A wound three-dimensional model reconstruction system based on reinforcement learning

Abstract

The application relates to the technical field of three-dimensional reconstruction, and discloses a wound three-dimensional model reconstruction system based on reinforcement learning, which comprises a wound data access module, a tissue context coding module, an exploration decision module, a three-dimensional Gaussian reconstruction module and a wound three-dimensional model output module; wherein the tissue context coding module models the space-time correlation between wound edges, wound surface depressions, exudation reflection areas and local occlusion areas through a space-time Transformer structure; the exploration decision module dynamically adjusts a supplementary viewing angle and a local reconstruction action in combination with a reinforcement learning strategy network; and the three-dimensional Gaussian reconstruction module performs Gaussian primitive adaptive densification processing based on an intrinsic signal-to-noise ratio constraint. The application improves the geometric recovery capability in a complex wound scene, and improves the reconstruction stability, geometric continuity and reconstruction precision of a wound three-dimensional model.

Description

Technical Field

[0001] This invention relates to the field of three-dimensional reconstruction technology, and in particular to a wound three-dimensional model reconstruction system based on reinforcement learning. Background Technology

[0002] With the continuous increase in the number of complex wound cases such as chronic wounds, postoperative wounds, diabetic foot ulcers, and pressure ulcers, digital wound assessment technology based on 3D reconstruction has gradually become an important research direction in the field of intelligent medical analysis. Existing intelligent 3D wound reconstruction technologies typically use depth cameras, multi-view image acquisition, point cloud reconstruction, or neural radiation fields to generate 3D wound models, and then combine mesh reconstruction, voxel optimization, or Gaussian representation methods to restore the wound's geometric structure. However, wound scenes themselves have significant complex tissue characteristics. Wound areas commonly exhibit exudate reflection, local occlusion, blurred boundaries, depth acquisition noise, and discontinuous tissue depressions, leading to problems such as missing boundaries, distorted reconstruction of depressed areas, local holes, and the diffusion of redundant noise structures in existing intelligent reconstruction methods. Furthermore, most existing technologies treat wound point clouds as static spatial data, lacking the ability to model the correlation of tissue changes between different acquisition times, making it difficult to utilize cross-time and cross-view contextual information to constrain and correct locally missing and unstable areas. In addition, existing intelligent 3D reconstruction systems typically employ fixed acquisition strategies and static densification mechanisms, which cannot dynamically adjust the acquisition perspective, lighting conditions, and local reconstruction parameters based on uncertain areas in the current reconstruction results. This can easily lead to problems such as repeated acquisition, ineffective densification, and accumulation of noise structures, thereby affecting the geometric integrity, boundary accuracy, and reconstruction stability of the wound 3D model. Summary of the Invention

[0003] This invention aims to address the problems in existing intelligent wound 3D reconstruction technologies, such as reconstruction distortion in reflective areas, geometrical loss in occluded areas, unstable boundary region recovery, and repetitive exploration and ineffective densification due to fixed re-sampling strategies. It provides a wound 3D model reconstruction system based on reinforcement learning. This invention constructs a wound tissue spatiotemporal context modeling mechanism to jointly model the changes in wound point clouds at different acquisition times and from different perspectives, enabling stable tissue structures in adjacent times and perspectives to provide geometric constraints on locally missing and unstable regions. Simultaneously, this invention incorporates a reinforcement learning exploration decision-making mechanism to dynamically adjust the re-sampling perspective, illumination status, and local reconstruction actions based on the current uncertain reconstruction areas, thereby improving the effective observation capability of complex wound regions. Furthermore, this invention performs Gaussian unit adaptive densification processing based on intrinsic signal-to-noise ratio constraints, reducing the impact of depth noise, local reflections, and occlusion disturbances on the 3D reconstruction results, thereby improving the geometric continuity and 3D model reconstruction accuracy of wound boundary regions, wound depression regions, and tissue defect regions.

[0004] This invention provides a wound 3D model reconstruction system based on reinforcement learning. The system includes a wound data access module, a tissue context encoding module, a reinforcement learning state construction module, an exploration decision module, a 3D Gaussian reconstruction module, and a wound 3D model output module. The wound data access module is used to acquire multi-source raw wound observation data corresponding to the target wound, and perform time alignment, camera calibration, wound region segmentation, and coordinate unification processing on the multi-source raw wound observation data to generate a wound spatiotemporal point cloud sequence. Based on the three-dimensional spatial coordinates, color features, depth confidence, and wound region segmentation results in the wound spatiotemporal point cloud sequence, the module determines the position parameters, scale parameters, covariance parameters, color parameters, transparency parameters, and uncertainty parameters of multiple wound three-dimensional Gaussian primitives, generating an initial set of wound three-dimensional Gaussian primitives. The multi-source raw wound observation data may include color images, depth images, wound region masks, acquisition pose data, and illumination status data. Each point in the wound spatiotemporal point cloud sequence includes three-dimensional spatial coordinates, normalized time coordinates, color features, wound mask identifiers, and acquisition viewpoint identifiers. The tissue context encoding module performs spatiotemporal point cloud patch encoding, cross-time-step attention aggregation, and contextual latent representation generation on the spatiotemporal point cloud sequence of the wound to obtain the wound tissue contextual latent representation. This module treats the wound point cloud not only as a static geometric point set, but also regards the changes in the point cloud at multiple acquisition times as a spatiotemporal observation process that can characterize the undulations of the wound tissue edges, wound depressions, local occlusions, and exudation reflection states. The resulting wound tissue contextual latent representation is used to provide tissue context conditions for the reinforcement learning reconstruction state construction module. The reinforcement learning state construction module consists of the current wound 3D Gaussian graph set, the historical reconstruction action record, and the current reconstruction uncertainty map, obtained by iteratively updating the initial wound 3D Gaussian graph set through historical reconstruction actions. Based on the wound tissue contextual latent representation, the current wound 3D Gaussian graph set, the historical reconstruction action record, and the current reconstruction uncertainty map, a wound reconstruction reinforcement learning state object is generated. The current wound 3D Gaussian graph set is used to express the 3D representation of the wound at the current reconstruction stage; the current reconstruction uncertainty map is used to characterize the reconstruction uncertainty of the wound boundary, recessed area, reflective area, and occluded area; and the historical reconstruction action record is used to characterize the previously performed acquisition view adjustment, illumination adjustment, re-acquisition, and densification operations. The exploration decision module constructs a reinforcement learning policy network and a set of reconstruction dynamic predictors. Based on the wound reconstruction reinforcement learning state object, a set of candidate reconstruction actions is constructed. The wound reconstruction reinforcement learning state object and candidate reconstruction actions are input into the set of reconstruction dynamic predictors to calculate the intrinsic reward value of the conditional information gain corresponding to the candidate reconstruction actions. A comprehensive reconstruction reward value is generated based on the intrinsic reward value of the conditional information gain and the extrinsic reward value of the wound geometric integrity. The reinforcement learning policy network is updated using the comprehensive reconstruction reward value, and the target reconstruction action is determined from the set of candidate reconstruction actions through the updated reinforcement learning policy network. The 3D Gaussian reconstruction module performs rendering error backpropagation, context adaptive moment estimation, signal-to-noise ratio gating, and structural densification on the current 3D Gaussian primitive set of the wound based on the target reconstruction action, generating a densified 3D Gaussian primitive set of the wound. This module changes the densification triggering basis of the 3D Gaussian primitive of the wound from a simple gradient magnitude to a statistical signal verification mechanism jointly determined by the first moment, second moment, quantile sorting, and intrinsic signal-to-noise ratio, thereby reducing redundant primitives caused by reflection, occlusion, and depth noise while preserving the real wound boundary and depression details. The wound 3D model output module is used to generate wound 3D model reconstruction results based on the dense wound 3D Gaussian tuple, and to generate and save wound reconstruction feedback data based on the wound 3D model reconstruction results and the subsequently collected wound feedback observation data. The wound 3D model reconstruction results include wound 3D surface model, wound boundary curve, wound surface area, wound maximum depth, wound volume, edge slope distribution, and wound color texture map.

[0005] By adopting the above solution, the beneficial effects achieved by the present invention are as follows: This invention constructs a spatiotemporal context modeling mechanism for wound tissue, enabling joint correlation analysis of wound tissue changes at different acquisition times and from different acquisition perspectives. This improves the geometric recovery capability of reflective areas, occluded areas, and blurred boundary areas in complex wound scenarios. It solves the problems of boundary breakage, depression distortion, and local holes caused by local tissue loss, exudate reflection, and unstable depth acquisition in existing intelligent wound 3D reconstruction technologies. It enhances the spatial continuity and structural integrity of wound boundary areas, wound depression areas, and tissue defect areas, thereby enabling the wound 3D model generated by this invention to more realistically reflect the changes in wound geometry.

[0006] This invention constructs a reinforcement learning-based exploration and decision-making mechanism to achieve dynamic identification and adaptive re-sampling control of uncertain areas in the current wound reconstruction. This improves the effective observation capability and reconstruction stability of complex wound areas and solves the problems of repeated exploration, invalid acquisition, and insufficient local information coverage that are easily generated by fixed re-sampling strategies in existing intelligent 3D reconstruction systems. It enhances the system's dynamic perception and autonomous exploration capabilities of the lower edge of the wound, areas with unstable depth, and locally occluded areas. As a result, this invention can automatically adjust the acquisition perspective, lighting conditions, and local reconstruction actions according to the current reconstruction status, thereby improving the overall reconstruction efficiency and accuracy of the 3D wound model.

[0007] This invention constructs an adaptive densification mechanism for Gaussian primitives based on intrinsic signal-to-noise ratio constraints. This mechanism enables dynamic statistical verification and selective densification control of the reliability of 3D Gaussian primitives in wounds, improving the geometric detail representation and local texture stability of complex wound regions. It also solves the problem of redundant Gaussian primitive diffusion caused by depth noise, local reflection, and occlusion disturbances in existing Gaussian reconstruction methods. Furthermore, it enhances the geometric continuity and 3D representation accuracy of wound edge contours, wound depression structures, and tissue texture regions, thereby improving the 3D model reconstruction quality and subsequent wound quantification assessment reliability in complex wound scenarios. Attached Figure Description

[0008] Figure 1 This is a schematic diagram of the module structure of a three-dimensional wound model reconstruction system based on reinforcement learning according to the present invention. Figure 2 This is a schematic diagram of the attention association modeling proposed in step A3 of Embodiment 3 of the present invention. The diagram illustrates the process by which the spatiotemporal Transformer encoder performs global spatiotemporal attention association between edge regions, concave regions, exudate / reflective regions, and occluded regions during wound tissue context modeling. Among them, edge region tokens (red) are used to represent wound contour change information, concave region tokens (green) are used to represent wound depth change information, exudate / reflective region tokens (blue) are used to represent local highlight and depth unstable region information, occluded region tokens (purple) are used to represent local missing region information, and other region tokens (gray) are used to represent normal tissue regions, transition regions, and background auxiliary structure regions around the wound. Detailed Implementation

[0009] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0010] Example 1, according to Figure 1 The present invention provides a wound 3D model reconstruction system based on reinforcement learning. The system includes a wound data access module, a tissue context encoding module, a reinforcement learning state construction module, an exploration decision module, a 3D Gaussian reconstruction module, and a wound 3D model output module. In this embodiment, a chronic wound formed by local dehiscence of the abdominal incision after surgery was used as the target wound. The wound was located in the lower right quadrant of the abdomen. Before collection, routine cleaning was performed by nursing staff. The visible area of ​​the target wound had a long axis of approximately 48.6 mm and a short axis of approximately 21.4 mm. There was a slight depression in the center of the wound, a local bulge at the wound edge, and a small amount of reflective exudate at the lower edge of the wound.

[0011] The wound data access module is used to acquire multi-source raw wound observation data corresponding to the target wound, and perform time alignment, camera calibration, wound region segmentation, and coordinate unification processing on the multi-source raw wound observation data to generate a wound spatiotemporal point cloud sequence. Based on the three-dimensional spatial coordinates, color features, depth confidence, and wound region segmentation results in the wound spatiotemporal point cloud sequence, the module determines the position parameters, scale parameters, covariance parameters, color parameters, transparency parameters, and uncertainty parameters of multiple wound three-dimensional Gaussian primitives, generating an initial set of wound three-dimensional Gaussian primitives. The multi-source raw wound observation data may include color images, depth images, wound region masks, acquisition pose data, and illumination status data. Each point in the wound spatiotemporal point cloud sequence includes three-dimensional spatial coordinates, normalized time coordinates, color features, wound mask identifiers, and acquisition viewpoint identifiers. The tissue context encoding module performs spatiotemporal point cloud patch encoding, cross-time-step attention aggregation, and contextual latent representation generation on the spatiotemporal point cloud sequence of the wound to obtain the wound tissue contextual latent representation. This module treats the wound point cloud not only as a static geometric point set, but also regards the changes in the point cloud at multiple acquisition times as a spatiotemporal observation process that can characterize the undulations of the wound tissue edges, wound depressions, local occlusions, and exudation reflection states. The resulting wound tissue contextual latent representation is used to provide tissue context conditions for the reinforcement learning reconstruction state construction module. In this embodiment, after receiving the wound spatiotemporal point cloud sequence, the tissue context encoding module combines the three-dimensional spatial coordinates and normalized temporal coordinates of each point into four-dimensional spatiotemporal coordinates. Based on the four-dimensional spatiotemporal coordinates, farthest point sampling is performed to select 512 wound spatiotemporal patch centers. For each patch center, neighboring points are searched in the four-dimensional spatiotemporal coordinate space, with a neighborhood radius of 4.0 mm and a temporal neighborhood span of 3 adjacent acquisition segments. A maximum of 64 points are retained for a single wound spatiotemporal patch. Subsequently, Fourier position encoding and local pooling are performed on the points in each wound spatiotemporal patch to generate a wound spatiotemporal patch token with a dimension of 128. Fourier position encoding is used to map local spatial coordinates, temporal coordinates, and small depth changes to a high-dimensional feature space, while local pooling is used to extract edge undulations, depth depressions, color changes, and local occlusion states within the patch.

[0012] In this embodiment, the spatiotemporal Transformer encoder is configured with 4 layers, each layer including 8 attention heads, and the hidden dimension is 256. After multiple wound spatiotemporal patch tokens are input into the spatiotemporal Transformer encoder, the self-attention mechanism models the spatiotemporal correlation between wound edges, wound depressions, reflective points in exudate areas, and occluded missing areas. For example, when a patch is located in the reflective area at the lower edge of the wound, the patch has depth instability in the single-frame depth map, but its boundary patches at adjacent times and adjacent viewpoints have continuous spatial curvature changes. Therefore, the spatiotemporal Transformer encoder can use adjacent visible patches to constrain the true geometric trend of the reflective area. After encoding, the tissue context encoding module performs attention pooling on the wound spatiotemporal token features to obtain a sequence latent representation with a dimension of 256. For the 5 wound point cloud sequences formed by the 5 acquisition viewpoints, learnable soft weight aggregation processing is further performed. After aggregation, a wound tissue context latent representation with a dimension of 256 is obtained.

[0013] The reinforcement learning state construction module consists of the current wound 3D Gaussian graph set, the historical reconstruction action record, and the current reconstruction uncertainty map, obtained by iteratively updating the initial wound 3D Gaussian graph set through historical reconstruction actions. Based on the wound tissue contextual latent representation, the current wound 3D Gaussian graph set, the historical reconstruction action record, and the current reconstruction uncertainty map, a wound reconstruction reinforcement learning state object is generated. The current wound 3D Gaussian graph set is used to express the 3D representation of the wound at the current reconstruction stage; the current reconstruction uncertainty map is used to characterize the reconstruction uncertainty of the wound boundary, recessed area, reflective area, and occluded area; and the historical reconstruction action record is used to characterize the previously performed acquisition view adjustment, illumination adjustment, re-acquisition, and densification operations. The exploration decision module constructs a reinforcement learning policy network and a set of reconstruction dynamic predictors. Based on the wound reconstruction reinforcement learning state object, a set of candidate reconstruction actions is constructed. The wound reconstruction reinforcement learning state object and candidate reconstruction actions are input into the set of reconstruction dynamic predictors to calculate the intrinsic reward value of the conditional information gain corresponding to the candidate reconstruction actions. A comprehensive reconstruction reward value is generated based on the intrinsic reward value of the conditional information gain and the extrinsic reward value of the wound geometric integrity. The reinforcement learning policy network is updated using the comprehensive reconstruction reward value, and the target reconstruction action is determined from the set of candidate reconstruction actions through the updated reinforcement learning policy network. In this embodiment, the exploration decision module constructs a reinforcement learning policy network and a set of reconstruction dynamic predictors. The set of reconstruction dynamic predictors includes five one-step predictors, each with the same network structure but different initialization parameters, used to predict the next reconstruction state after executing a candidate reconstruction action. The calculated conditional information gain intrinsic reward values ​​for some candidate reconstruction actions are as follows: ; The calculation results show that the action of lowering the camera height by 20mm and performing supplementary sampling corresponds to the highest comprehensive reconstruction reward value, and its conditional information gain intrinsic reward value is also relatively high. This indicates that this action can provide new observational information for the lower edge and depression areas of the wound, which currently have high uncertainty, while having low redundancy with historical actions. Therefore, the exploration decision module identifies "lowering the camera height by 20mm and performing supplementary sampling on the lower edge of the wound" as the target reconstruction action. After executing the target reconstruction action, the system added 42 valid frames and approximately 138,000 valid point clouds. After registration and downsampling, 3,260 new points were added to the wound spatiotemporal point cloud sequence. After supplementary sampling, the mean uncertainty of the lower edge region of the wound decreased from 0.72 to 0.46, and the mean uncertainty of the central depression area of ​​the wound decreased from 0.64 to 0.49.

[0014] The 3D Gaussian reconstruction module performs rendering error backpropagation, context adaptive moment estimation, signal-to-noise ratio gating, and structural densification on the current 3D Gaussian primitive set of the wound based on the target reconstruction action, generating a densified 3D Gaussian primitive set of the wound. This module changes the densification triggering basis of the 3D Gaussian primitive of the wound from a simple gradient magnitude to a statistical signal verification mechanism jointly determined by the first moment, second moment, quantile sorting, and intrinsic signal-to-noise ratio, thereby reducing redundant primitives caused by reflection, occlusion, and depth noise while preserving the real wound boundary and depression details. The wound 3D model output module is used to generate a wound 3D model reconstruction result based on the dense wound 3D Gaussian unit set, and to generate wound reconstruction feedback data based on the wound 3D model reconstruction result and subsequently collected wound feedback observation data. The wound reconstruction feedback data is used to update the tissue context encoding module, the exploration decision module, and the 3D Gaussian reconstruction module. The wound 3D model reconstruction result includes a wound 3D surface model, wound boundary curve, wound surface area, maximum wound depth, wound volume, edge slope distribution, and wound color texture map.

[0015] Example 2, based on Example 1, describes the process by which the wound data access module generates a spatiotemporal point cloud sequence of wounds, specifically including the following steps: Step S1: Acquire color images, depth images, wound area masks, acquisition posture data, and illumination status data of the same target wound at multiple acquisition times and from multiple acquisition perspectives to generate multi-source raw wound observation data; In this embodiment, multiple acquisition perspectives are acquired by a depth camera on a fixed support according to a preset trajectory; Acquisition posture data includes the distance, pitch angle, yaw angle, and roll angle of the camera relative to the target wound. Step S2: Based on the acquired pose data and camera intrinsic and extrinsic parameters, perform spatial registration processing on the color image and depth image, and map the wound area mask onto the registered depth image to obtain the wound area depth observation data; this step ensures that subsequent point cloud generation is performed only on the target wound and its adjacent skin boundaries, reducing the impact of background sheets, dressings and hands on the reconstruction of the 3D model of the wound. Step S3: Perform outlier removal, local hole completion, scale normalization, and coordinate transformation on the wound area depth observation data to obtain the wound local point cloud frames corresponding to each acquisition time. Outlier removal is performed based on neighborhood distance statistics, local hole completion is performed based on neighborhood plane interpolation or point cloud resampling, scale normalization is used to convert the wound point clouds at different acquisition distances to a uniform scale, and coordinate transformation is used to convert point cloud frames from different viewpoints to the same wound local coordinate system. Step S4: Perform temporal organization processing on multiple local point cloud frames of the wound according to the acquisition time sequence, and add three-dimensional spatial coordinates, normalized time coordinates, color features, wound mask identifiers and acquisition viewpoint identifiers to each point to generate a spatiotemporal point cloud sequence of the wound.

[0016] Example 3, according to Figure 2 This embodiment is based on Embodiment 2. In this embodiment, the process of obtaining the potential representation of wound tissue context by the tissue context encoding module specifically includes the following steps: Step A1: Combine the three-dimensional spatial coordinates and normalized time coordinates in the wound spatiotemporal point cloud sequence into four-dimensional spatiotemporal coordinates, and perform farthest point sampling processing based on the four-dimensional spatiotemporal coordinates to obtain multiple wound spatiotemporal patch centers; by sampling in the four-dimensional spatiotemporal coordinates, the system can simultaneously focus on the spatial distribution and cross-time changes of the wound surface. Step A2: For each wound spatiotemporal patch center, perform a neighbor search process in the four-dimensional spatiotemporal coordinates to obtain the wound spatiotemporal patch, and perform Fourier position encoding and local pooling processing on the points in the wound spatiotemporal patch to generate a wound spatiotemporal patch token. Step A3: Construct a spatiotemporal Transformer encoder based on the spatiotemporal Transformer network structure; input multiple wound spatiotemporal patch tokens into the spatiotemporal Transformer encoder, and use the self-attention mechanism to model the spatiotemporal correlation between wound edges, wound depressions, reflective points in the exudate area, and occluded missing areas to generate wound spatiotemporal token features; the spatiotemporal Transformer encoder includes multiple self-attention layers, a feedforward network, and a normalization layer. The wound spatiotemporal patch tokens establish a global correlation through attention weights, thereby constraining the shape estimation of the missing area by the visible edges near the occluded area and the point cloud at adjacent time points; Step A4: Perform attention pooling on the spatiotemporal token features of the wound to obtain the sequence latent representation corresponding to a single wound point cloud sequence; when there are multiple wound point cloud sequences for the same target wound, perform learnable soft weight aggregation on the multiple sequence latent representations to generate an initial context latent representation; Step A5: During the training phase, auxiliary supervision processing is performed on the initial contextual latent representation based on the wound tissue attribute labels, and symbolic distance field reconstruction supervision processing is performed on the wound spatiotemporal patch token based on the wound symbolic distance field labels, so as to jointly optimize the wound spatiotemporal patch encoding, spatiotemporal Transformer feature extraction and contextual latent representation generation process, so as to obtain the wound tissue contextual latent representation, which simultaneously retains wound tissue attribute information and fine-grained geometric boundary information.

[0017] Example 4, based on Example 3, explores the process by which the decision-making module generates the conditional information gain intrinsic reward value and the target reconstruction action, specifically including the following steps: Step B1: Construct a set of candidate reconstruction actions based on the wound reconstruction reinforcement learning state object. The candidate reconstruction actions in the set include the next acquisition view adjustment action, the illumination state adjustment action, the point cloud re-acquisition action, the Gaussian unit densification action, and the local reconstruction parameter update action. Step B2: Input the wound reconstruction reinforcement learning state object and candidate reconstruction actions into a set of reconstruction dynamic predictors consisting of multiple one-step predictors, obtain the prediction result of each one-step predictor for the next reconstruction state, and calculate the prediction bias vector based on multiple prediction results. Step B3: Construct an ensemble bifurcation kernel matrix based on the prediction bias vectors corresponding to different candidate reconstruction actions, and introduce observation noise ridge terms obtained from historical reconstruction residual estimation into the ensemble bifurcation kernel matrix to obtain a conditional information gain kernel matrix with noise ridge terms; wherein, the observation noise ridge terms obtained from historical reconstruction residual estimation are used to characterize the intensity of historical observation noise caused by depth acquisition error, local reflection, tissue exudation highlighting, boundary blurring, and occlusion loss during wound 3D reconstruction; the historical reconstruction residuals include depth reconstruction residuals, wound boundary reconstruction residuals, and reconstruction uncertainty of missing regions in multiple consecutive historical reconstruction cycles; by performing statistical estimation and smoothing update on the historical reconstruction residuals in multiple consecutive historical reconstruction cycles, the observation noise ridge terms corresponding to the current reconstruction cycle are obtained, and the ensemble bifurcation kernel matrix is ​​diagonally stabilized using the observation noise ridge terms to reduce the impact of local abnormal observation noise on the conditional information gain reward calculation results.

[0018] Construct a set of divergence kernel matrices based on the prediction bias vectors of different candidate reconstruction actions: ; in, Represents the set of branching kernel matrix The Middle Line 1 The elements of the column are used to represent the first... The candidate reconstruction action and the first The degree of correlation between the predicted discrepancies among the candidate reconstruction actions; This indicates the number of one-step predictors in the set of dynamic predictors for reconstruction. Indicates the one-step predictor number; Indicates the first One-step predictor for The prediction bias vector generated by each candidate reconstruction action is used to describe the degree of deviation of the predictor's output from the average prediction result of multiple predictors. Indicates the transpose symbol; Indicates the first The one-step predictor targets the first... The prediction bias vector generated by each candidate reconstruction action; Step B4: Perform Cholesky decomposition on the conditional information gain kernel matrix with noisy ridge terms, and generate the intrinsic reward value of conditional information gain for each candidate reconstruction action based on the decomposed diagonal terms, so that the reward of the current candidate reconstruction action is simultaneously constrained by historical wound reconstruction experience and the prefix redundancy of the current action. Based on the redundancy relationship between candidate reconstruction actions and the current action prefix, the formula for determining the conditional information gain intrinsic reward value of candidate reconstruction actions is as follows: ; in, Indicates the first One candidate reconstruction action The corresponding conditional information gain intrinsic reward value; Represents the first branch in the set of the kernel matrix. The diagonal elements corresponding to each candidate reconstruction action are used to characterize the prediction divergence strength of that candidate reconstruction action itself. This represents the observation noise ridge term obtained from the historical reconstruction residual estimation, used to suppress random prediction disturbances caused by local reflections, occlusion, depth acquisition errors, and exudate highlights; Indicates the first The discrepancy correlation vector between each candidate reconstruction action and the evaluated candidate reconstruction actions in the current action prefix. This represents the inverse matrix of the conditional information gain kernel matrix corresponding to the current action prefix. Indicates the first The prediction redundancy that can be explained by the current action prefix in each candidate reconstruction action is used to suppress repeated exploration of uncertain wound areas that have been covered by historical reconstruction actions. Step B5: Based on the reconstructed mesh corresponding to the current wound 3D Gaussian trigonometric unit set, and the Chamfer distance, boundary coverage, and missing region completion rate between the mesh and the target wound observation point cloud in the wound spatiotemporal point cloud sequence, calculate the external reward value for wound geometric integrity; generate a comprehensive reconstruction reward value based on the conditional information gain intrinsic reward value and the external reward value for wound geometric integrity, and update the reinforcement learning policy network using the comprehensive reconstruction reward value; Step B6: Use the updated reinforcement learning policy network to perform action selection processing on the candidate reconstruction action set to generate the target reconstruction action.

[0019] Example 5, based on Example 4, describes the process of generating a denser 3D Gaussian primitive set for the wound using the 3D Gaussian reconstruction module, which includes the following steps: Step E1: Generate a wound rendering image based on the target reconstruction action and the current wound 3D Gaussian primitive set. Based on the rendering error between the wound rendering image and the corresponding real observation image, perform gradient backpropagation processing on the position parameters of each wound 3D Gaussian primitive to obtain the primitive position gradient. Each wound 3D Gaussian primitive in the current wound 3D Gaussian primitive set includes center position, covariance, color, transparency, and scale parameters. The primitive position gradient is used to characterize the geometric correction direction of the primitive to the current rendering error. Step E2: Perform first-order moment estimation on the primitive position gradient to generate primitive gradient momentum vector, so that the wound geometric correction signal with consistent direction is accumulated, while the instantaneous noise gradient caused by random viewpoint, local reflection, exudate highlight and occlusion is canceled. Step E3: Perform second-order moment estimation on the primitive position gradient to generate a primitive gradient fluctuation estimate, and calculate the intrinsic signal-to-noise ratio of the primitive based on the primitive gradient momentum vector and the primitive gradient fluctuation estimate; Intrinsic signal-to-noise ratio formula for elementary elements: ; in, Indicates the first The three-dimensional Gaussian element of the wound in the first The intrinsic signal-to-noise ratio of the primitive corresponding to each reconstruction cycle is used to measure whether the geometric correction signal of the three-dimensional Gaussian primitive of the wound is stable and reliable; the larger the value, the more stable the gradient correction direction of the primitive is, the smaller the random disturbance, and the more suitable it is as a reliable compaction object. Indicates the first The three-dimensional Gaussian element of the wound in the first The primitive gradient momentum vector in each reconstruction cycle is obtained by estimating the first moment of the primitive position gradient. This represents the L2 norm of the elementary gradient momentum vector, i.e., the strength of the stable geometric correction signal; Indicates the first The three-dimensional Gaussian element of the wound in the first The primitive gradient fluctuation estimate in each reconstruction cycle is obtained by estimating the second moment of the primitive position gradient; Indicates a stable term; express The three-dimensional Gaussian element of the wound in the first The wound boundary proximity enhancement factor in each reconstruction cycle is used to prioritize the densification of Gaussian elements near the wound boundary. Since the wound boundary typically determines wound contour, area, depth assessment, and healing trend analysis, Gaussian elements near the boundary require higher reconstruction accuracy. express The three-dimensional Gaussian element of the wound in the first The enhancement factor for wound depressions during each reconstruction cycle is used to prioritize the densification of Gaussian elements in wound depressions, tissue defects, grooves, or low-lying areas. These areas typically exhibit significant three-dimensional morphological variations and are important regions for assessing wound depth, volume, and the extent of tissue loss. Step E4: Generate a context-adaptive candidate densification set based on the quantile ranking of the primitive gradient momentum vector in the current wound 3D Gaussian primitive set. In this embodiment, the context-adaptive candidate densification set is obtained not by selecting Gaussian primitives using a fixed gradient threshold, but by ranking the primitive gradient momentum distribution in the current wound 3D Gaussian primitive set according to the quantile. Its purpose is to dynamically select the primitives that need more refinement based on the current overall reconstruction state in each reconstruction cycle, rather than using a fixed threshold. Step E5: Perform SNR gating on the context-adaptive candidate compaction set using the intrinsic SNR of the primitives to obtain a reliable compaction primitive set; only wound 3D Gaussian primitives that simultaneously satisfy the quantile sorting condition and the intrinsic SNR threshold condition are added to the reliable compaction primitive set; when a primitive has a high gradient magnitude but severe directional fluctuations, its intrinsic SNR is low, so compaction will not be triggered. Step E6: For each wound 3D Gaussian primitive in the reliable compacted primitive set, determine the splitting or cloning operation based on the spatial scale of the wound 3D Gaussian primitive, and perform selective transparency reset processing on wound 3D Gaussian primitives with an intrinsic signal-to-noise ratio lower than a preset signal-to-noise ratio threshold; based on the splitting operation, cloning operation, and selective transparency reset processing, update the current wound 3D Gaussian primitive set to obtain a compacted wound 3D Gaussian primitive set; the splitting operation is used to process primitives with larger spatial scales, the cloning operation is used to retain primitives with smaller scales but reliable geometric correction signals; the selective transparency reset processing is used to suppress noise-dominated Gaussian primitives, without performing a uniform transparency reset on all primitives.

[0020] Example 6, based on Example 5, describes the process by which the wound 3D model output module generates the wound 3D model reconstruction result and wound reconstruction feedback data, specifically including the following steps: Step S61: Perform multi-view rendering and surface extraction processing based on the dense wound 3D Gaussian tuple to obtain a 3D surface model of the wound; multi-view rendering is used to check the color consistency and geometric continuity of the wound 3D Gaussian tuple under different viewing directions, and surface extraction processing is used to obtain a wound surface mesh or point cloud surface that can be used for measurement from the Gaussian tuple representation. Step S62: Calculate the wound boundary curve, wound surface area, maximum wound depth, wound volume, edge slope distribution, and wound color texture map based on the 3D surface model of the wound to generate the 3D model reconstruction result of the wound; the wound boundary curve is used to describe the boundary between the wound and the surrounding skin, the wound surface area and wound volume are used to quantify the wound size, the maximum wound depth and edge slope distribution are used to characterize the degree of depression and the state of edge elevation, and the wound color texture map is used to help record the color changes of granulation tissue, necrosis, exudation, and surrounding skin; Step S63: Register the 3D model reconstruction result of the wound with the subsequently collected wound feedback observation data, and calculate the reconstruction error map, boundary deviation map and depth deviation map; the reconstruction error map is used to identify areas where there is a large difference between the rendering result and the actual observation, the boundary deviation map is used to identify areas where the wound boundary estimation is off, and the depth deviation map is used to identify areas where the wound depression estimation is unstable. Step S64: Generate wound reconstruction feedback data based on the reconstruction error map, boundary deviation map, and depth deviation map; the wound reconstruction feedback data includes the location of the feedback area, the type of feedback error, the intensity of the feedback error, and the type of suggested action; the suggested action type is to re-sample, increase densification, reduce the weight of noise primitives, update the local texture, or terminate reconstruction.

[0021] The present invention and its embodiments have been described above. This description is not restrictive. The accompanying drawings are only one embodiment of the present invention, and the actual structure is not limited thereto. In short, if a person skilled in the art is inspired by this description and designs a similar structure and embodiment without departing from the spirit of the present invention, such design should fall within the protection scope of the present invention.

Claims

1. A wound 3D model reconstruction system based on reinforcement learning, characterized in that, The system includes: The wound data access module generates a spatiotemporal point cloud sequence of the wound. The tissue context encoding module performs spatiotemporal point cloud patch encoding, cross-time-step attention aggregation, and contextual latent representation generation on the wound spatiotemporal point cloud sequence to obtain the wound tissue contextual latent representation. The reinforcement learning state construction module obtains the current three-dimensional Gaussian set of wound meta-elements and generates a wound reconstruction reinforcement learning state object by combining the latent representation of wound tissue context. Explore the decision-making module, construct a reinforcement learning policy network and a set of reconstruction dynamic predictors; construct a set of candidate reconstruction actions based on the wound reconstruction reinforcement learning state object, and input the wound reconstruction reinforcement learning state object and candidate reconstruction actions into the set of reconstruction dynamic predictors to generate a comprehensive reconstruction reward value, update the reinforcement learning policy network, and determine the target reconstruction action from the set of candidate reconstruction actions through the updated reinforcement learning policy network. The 3D Gaussian reconstruction module performs rendering error backpropagation, context adaptive moment estimation, signal-to-noise ratio gating, and structural densification on the current 3D Gaussian primitive set of the wound according to the target reconstruction action, generating a densified 3D Gaussian primitive set of the wound. The wound 3D model output module generates the wound 3D model reconstruction result based on the dense wound 3D Gaussian element set.

2. The wound 3D model reconstruction system based on reinforcement learning according to claim 1, characterized in that: The wound tissue context encoding module, which obtains a latent representation of the wound tissue context, specifically includes the following steps: Step A1: Combine the three-dimensional spatial coordinates and normalized time coordinates in the wound spatiotemporal point cloud sequence into four-dimensional spatiotemporal coordinates, and perform farthest point sampling processing based on the four-dimensional spatiotemporal coordinates to obtain the center of the wound spatiotemporal patch. Step A2: For each wound spatiotemporal patch center, perform a neighbor search process in the four-dimensional spatiotemporal coordinates to obtain the wound spatiotemporal patch, and perform Fourier position encoding and local pooling processing on the points in the wound spatiotemporal patch to generate a wound spatiotemporal patch token. Step A3: Construct a spatiotemporal Transformer encoder, input the wound spatiotemporal patch token into the spatiotemporal Transformer encoder, and generate the wound spatiotemporal token feature; Step A4: Perform attention pooling on the spatiotemporal token features of the wound to obtain the sequence latent representation corresponding to a single wound point cloud sequence; when there are multiple wound point cloud sequences for the same target wound, perform learnable soft weight aggregation on the multiple sequence latent representations to generate an initial context latent representation; Step A5: During the training phase, auxiliary supervision processing is performed on the initial contextual latent representation using wound tissue attribute labels, and symbolic distance field reconstruction supervision processing is performed on the wound spatiotemporal patch token using wound symbolic distance field labels to obtain the wound tissue contextual latent representation.

3. The wound 3D model reconstruction system based on reinforcement learning according to claim 2, characterized in that: Step A3 specifically includes: inputting the wound spatiotemporal patch token into the spatiotemporal Transformer encoder, and using a self-attention mechanism to model the spatiotemporal correlation between different wound spatiotemporal patch tokens, wound edges, wound depressions, reflective points in the exudate area, and occluded missing areas, thereby generating wound spatiotemporal token features.

4. The wound 3D model reconstruction system based on reinforcement learning according to claim 1, characterized in that: The candidate reconstruction actions in the candidate reconstruction action set include at least one of the following: next acquisition viewpoint adjustment action, illumination state adjustment action, point cloud re-acquisition action, Gaussian pixel densification action, and local reconstruction parameter update action.

5. The wound 3D model reconstruction system based on reinforcement learning according to claim 4, characterized in that: The process of acquiring the target reconstruction action specifically includes the following steps: Step R1: Input the wound reconstruction reinforcement learning state object and candidate reconstruction actions into a set of reconstruction dynamic predictors consisting of multiple one-step predictors, obtain the prediction result of each one-step predictor for the next reconstruction state, and calculate the prediction bias vector based on multiple prediction results. Step R2: Construct a set of divergence kernel matrices based on the prediction bias vectors corresponding to different candidate reconstruction actions, and introduce observation noise ridge terms obtained from historical reconstruction residual estimation into the set of divergence kernel matrices to obtain a conditional information gain kernel matrix with noise ridge terms; Step R3: Perform Cholesky decomposition on the conditional information gain kernel matrix with noisy ridge terms, and generate the intrinsic reward value of conditional information gain for each candidate reconstruction action based on the decomposed diagonal terms. Step R4: Based on the reconstructed mesh corresponding to the current wound 3D Gaussian trigonometric unit set, the Chamfer distance, boundary coverage, and missing region completion rate between the mesh and the target wound observation point cloud in the wound spatiotemporal point cloud sequence, calculate the external reward value for wound geometric integrity; generate a comprehensive reconstruction reward value based on the conditional information gain intrinsic reward value and the external reward value for wound geometric integrity, and update the reinforcement learning policy network using the comprehensive reconstruction reward value; Step R5: Use the updated reinforcement learning policy network to perform action selection processing on the candidate reconstruction action set to generate the target reconstruction action.

6. The wound 3D model reconstruction system based on reinforcement learning according to claim 5, characterized in that: The observation noise ridge term obtained from the historical reconstruction residual estimation is used to characterize the intensity of historical observation noise caused by depth acquisition error, local reflection, tissue exudation highlighting, boundary blurring and occlusion loss during the three-dimensional reconstruction of the wound.

7. The wound 3D model reconstruction system based on reinforcement learning according to claim 1, characterized in that: The process of obtaining the three-dimensional Gaussian element set of the compacted wound includes the following steps: Step E1: Generate a wound rendering image based on the target reconstruction action and the current wound 3D Gaussian primitive set, and perform gradient backpropagation processing to obtain the primitive position gradient; Step E2: Perform first-order moment estimation on the primitive position gradient to generate the primitive gradient momentum vector; Step E3: Perform second-order moment estimation on the primitive position gradient to generate the primitive gradient fluctuation estimate, and calculate the intrinsic signal-to-noise ratio of the primitive based on the primitive gradient momentum vector and the primitive gradient fluctuation estimate; Step E4: Generate a context-adaptive candidate compactification set based on the quantile sorting results of the primitive gradient momentum vector in the current wound 3D Gaussian primitive set; Step E5: Perform SNR gating on the context-adaptive candidate compaction set using the intrinsic SNR of the primitives to obtain a reliable compaction primitive set; Step E6: For each wound 3D Gaussian primitive in the reliable densification primitive set, determine the splitting or cloning operation based on the spatial scale of the wound 3D Gaussian primitive, and perform selective transparency reset processing on wound 3D Gaussian primitives with intrinsic signal-to-noise ratio lower than a preset signal-to-noise ratio threshold to obtain a densified wound 3D Gaussian primitive set.

Images