity stage; 27,860 frames of samples with manually annotated frost boundaries frame by frame; and 412 sets of multi-view samples for 3D reconstruction verification. The test sample set contains 37 sets of cryoablation sequences, totaling 21,480 frames. All methods were compared on the same training sample set, the same hardware conditions, and the same test sample set. The comparison methods included histogram equalization + U-Net, a single defogging network + U-Net, Retinex enhancement + DeepLabV3 +, and the method of this invention. The test results are shown in Table 1 below:

[0209] Table 1. Data Comparison between the Invention and Traditional Methods

[0210]

[0211] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for recognizing the frosting state of the target area of ​​a cryoablation needle based on image processing, characterized in that, include: During cryoablation surgery, multimodal raw visual data of the target area were collected, and time synchronization registration and spatial geometric correction were performed to construct a raw image sequence of the target area under a unified spatiotemporal coordinate system. Degradation stage labels are generated based on the original image sequence of the target area, and the corresponding degradation stage information is output. The original image sequence of the target area and the degradation stage information are input into the stage adaptive DiffUIR model to obtain a high-fidelity clear image sequence of the target area. Reversible residual hourglass mapping is performed on high-fidelity target area clear image sequences to decompose the latent feature space into signal subspace and noise subspace. The signal subspace features are then aggregated through a channel gating mechanism to obtain a signal enhancement feature map set. Input the signal enhancement feature map set into the weak feature multi-scale coding network to generate a multi-scale semantic feature tensor of the target region. Input the multi-scale semantic feature tensor of the target area into the iterative topologically consistent segmentation and decoding network, and output the initial frost probability distribution map of the target area; A morphological prior scheduler is constructed, and the connectivity constraints, curvature smoothing constraints, and volume growth constraints output by the morphological prior scheduler are applied to the initial frosting probability distribution map of the target area. Morphological constraint segmentation optimization is performed to obtain a morphologically consistent frosting region mask. Optical flow reverse mapping comparison and anomaly fine-tuning are performed on the frost region mask with the frost region mask with the same morphology in the previous frame to obtain a temporally stable frost region mask sequence and construct a dynamic evolution model of frost state. The dynamic evolution model of frost state is then used to interact with the real-time surgical monitoring system to display early warning information in real time.

2. The method for recognizing the frosting state of the target area of ​​a cryoablation needle based on image processing according to claim 1, characterized in that, The multimodal raw visual data of the target area includes synchronized endoscopic images, ultrasound images, optical coherence tomography data, and corresponding temperature-time curves.

3. The method for recognizing the frosting state of the target area of ​​a cryoablation needle based on image processing according to claim 1, characterized in that, The generation of degradation stage labels based on the original image sequence of the target region includes: For each frame of the original image of the target area, a degradation characterization quantity is extracted, and a degradation feature vector consisting of fogging response quantity, refraction enhancement response quantity, and ice crystal high reflectivity response quantity is constructed. The input sequence of the degradation stage detection network is constructed using the degradation feature vectors of consecutive frames. The input sequence is then fed into the degradation stage detection network, which outputs the degradation stage prediction vector corresponding to the original image of the target area, which is constrained by the monotonicity of the unidirectional thermodynamic phase transition. Based on the degradation stage prediction vector, degradation stage labels corresponding to the original image of the target area are generated. Temporal smoothing is then performed on the degradation stage labels to obtain smoothed degradation stage labels. Based on the smoothed degradation stage label and degradation stage prediction vector, the corresponding degradation stage information is output.

4. The method for recognizing the frosting state of the target area of ​​a cryoablation needle based on image processing according to claim 1, characterized in that, The step of inputting the original image sequence of the target region and degradation stage information into the stage-adaptive DiffUIR model includes: Based on the smoothed degradation stage label corresponding to the original image of the target region in frame t, a diffusion coefficient is constructed for the original image of the target region in frame t. A forward diffusion process is performed on the original image of the target region in frame t based on the diffusion coefficient to generate a diffusion state sequence; Based on the degradation stage information, noise estimation is performed on each diffusion state in the diffusion state sequence to generate the corresponding noise estimation results and cumulative retention coefficients. Based on the degradation stage information of frame t, a noise reduction intensity adjustment coefficient is constructed in the reverse diffusion process; Based on the noise estimation results, cumulative preservation coefficients, and denoising intensity adjustment coefficients, a reverse diffusion process is performed to obtain the clear image reconstruction result corresponding to the original image of the target region in frame t. The reconstruction results of each clear image are output frame by frame to construct a high-fidelity clear image sequence of the target area.

5. The method for recognizing the frosting state of the target area of ​​a cryoablation needle based on image processing according to claim 1, characterized in that, The reversible residual hourglass mapping process for high-fidelity target region clear image sequences includes: The sharp image reconstruction result in the high-fidelity target region sharp image sequence is input into the reversible residual hourglass mapping processing unit to generate the corresponding initial sharp feature map; Perform reversible residual hourglass mapping on the initial clear feature map to obtain the corresponding bottleneck potential features; Based on the bottleneck potential features, the potential feature space is decomposed to obtain signal subspace features and noise subspace features; Channel gating coefficients are constructed based on signal subspace features and noise subspace features. The channel gating coefficients are then used to enhance the signal subspace features while suppressing the noise subspace features, thus obtaining gated signal features. Hourglass decoding and reconstruction are performed on the gated signal features of frame t to obtain the signal enhancement feature set of frame t. The signal enhancement feature sets of each frame are output frame by frame to construct the total signal enhancement feature set.

6. The method for recognizing the frosting state of the target area of ​​a cryoablation needle based on image processing according to claim 1, characterized in that, The step of inputting the signal enhancement feature map set into the weak feature multi-scale coding network includes: The signal enhancement feature map set is input into the weak feature multi-scale coding network, and fractional gradient response is extracted for the signal enhancement feature map at each scale level to obtain the phase transition microtexture features at the corresponding scale. A spatial offset field is generated based on phase transition microtexture features, and deformable convolutional coding is performed on the signal enhancement feature map using the spatial offset field to obtain morphology-aware features that adaptively fit the ice crystal growth morphology. Cross-scale nonlocal attention interaction is performed on morphological perception features at various scales, and global topological context and local micro-texture are fused to generate multi-scale semantic feature tensors for the target region.

7. The method for recognizing the frosting state of the target area of ​​a cryoablation needle based on image processing according to claim 1, characterized in that, The step of inputting the multi-scale semantic feature tensor of the target region into the iterative topologically consistent segmentation and decoding network includes: The multi-scale semantic feature tensor of the target area is input into an iterative topologically consistent segmentation and decoding network to perform cross-scale feature concatenation decoding and feature resolution restoration, resulting in full-resolution decoded features consistent with the spatial scale of the original image of the target area. Extract the structural topology attributes of the full-resolution decoded features, construct a boundary-aware gating mechanism, perform topology feature correction on the full-resolution decoded features, and obtain topology-consistent boundary enhancement features; The topological consistency boundary enhancement feature is mapped to a single-channel probability space to generate an initial frost probability distribution map of the target area.

8. The method for recognizing the frosting state of the target area of ​​a cryoablation needle based on image processing according to claim 1, characterized in that, The process of applying the connectivity constraints, curvature smoothing constraints, and volume growth constraints output by the morphological prior scheduler to the initial frosting probability distribution map of the target region includes: A morphological prior scheduler is constructed, and a corresponding physical morphological prior field, as well as connectivity constraint coefficient, curvature smoothing constraint coefficient, and volume growth constraint coefficient, are generated based on the initial frost probability distribution map of the target area. Based on the connectivity constraint coefficient and the physical morphology prior field, the initial frost probability distribution map of the target area is optimized by connectivity constraint to obtain the connectivity constraint probability map; Based on the curvature smoothing constraint coefficient, curvature smoothing constraint optimization is performed on the connectivity constraint probability graph to obtain the curvature smoothing probability graph. Based on the volume growth constraint coefficient, volume growth constraint optimization is performed on the curvature smoothing probability map to obtain a frosting region mask with consistent shape. Frame by frame, output the frost region mask with consistent shape in each frame to construct a sequence of frost region masks with consistent shape.

9. The method for recognizing the frosting state of the target area of ​​a cryoablation needle based on image processing according to claim 1, characterized in that, The process involves performing optical flow reverse mapping comparison and anomaly fine-tuning on the morphologically consistent frost region mask and the morphologically consistent frost region mask of the previous frame to obtain a temporally stable frost region mask sequence and constructing a dynamic evolution model of the frost state, including: Optical flow reverse mapping comparison and anomaly fine-tuning are performed on the frost region mask with the frost region mask with the same shape as the previous frame to obtain a temporally stable frost region mask sequence. Based on a time-stable frosting region mask sequence, a three-dimensional frosting state model is constructed, and a set of frosting quantification indicators for the target area is calculated. A dynamic evolution model of frost state is generated based on the three-dimensional frost state model and the frost quantification index set. The dynamic evolution model of frost state is then used to interact with the real-time surgical monitoring system to perform monitoring, intervention and early warning.

10. The method for recognizing the frosting state of the target area of ​​a cryoablation needle based on image processing according to claim 9, characterized in that, The monitoring, intervention, and early warning systems include: When the frost expansion rate exceeds the preset safe expansion rate threshold, or the minimum spatial spacing is less than the preset safe distance threshold and spatial collision and overlap occur, a surgical monitoring early warning signal containing a cooling and decompression command or an equipment shutdown command is generated. The surgical monitoring early warning signal is sent to the surgical real-time monitoring system to trigger the corresponding emergency highlight rendering and forced equipment flow interruption intervention. When the frost expansion rate is less than or equal to the preset safe expansion rate threshold, and the minimum spatial spacing is greater than or equal to the preset safe distance threshold, a rendering monitoring signal is generated and sent to the surgical real-time monitoring system to display the frost range and frost speed in real time.

Images