A novel method and system for measuring bladder volume using ultrasound.

Abstract

This invention discloses a novel method and system for ultrasound measurement of bladder volume. The method includes: acquiring dense point cloud data of the bladder from CT scans, and generating a dataset of incomplete meridians simulating ultrasound based on the characteristics of ultrasound imaging; constructing a point cloud completion model based on deep learning, training the model using CT data to achieve the mapping from incomplete point clouds to dense point clouds; and using the Crust algorithm to perform three-dimensional reconstruction and volume calculation on the completed dense point cloud. This invention corrects the systematic errors of the ultrasound system by leveraging the high precision characteristics of CT data, achieving high-precision measurement of bladder volume under radiation-free and low-cost conditions. The measurement error can be controlled within 10%, making it suitable for rapid and accurate assessment of bladder volume before radiotherapy.

Description

Technical Field

[0001] This invention relates to the field of medical image processing and ultrasound measurement technology, specifically to a method and system for ultrasound measurement of bladder volume that combines CT data-assisted point cloud completion and Crust reconstruction, which is particularly suitable for rapid, radiation-free assessment of bladder volume before pelvic radiotherapy. Background Technology

[0002] In precision radiotherapy systems, CT localization and irradiation for pelvic radiotherapy require maintaining approximate volume and distribution consistency among human organs. Consistent bladder volume is particularly crucial in this process, necessitating precise measurement of the patient's bladder volume before localization or treatment. While CT can accurately measure bladder volume, its high radiation dose, high cost, and complex operation make it unsuitable for repeated use before each radiotherapy session.

[0003] Ultrasound equipment offers radiation-free imaging, ease of operation, and low cost, making it more suitable for frequent monitoring. However, its two-dimensional imaging characteristics and probe angle limitations result in the incompleteness of the generated measurement curves, leading to significant deviations between actual volume estimations and clinical requirements. This is especially true when using handheld ultrasound equipment, where the actual clinical error can exceed 20%.

[0004] Traditional methods for calculating bladder volume using ultrasound images involve measuring the bladder's length, width, and height, then calculating using the formula "length × width × height × correction factor." The correction factor varies depending on the bladder shape, typically 0.72 (for unknown shapes), 0.66 for other shapes such as triangular prisms, 0.81 for cylinders (ellipsoids), 0.89 for cuboids, and 0.52 for spheres. This method relies on operator experience and assumptions about the bladder shape, leading to significant errors in testing irregularly shaped bladders.

[0005] In recent years, ultrasound three-dimensional reconstruction methods based on deep learning have gradually emerged. For example, Peng et al. (2024) proposed a bladder volume measurement method based on weakly supervised deep learning, which reconstructs the three-dimensional structure of the bladder through 12 two-dimensional ultrasound images. However, this method still relies on a large amount of labeled data and has high requirements for image quality.

[0006] Furthermore, while existing ultrasound equipment, such as bladder volume analyzers, can acquire multiple cross-sectional images simultaneously, they often suffer from issues like missing edges and image distortion. Post-loading ultrasound images are clear but have low acquisition efficiency. Additionally, current technologies generally overlook the data mismatch caused by changes in organ morphology during CT and ultrasound acquisition. As a deformable hollow organ, the bladder exhibits significant differences in geometry between CT scans (typically in a supine, full position) and ultrasound scans (potentially due to probe pressure and changes in body position leading to localized non-rigid deformation). If dense CT point clouds are directly used as supervisory signals to train the ultrasound completion network, the network will learn a distorted mapping relationship, thus affecting the accuracy of the final volume measurement. Therefore, there is an urgent need for a bladder volume measurement method that retains the radiation-free advantages of ultrasound, approaches the accuracy of CT, and effectively addresses probe pressure deformation. Summary of the Invention

[0007] To address the issues of existing CT scans increasing radiation exposure, high costs, and complex procedures, making them unsuitable for frequent repetition before radiotherapy, while commercial ultrasound uses simplified geometric models and suffers from large measurement errors, this invention provides a method and system for ultrasound measurement of bladder volume that combines CT data-assisted point cloud completion and Crust reconstruction. By introducing a data preprocessing strategy based on non-rigid deformation registration and combining it with the high-precision characteristics of CT data to correct systematic errors in the ultrasound system, this invention achieves radiation-free, low-cost, high-precision, repeatable bladder volume assessment that effectively adapts to clinical probe pressure deformation. The measurement error can be controlled within 10%, making it suitable for rapid and accurate assessment of bladder volume before radiotherapy.

[0008] To achieve the above objectives, the present invention adopts the following technical solution: This invention provides a novel method for measuring bladder volume using ultrasound, comprising the following steps: S1. Constructing a CT-like ultrasound cross-section dataset: Three-dimensional CT images of the bladder are acquired and high-precision dense point cloud data is generated. To overcome the differences in non-rigid deformation caused by probe pressure and other factors during CT and ultrasound acquisition of the bladder geometry, this invention performs deformation preprocessing on the dense point cloud data. Specifically, a bladder geometric model is constructed using Blender 3D modeling software. Based on the principal component analysis (PCA) algorithm, the principal component directions of the bladder point cloud data are calculated to determine the main direction axes of deformation. Using the pressure center axis as a reference, the model is stretched or compressed at multiple angles and to varying degrees along the meridian direction to simulate probe pressure deformation in different directions and intensities. Through this operation, 120 original CT data samples are augmented from multiple angles, generating a total of 360 deformation samples along the meridian direction. Finally, a total of 3335 training data sets are constructed, including the original data and the deformation-processed data. Subsequently, a meridian dataset simulating ultrasound scanning is extracted from the deformed point cloud data. The meridian dataset includes incomplete and complete meridian datasets, and incomplete-dense point cloud data pairs are established.

[0009] S2. Training a multi-path point cloud completion model: A deep learning network is used to train the incomplete-dense point cloud data pair to obtain the mapping weights from the incomplete point cloud to the complete point cloud.

[0010] S3. Ultrasonic Data Quality Grading and Preprocessing: Acquire actual ultrasound scan data and classify them into high-quality image sets and low-quality image sets according to image quality; for high-quality image set data, perform contour registration directly to generate partial point clouds; for low-quality image set data, perform 2D boundary completion and stabilized 3D assembly to generate partial point clouds.

[0011] S4. Ultrasonic Data 3D Reconstruction: The generated partial point cloud is input into the trained multi-path point cloud completion model to complete it into a 3D dense point cloud.

[0012] S5. Bladder volume calculation: The Crust algorithm is used to perform surface reconstruction and volume integration on the completed 3D dense point cloud to obtain the estimated bladder volume.

[0013] Further, in step S1, the generation of the complete meridian dataset includes: Determine the central axis of the bladder, which includes the coordinate axes in the three-dimensional coordinate system and the diagonal direction of the cube; For each central axis, an initial normal vector is generated, and the initial normal vector is rotated around the central axis. A normal vector is obtained every 15° of rotation, for a total of 12 normal vectors. For each normal vector, a cross-sectional plane passing through the center of the bladder and perpendicular to the normal vector is determined. The set of points intersecting with the cross-sectional plane is extracted from the dense point cloud data to form a complete meridian.

[0014] Further, in step S1, the generation of the incomplete meridian dataset includes: collecting incomplete contour samples existing in actual ultrasound equipment, sorting and removing regions of the contour points, simulating different missing modes, and mapping the two-dimensional incomplete contours to three-dimensional space to form incomplete meridian data.

[0015] Furthermore, during the generation of the incomplete meridian dataset, the contour points in the ultrasound image are sorted counterclockwise. Using the top point as a reference, the point set is batch-removed according to the preset incomplete interval to simulate the contour loss caused by probe obstruction, tissue interference, or equipment limitations in clinical practice.

[0016] Furthermore, in step S4, the multi-path point cloud completion model includes at least one of the following completion paths: Path 1: For high-quality image datasets, input the relatively complete data of the 12 meridians into the deep learning model and directly output dense point cloud data; Path 2: For low-quality image set data, the incomplete 12 meridian data are first repaired into complete 12 meridian data through image segmentation or traditional geometric algorithms, and then input into the deep learning model to output dense point cloud data.

[0017] Furthermore, the point cloud completion model adopts an architecture based on Transformer or convolutional neural networks and uses chamfer distance as the loss function for training.

[0018] Furthermore, in step S5, the Crust algorithm includes: Delaunay triangulation is performed on dense point clouds. The concave and convex structures are identified by judging the positional relationship between the center of the circumcircle of adjacent triangles and the common chord. Triangular facets of concave regions are retained, invalid facets are removed, and a closed triangular mesh of the bladder surface is constructed. The total volume of the bladder can be calculated by constructing a tetrahedron from the surface triangular facets and the internal reference points and summing their volumes, or by directly summing the volumes of the internal tetrahedrons.

[0019] Furthermore, it also includes using a phantom to calibrate ultrasound-CT data before model training in order to estimate systematic errors and perform background correction.

[0020] The present invention also provides a system for performing the above-described method, comprising: The data acquisition module is used to acquire CT three-dimensional image data and ultrasound image data of the bladder; The point cloud generation module is used to generate dense point cloud data of the bladder based on CT 3D image data; The non-rigid deformation registration module is used to preprocess dense point cloud data to simulate probe pressure deformation. The meridian dataset generation module is used to extract a meridian dataset simulating ultrasound scanning from the dense point cloud data based on the principle of ultrasound imaging. The ultrasound data quality grading module is used to grade actual ultrasound images into high-quality / low-quality image sets and perform corresponding preprocessing to generate partial point clouds. The point cloud completion module includes a deep learning model used to complete partial point clouds into dense point cloud data. The 3D reconstruction and volume calculation module uses the Crust algorithm to reconstruct the 3D surface of the completed dense point cloud data and calculate the bladder volume.

[0021] Compared with the prior art, the present invention has the following beneficial effects: (1) In traditional bladder volume measurement, CT and ultrasound are two independent and interchangeable modalities. The core innovation of this invention lies in breaking down this modal barrier and constructing a collaborative paradigm of "CT offline training and ultrasound online application". Its innovation lies not only in the simple combination of data, but also in the construction of a high-precision "three-dimensional prior knowledge base" using CT data. Through a deep learning model, the rules and details about the complex morphology of the bladder (such as common depressions, protrusions, and asymmetries) learned from CT data are encoded as network weights. When the model processes incomplete ultrasound data, this prior knowledge is activated to guide and constrain the three-dimensional reconstruction process, thereby accurately "inferring" structures close to the accuracy of CT on ultrasound images lacking complete three-dimensional information.

[0022] (2) The multi-path point cloud completion strategy proposed in this invention designs differentiated processing paths for ultrasound data with different levels of completeness, improving the robustness and clinical applicability of the system. It does not simply provide multiple independent models, but constructs an intelligent system that can adaptively select or weightedly fuse the optimal processing flow based on the characteristics of the input data (such as point cloud density, distribution continuity, and size of missing regions). This multi-path point cloud completion strategy mimics the processing thinking of experienced physicians when facing images of different quality—directly evaluating images of good quality, repairing partially missing images first, and requiring a more comprehensive judgment for images of very poor quality. This data-driven, expert-decision-simulating flexible processing framework significantly improves the system's adaptability to complex clinical environments and ensures the stability and reliability of measurements.

[0023] (3) General 3D reconstruction algorithms are prone to producing holes, self-intersections, or volume calculation errors when dealing with thin-walled, irregular soft tissue organs such as the bladder that may have local depressions. This invention uses an improved Crust algorithm to deeply analyze the geometric features of the concave and convex areas on the bladder surface. It takes the geometric criterion of "the spatial positional relationship between the center of the circumcircle of adjacent triangles and the common chord" as the core rule for identifying and filtering erroneous triangular facets, accurately distinguishing between real depressions and reconstruction artifacts. Based on the mature Delaunay triangulation, local constraints based on medical image priors are introduced to form a hybrid algorithm of "global optimal triangulation + local geometric verification". This algorithm customization and optimization for specific organ anatomical features improves the anatomical accuracy and volume calculation precision of the reconstruction.

[0024] (4) The present invention introduces a phantom calibration step, using a phantom with known physical properties, and takes CT measurement value as the "gold standard" to perform overall system-level calibration of the entire process of ultrasound measurement (from image acquisition to algorithm output), thereby improving the repeatability and measurement consistency of the method.

[0025] (5) The present invention achieves seamless full-process automation from ultrasound equipment image acquisition, DICOM data transmission, automatic contour extraction / segmentation, deep learning completion reconstruction, intelligent volume calculation, to final result visualization and report generation. This minimizes the manual intervention links and eliminates the source of subjective error.

[0026] (6) This invention innovatively introduces non-rigid deformation registration based on the PCA algorithm in the dataset construction stage. This strategy is the first to systematically incorporate the clinically relevant factor of bladder deformation caused by ultrasound probe pressure into the training process in a data-driven manner. By generating 360 deformation samples along the meridian direction and constructing an enhanced training set containing 3335 data points, this invention fundamentally solves the essential mismatch between CT gold standard data and ultrasound acquisition data in terms of geometric morphology. This enables the deep learning model to learn the precise recovery rules from the state of pressure deformation to the true complete morphology, significantly improving the model's generalization ability and robustness in complex clinical scenarios, and providing a key data foundation for high-precision bladder capacity measurement. Attached Figure Description

[0027] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0028] Figure 1 CT scan images of the phantom; Figure 2 Images obtained from a bladder volume analyzer for the phantom; Figure 3 Images obtained from handheld ultrasonic testing of the phantom; Figure 4To simulate ultrasound imaging point cloud sampling of CT data, 12 cross sections resembling ultrasound were collected from the complete point cloud data; Figure 5 To show the completion effect of CT scans trained by the network, yellow dots represent complete CT data, and blue dots represent sampled cross-sectional data. Figure 6 This is a rarefaction cloud image from a 3D ultrasound study; the blue area represents the 3D incomplete or missing portion. Figure 7 A three-dimensional handheld ultrasound reconstruction image; Figure 8 These are common bladder ultrasound images seen in patients over the past two years; Figure 9 CT data (top) and ultrasound data (bottom) were obtained. Figure 10 Ultrasound images acquired by a handheld ultrasound device (top) and a bladder volume meter (bottom); Figure 11 To perform semi-automatic bladder segmentation in 3DSlicer (a), 12 complete meridian sections were obtained by ultrasound-like sampling of CT data (b) (c). Figure 12 The images show the bladder surface reconstruction model (a) using the Alpha Shape algorithm and the bladder surface reconstruction model (b) using the Crust algorithm. Detailed Implementation

[0029] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0030] This invention provides a method for calculating bladder volume during radiotherapy using B-mode ultrasound. This method can obtain a relatively accurate volume for bladders of different shapes, with the measurement error controlled within 10%.

[0031] The core technical idea is to use a combination of deep learning and traditional reconstruction methods, which utilizes CT data to assist in point cloud completion and Crust reconstruction to calculate bladder volume during radiotherapy. The specific steps are as follows: S1. Using a retrofitted ultrasonic scanner, point cloud data is generated by scanning the phantom. The volume of the phantom is reconstructed from the point cloud data using the Crust algorithm. The phantom is then placed on a localized CT scanner for measurement, obtaining the volume from the CT data. The volume obtained from ultrasound and the volume obtained from CT are compared, with the CT volume used as the gold standard, to calibrate the ultrasound data and estimate background error. The result is as follows: Figure 1-3 The image shown ( Figure 1 These are the three views of the test phantom, obtained using CT. Figure 2 Images obtained from the bladder volume meter of the phantom. Figure 3 The images of the phantom obtained by handheld ultrasound were analyzed to identify errors caused by image distortion in two ultrasound devices (bladder volume meter and handheld ultrasound). The bladder volume meter showed obvious image concave distortion and missing edges, which often accompanied image quality problems such as missing edges in clinical operation, resulting in poor three-dimensional reconstruction effect. Clinically, the volume accuracy was not high and the repeatability was poor.

[0032] S2. Collect the patient's corresponding ultrasound and CT data (see...) Figure 9 A dataset was created, and the CT data was elastically corrected based on the ultrasound data. Then, based on the format of the ultrasound data, ultrasound-like data extraction and simulation were performed on the CT data, dividing it into incomplete data and dense data (see...). Figure 10 (This refers to the database of CT data, and there is a one-to-one correspondence between the two databases.)

[0033] To construct a high-quality training dataset, deformation preprocessing was first performed on the 120 dense point cloud data obtained from CT scans.

[0034] (1) Deformation tool and principal direction determination: The operation is performed in Blender 3D modeling software. For each CT point cloud data, the principal component analysis (PCA) algorithm is first applied to calculate the three principal component directions of the point cloud data. The first principal component direction is determined as the main direction of bladder deformation (i.e., the principal axis direction of probe pressure), which is used as the pressure center axis.

[0035] (2) Multi-angle deformation amplification: Based on the determined pressure center axis, the model is non-rigidly deformed along the meridian direction in Blender to simulate the local compression or stretching of the bladder when the ultrasound probe applies pressure at different angles. Specifically, starting from the center axis, the CT model is augmented with data from multiple angles, generating a total of 360 meridians, and each meridian is subjected to different degrees of deformation processing.

[0036] (3) Dataset construction: After this operation, the original 120 CT data were expanded into a total of 3335 training data, including those without deformation and those with multiple deformation states. This step ensured that the "incomplete meridian" data in the training set were highly matched with the real ultrasound acquisition data in terms of morphology, laying the foundation for the network to learn accurate geometric mapping in the future.

[0037] To address the inherent variability in acquisition quality, two distinct processing flows were implemented: (1) High-quality image set data: applied to scans with relatively complete 2D cross-sections and clear bladder boundaries. For these scans, the contours were semi-automatically delineated and geometrically registered to a common 3D coordinate system, generating an initial sparse point cloud that preserved global shape features and cross-sectional consistency. (2) Low-quality image set data: applied to scans with fragmented, discontinuous, or partially missing boundaries due to noise, attenuation, and artifacts. A two-stage stabilization flow was used: (i) an efficient 2D boundary completion step to restore contour continuity and suppress gaps, followed by (ii) a coarse 3D assembly stage with optional point set refinement to enhance spatial consistency and reduce outliers. This quality grading approach robustly normalizes heterogeneous ultrasound observations, providing well-conditional partial point clouds for subsequent learning-based shape completion.

[0038] S3, where the incomplete data is further divided into complete datasets of the 12 meridians (see...). Figure 11 ) and an incomplete dataset of 12 meridians (obtained by processing the complete meridians).

[0039] (1) The complete meridian dataset was obtained by geometrically segmenting the dense point cloud. The bladder was modeled as a three-dimensional point cloud structure, with its central axis representing the reference direction during image acquisition. The selection of the central axis needed to cover multiple possible directions to simulate the movement and rotation of the actual ultrasound probe. Multiple cross-sectional planes were generated through rotational normal vectors, with each plane corresponding to a meridian, ensuring that the 12 meridians were evenly distributed within a 360-degree range. See Figure 4 .

[0040] Specifically, it includes: Identify the bladder's central axis: Define the central axis as the main axes in a three-dimensional coordinate system, including the X-axis, Y-axis, Z-axis, and the four diagonals of the cube model (i.e., the diagonal vectors from the origin to the vertices of the cube). The diversity of these axes is used to simulate the angular changes of the ultrasound probe during measurement, such as tilting or rotating the probe in different directions. Assume the bladder center is located at the origin (0,0,0), and all axes pass through this point.

[0041] Calculate the initial normal vector: For each selected central axis, choose a base vector that is not collinear with it (e.g., if the central axis is the Z-axis, then the X-axis vector can be chosen as the base). Calculate the initial normal vector using an orthogonalization method, ensuring it is perpendicular to the central axis. The initial normal vector is defined as the starting direction (rotation angle of 0 degrees).

[0042] Generate multiple normal vectors and cross-sectional planes: Rotate 15 degrees at a time around the central axis, starting from the initial normal vector (i.e., 360 degrees / 24 = 15 degrees, but actually generate 12 directions, possibly through selective sampling to achieve uniform distribution). The rotation process starts at 0 degrees, generating 12 normal vectors sequentially: 0°, 15°, 30°, ..., 165° (Note: Due to symmetry, the 12 directions cover a 180-degree range, but full coverage must be ensured). Each normal vector defines a cross-sectional plane: this plane passes through the origin (bladder center) and is perpendicular to the normal vector.

[0043] Point cloud segmentation: For each cross-sectional plane, extract the set of points intersecting the plane from the dense point cloud to form a complete meridian. Ensure sufficient point cloud density (each meridian contains at least 86 points) and apply noise filtering to improve accuracy. Output 12 complete meridian datasets, each meridian represented as an ordered sequence of points for subsequent 3D reconstruction.

[0044] (2) The incomplete meridian dataset was generated by processing ultrasound images acquired by a bladder volume analyzer. Ultrasound images are often incomplete (black border areas) due to probe position or tissue obstruction. By searching and processing the contours of these images, incomplete meridians were simulated. After sorting the contour points by angle, the regions were removed to generate 12 incomplete contours, which were then converted into three-dimensional meridian data.

[0045] Specifically, it includes: Image search and sample confirmation: Search for images with black borders on both sides from the set of ultrasound images acquired from the bladder volume analyzer, and confirm ten common sample images that should represent typical incomplete patterns (such as asymmetrical black borders, partial occlusion). For each sample, extract the bladder contour: use threshold segmentation to obtain a two-dimensional point set.

[0046] like Figure 8 As shown in the figure, the ten common incomplete samples for bladder volume analyzers are ten relatively common samples found from patients using bladder volume analyzers in the hospital in the past two years. These ten samples are used as a template for sampling CT data.

[0047] Contour point numbering and sorting: Identify the "top point" as the reference: prioritize the point with the maximum Y-coordinate; if the Y-coordinates are equal, select the point with the minimum X-coordinate. Sort the point set in a counterclockwise direction to form an ordered sequence (index starts from 1).

[0048] Region Data Removal: Define 12 preset index intervals (based on cardinality 1), which correspond to the missing regions to be removed. For example, the intervals can be [1-10, 20-30, ...], specifically predefined according to the sample's missing pattern. For each preset interval, batch remove the corresponding points in the sorted point set, outputting 12 incomplete contours, each contour being a sequence of removed points, simulating different degrees of incompleteness.

[0049] 3D Placement and Meridian Generation: Each incomplete contour is placed in 3D space: assuming the contour is a projection onto the cross-sectional plane, it is placed along the central axis. A transformation matrix is ​​applied to expand the 2D contour into 3D meridians, for example, by rotation and translation to ensure alignment with the complete meridian dataset. This outputs a dataset of 12 incomplete meridians, each represented as a sequence of incomplete meridians, used for comparative analysis and model training.

[0050] S4. The incomplete data is supplemented into complete data through traditional algorithms and deep learning point clouds, transforming incomplete point clouds into dense point clouds.

[0051] The above specific process can be divided into two types: (1) The relatively complete data of 12 meridians were supplemented into a dense point cloud through deep learning.

[0052] The deep learning model implements an end-to-end system that takes a partial point cloud as input and outputs a complete point cloud. It mainly consists of four parts: data processing, model building, training, and testing.

[0053] Data processing: Point cloud data is read from the file and normalized (point cloud coordinates are standardized to a specific range) to ensure that the number of input and output point clouds is fixed (512 and 2048 points respectively). For point clouds with too many points, random downsampling is performed to reach the target number; for point clouds with insufficient points, existing points are randomly copied to reach the target number. Data diversity is increased and the model's generalization ability is improved by random rotation (small-angle rotation around three axes) and the addition of Gaussian noise.

[0054] Model Architecture: A Transformer-based point cloud completion model. First, 3D coordinates (x, y, z) are mapped to a high-dimensional feature space (128-dimensional) to stabilize the training process and accelerate convergence. A self-attention mechanism is used to capture the global relationships between points, comprising two encoder layers, each with an 8-head attention mechanism, mapping the Transformer output features to 2048 3D point coordinates (2048×3). The core idea of ​​the model is to utilize the Transformer's attention mechanism to understand the spatial relationships between points in the point cloud, thereby inferring the complete structure from a partial point cloud. Chamfer distance is used as the loss function, a commonly used loss function in point cloud processing: calculating the distance from each point in the predicted point cloud to the nearest point in the real point cloud, and calculating the distance from each point in the real point cloud to the nearest point in the predicted point cloud, taking the average of these two distances as the loss value. This loss function effectively measures the similarity between two point clouds, ensuring that the predicted result closely approximates the true value in terms of spatial structure.

[0055] Training process: Using the Adam optimizer and cosine annealing learning rate scheduler, the data is reloaded at each epoch (this can be modified to a more efficient one-time loading method), the chamfer distance between the predicted point cloud and the real point cloud is calculated and the parameters are updated by backpropagation, and the training loss curve and the final model weights are saved.

[0056] Testing and Visualization: Load the trained model, perform completion predictions on the new partial point cloud, and visualize the predicted complete point cloud using the Open3D library. See [link / description]. Figure 5 ( Figure 5 Yellow point clouds represent complete point clouds, while blue point clouds represent partial point clouds.

[0057] (2) The incomplete 12 meridian data are processed by traditional or deep learning algorithms to become complete 12 meridian data, and then the data is completed into a dense point cloud through deep learning.

[0058] To address the issue of incomplete 2D meridian images, a preprocessing method based on medical image segmentation is employed. Since images acquired by bladder volume analyzers often exhibit edge defects (manifested as black boundary regions), this method introduces a manual completion step before training the segmentation model to restore the incomplete image to a complete 2D meridian. This preprocessing approach ensures that the weights of the final trained segmentation model inherently possess image completion capabilities, thereby achieving end-to-end segmentation of the original incomplete image.

[0059] The complete training process of the medical image segmentation model mainly revolves around the segmentation task of bladder-related images, and is divided into data processing, model building, training process, and result recording.

[0060] Data processing: Set data preprocessing rules, including center cropping of images (to unify input size), converting NumPy format data to tensors (to adapt to PyTorch computation), and format conversion of masks, to ensure that the data format input to the model is consistent. Based on a custom bladder dataset class, construct training and validation sets respectively.

[0061] Model Building: Select the UNet network based on the configuration, set the number of output classes (matching the classes in the bladder dataset), and load the model onto the GPU for accelerated computation. Loss Function and Optimizer Configuration: Select the Dice loss function suitable for medical segmentation tasks, paired with the specified Adam optimizer, and set the initial learning rate and weight decay parameters for updating model parameters. Configure the learning rate scheduler to dynamically adjust the learning rate to optimize convergence; simultaneously initialize an early stopping mechanism to avoid overfitting, automatically stopping training when validation metrics show no improvement for several consecutive rounds.

[0062] Training Process: In each training epoch, the model is switched to training mode, the training dataset is traversed, and forward propagation is performed on the input images to obtain the model's segmentation prediction results. The loss between the prediction results and the ground truth mask is calculated, and the model parameters are updated through backpropagation. The training loss and Dice coefficients for each category (core metrics for evaluating segmentation accuracy) are calculated and recorded in real time to monitor the training effect. After each training epoch, the model is switched to evaluation mode (gradient calculation is disabled to save resources), the validation dataset is traversed, and forward propagation is performed on the validation images to obtain the prediction results. The validation loss and Dice coefficients are calculated to evaluate the model's generalization ability on unseen data. The validation metrics are written to the log for subsequent analysis and early stopping detection.

[0063] Results Recording: After each round of validation, the average Dice coefficient of the validation set is used to determine whether an early stopping mechanism is triggered (if the metric does not improve for several consecutive rounds or the learning rate is too low). If triggered, training is stopped to avoid ineffective training and overfitting. After training terminates, the final saved optimal model round number, the training round number at which training stopped, and other key information are output, completing the entire training process.

[0064] After segmentation and 2D completion, the 2D image is arranged in 3D, followed by 3D reconstruction, see... Figure 6 and Figure 7 .

[0065] In terms of model structure, an improved PoinTr framework was adopted: Input projection layer: First, the three-dimensional coordinates are mapped to a high-dimensional feature space through linear transformation, providing a foundation for subsequent feature extraction.

[0066] The CNN module performs point-by-point nonlinear mapping on point features through stacked layers of one-dimensional convolutions (Conv1d), batch normalization (BatchNorm), and nonlinear activation (ReLU). This module can efficiently extract local geometric patterns and improve the stability of feature representation, essentially performing "local shape encoding" at the point cloud level.

[0067] The Transformer encoder, after obtaining the local features of points, uses a self-attention mechanism to capture the global dependencies between points. This allows the network to not only focus on the local details of the point cloud but also integrate the overall structural information of the point cloud, thus learning a more globally consistent point cloud representation.

[0068] Decoder: The globally aggregated features are further mapped to the target point cloud (2048×3), enabling prediction from incomplete point cloud to complete point cloud.

[0069] During training, Chamfer distance is used as the loss function to measure the geometric difference between the predicted and real point clouds. This is combined with the Adam optimizer, cosine annealing learning rate scheduling strategy, and early stopping mechanism to improve convergence speed and model generalization performance. After training, the system saves the parameters of the best-performing model and plots the loss curve for analysis. Overall, this network can automatically learn the mapping relationship from partial point clouds to complete point clouds, achieving high-quality completion of 3D point clouds.

[0070] S5. Using the Crust algorithm, calculate the volume of the dense point cloud, i.e., the bladder volume. The Crust algorithm calculates the bladder volume by constructing triangular prisms from the points in the point cloud data to fill the bladder outline and then summing the volumes of all the triangular prisms.

[0071] This invention measures bladder volume during radiotherapy by training CT bladder data to obtain weights from point cloud data ranging from incomplete to dense. Compared with existing technologies, which show greater error with larger bladder shapes when measuring human bladder volume, this invention achieves more accurate and consistent results, regardless of bladder shape, with an error of less than 10% between the measurement results and CT results (obtained by performing ultrasound scans before CT scans to verify the results).

[0072] This invention uses two instruments to acquire bladder images: a bladder volume analyzer and an afterloading ultrasound instrument. The bladder volume analyzer is a common instrument used in radiotherapy to measure bladder volume, capable of obtaining 12 cross-sectional ultrasound images at 15° intervals in a single measurement. However, the imaging is relatively poor, and large bladders may exhibit irradiation blind spots, resulting in black borders. The afterloading ultrasound instrument is mainly used in gynecologic oncology, providing clearer images, but it can only measure one cross-section at a time. Acquiring 12 cross-sections requires a relatively long collection period, and angle adjustment heavily relies on the technician's experience.

[0073] The CT data was acquired by a Siemens Healthineers CT computed tomography scanner scanning the patient. The software automatically delineated the data, then exported the bladder area in STL format. Finally, the data was placed in MeshLab software and a fixed number of point cloud data (2048 points) were obtained by applying a Poisson distribution to the bladder surface. This is the original dense point cloud data.

[0074] After acquiring the complete dense CT data, the dense point cloud data needs to be converted into an ultrasound-like format to obtain incomplete longitudinal data, similar to the method used for acquiring ultrasound images. Therefore, the central axes of the bladder are determined, encompassing the XYZ axes of the three-dimensional coordinate system and the four diagonals of the cube. Different central axes are chosen because the movement of the probe during ultrasound measurement will produce different angles, requiring adaptation to the actual measurement scenario. Next, a base vector that is not collinear with the current central axis is selected, and the initial normal vector is calculated. Then, this initial normal vector is rotated around the central axis by 15 degrees each time, starting from 0 degrees, resulting in 12 normal vectors with different directions. Each normal vector corresponds to a cross-sectional plane, and these cross-sections correspond to the 12 cross-sections of ultrasound.

[0075] After acquiring the incomplete data, there are three processing methods: (1) Train a deep learning model for ultrasound images with good imaging (clear and complete contours), directly training from incomplete meridians to complete dense point cloud data. (2) Train a deep learning model for ultrasound images with missing data, using traditional or deep learning algorithms to transform the incomplete 12 meridians into complete 12 meridians, and then using deep learning to complete the dense point cloud. (3) Directly use deep learning to complete the incomplete 12 meridians into a dense point cloud.

[0076] The incomplete model uses classic ultrasound cases to incompletely represent training data. The ultrasound image has 12 cross-sectional images, and the contours of these 12 ultrasound images are evenly divided into 86 points. Starting from the top point, the points are counted counterclockwise, and the intervals that do not show the complete contour of the bladder are recorded. When processing CT data, these template intervals are referenced to incomplete the points inside the contours of the CT data, thereby achieving an effect similar to ultrasound data.

[0077] Finally, the volume is obtained by calculating the dense point cloud. Since the bladder differs from a typical ellipsoid (it is a quasi-ellipsoid with indentations and varying shapes), this invention uses the Crust algorithm for volume calculation. The Delaunay triangulation volume (sumP1) method directly triangulates the original point cloud, dividing it into non-overlapping tetrahedrons. The volume of each tetrahedron is then calculated and summed to obtain a total volume estimate including the indented areas. Secondly, the MyCrust algorithm volume (vol) adds "shield" points to the outside of the original point cloud to enclose the space. It uses the relationship between the circumscribed sphere radius and the center point to determine whether each tetrahedron is inside the object, accumulating only the volumes of the inner tetrahedrons to obtain a volume estimate closer to the actual object boundary. Finally, the surface triangular mesh-based volume calculation (sumvlm1, sumvlm2) method constructs tetrahedrons from the surface triangles output by MyCrust and a reference point (such as the geometric center of the point cloud or any point), calculating the volume of each tetrahedron and summing them. This method relies on the reconstructed closed surface and is suitable for calculating the shell volume.

[0078] Current ultrasound methods for estimating bladder volume mainly include commercial handheld devices, flexible wearable devices, and research on novel algorithms. Literature shows that commercial systems such as Bladderscan, Bardscan, and 3D-US all exhibit significant correlations in estimating bladder emptying volume (r values ​​of 0.97, 0.98, and 0.99, respectively). 3D-US demonstrates the highest accuracy, with a Bland-Altman consistency limit of −28.9 to 40.0 mL. In contrast, Bardscan and Bladderscan have consistency limits of −64.5 to 107.2 mL and −73.7 to 88.4 mL, respectively, exhibiting larger systematic errors and greater dispersion. Furthermore, Bardscan underestimates urine volume by an average of 21.4 mL.

[0079] Because the bladder has deformable and typically non-convex geometry, coupled with sparse sampling, missing regions, and noise, directly calculating the volume from the raw point cloud is prone to errors. To address this issue, this application employs a geometry-driven two-step method: first, a closed (watertight) surface mesh is reconstructed, followed by volume calculation based on the mesh. Various surface reconstruction algorithms were tested on complete bladder point clouds with varying sparsity, including Delaunay triangulation and Alpha Shape (…). Figure 12 a) Power Crust and Crust algorithm.

[0080] Comparative analysis shows that while Delaunay triangulation is robust to well-sampled convex polyhedra, it introduces an artificial "bridging" phenomenon (i.e., incorrectly filling in depressions) when processing concave areas. Alpha Shape and PowerCrust, while effective in certain specific scenarios, are prone to producing holes or losing fine details under conditions of common noise and non-uniform sampling in anatomical data. Based on the above comparisons, the Crust algorithm was ultimately chosen for surface reconstruction. Figure 12 (b) because it can stably generate closed meshes from imperfect point sets while fully preserving the concave details in the anatomical structure.

[0081] In the volume calculation stage, comparative experiments were conducted on various reconstruction methods such as Alpha shape and Power crust to analyze the differences in bladder surface undulation recovery, closed surface construction, and volume error control. Based on this, this application selected and optimized the Crust algorithm to make it more suitable for bladder three-dimensional surface reconstruction and volume integration. Therefore, the application of the Crust algorithm in this application is not an equivalent substitution that can be arbitrarily made by those skilled in the art based on conventional knowledge, but rather a technical choice formed to address a specific technical problem.

[0082] The effectiveness of the present invention will be further verified below with specific experimental examples: Before undergoing radiotherapy, the shape of the bladder needs to be fixed. This is often done by having the patient drink 500mL of water, and then measuring the bladder volume using ultrasound after 10 minutes. After that, a CT scan is performed to obtain the location of the tumor. Then, the bladder volume is measured again before radiotherapy. If the volumes are similar, the bladder shapes will also be similar, and the radiotherapy to the target area will be more accurate.

[0083] Therefore, the validation data for this invention consisted of 13 patients who were about to undergo pelvic radiotherapy. Before CT localization, bladder ultrasound images were collected using a bladder volume analyzer (Chongqing Kangchao Medical Technology Co., Ltd.) according to standard operating procedures (such as fixed time after drinking water). CT scans were then performed immediately afterward.

[0084] Comparison Method: To verify the improvement in volume measurement accuracy of this invention (especially the deformation preprocessing step), the following comparative experiment was conducted. Using CT volume as the gold standard, the method of this invention (denoted as HQIA) was compared with existing built-in algorithms in volumetric instruments (denoted as the volumetric instrument) and a low-quality algorithm trained with unprocessed deformation preprocessing data (denoted as LQIA). The experimental results are shown in Table 1 below: Data processing: The CT volume is automatically delineated and calculated by the CT scanning system, and is used as the gold standard V. ct The other three methods calculate the volume, denoted as V. other .

[0085] Evaluation metric: Relative error percentage = |V other -V ct | / V ct ×100%.

[0086] The test results are shown in Table 1 below. The relative error percentage of the method of the present invention (My-US) is significantly lower than that of the traditional ultrasound method, and the error range is more concentrated.

[0087] Table 1. Bladder volume values ​​measured by different methods (unit: mL) Serial Number CT Volumetric meter Error value 1 HQIA Error value 2 LQIA Error value 3 1 375.59 441 -65.41 363.11 12.48 366.35 9.24268 2 487.29 509 -21.71 467.31 19.98 437.17 50.12496 3 395.49 488 -92.51 409.156 -13.666 406.14 -10.6483 4 403.4 436 -32.60 429.396 -25.996 425.98 -22.5776 5 239.421 263 -23.58 231.728 7.693 175.21 64.21164 6 472.291 499 -26.71 445.659 26.632 460.41 11.88268 7 373.858 431 -57.14 361.193 12.665 351.75 22.10534 8 364.318 388 -23.68 373.745 -9.427 339.93 24.39074 9 415.801 423 -7.20 432.610 -16.809 413.76 2.04132 10 381.09 481 -99.91 352.735 28.355 407.20 -26.11448 11 168.083 235 -66.92 201.14 -33.057 103.11 64.9714 12 124.866 161 -36.13 143.963 -19.097 116.91 7.95134 13 381.985 410 -28.02 375.45 6.535 413.97 -31.98668 Table 2 index Volumetric meter LQIA HQIA Average Error Rate (MAPE) 14.97% 9.52% 6.18% Median of error rate 9.85% 6.37% 4.04% Maximum error rate 39.81% 38.65% 19.67% Minimum error rate 1.73% 0.49% 1.71% Standard deviation (degree of volatility) 11.52% 10.92% 5.33% Experimental results show that the method of this invention (HQIA) is significantly superior to existing technologies in terms of average error, error stability (standard deviation), and maximum error control. In particular, compared with LQIA without deformation preprocessing, the average error rate of this invention is further reduced, and the maximum error rate is effectively controlled, demonstrating that PCA-guided non-rigid deformation registration preprocessing plays a crucial role in improving the measurement accuracy of the model in real clinical scenarios (where probe pressure deformation exists).

[0088] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A novel method for measuring bladder volume using ultrasound, characterized in that, Includes the following steps: S1. Construct a CT-like ultrasound section dataset: acquire bladder CT three-dimensional images and generate high-precision dense point cloud data; perform non-rigid deformation registration on the dense point cloud data to simulate bladder deformation caused by ultrasound probe pressure and generate deformation point cloud data. Meridian datasets for simulated ultrasonic scanning are extracted from deformed point cloud data. These datasets include incomplete and complete meridian datasets, and incomplete-dense point cloud data pairs are established. S2. Training a multi-path point cloud completion model: A deep learning network is used to train the incomplete-dense point cloud data pair to obtain the mapping weights from the incomplete point cloud to the complete point cloud. S3. Ultrasonic Data Quality Grading and Preprocessing: Acquire actual ultrasound scan data and classify them into high-quality image sets and low-quality image sets according to image quality; for high-quality image set data, perform contour registration directly to generate partial point clouds; for low-quality image set data, perform 2D boundary completion and stabilized 3D assembly to generate partial point clouds. S4. Ultrasonic data 3D reconstruction: The generated partial point cloud is input into the trained multi-path point cloud completion model to complete it into a 3D dense point cloud; S5. Bladder volume calculation: The Crust algorithm is used to perform surface reconstruction and volume integration on the completed 3D dense point cloud to obtain the estimated bladder volume.

2. The method according to claim 1, characterized in that, In step S1, the non-rigid deformation registration includes: constructing a geometric model of the bladder using three-dimensional modeling software, and using the central axis of the bladder as a reference, performing multi-angle deformation amplification of the model along the meridian direction to simulate probe pressure deformation in different directions and to varying degrees.

3. The method according to claim 2, characterized in that, The multi-angle deformation amplification process includes: multi-angle data amplification of the CT point cloud from a central axis, generating a total of 360 meridians, performing deformation processing on each meridian, and finally obtaining multiple training data containing no deformation and multiple deformation states.

4. The method according to claim 1, characterized in that, In step S1, the generation of the complete meridian dataset includes: Determine the central axis of the bladder, which includes the coordinate axes in the three-dimensional coordinate system and the diagonal direction of the cube; For each central axis, an initial normal vector is generated, and the initial normal vector is rotated around the central axis. A normal vector is obtained every 15° of rotation, for a total of 12 normal vectors. For each normal vector, determine a cross-sectional plane that passes through the center of the bladder and is perpendicular to the normal vector. Extract the set of points that intersect with the cross-sectional plane from the dense point cloud data to form a complete meridian. The generation of incomplete meridian datasets includes: collecting incomplete contour samples from actual ultrasound equipment, sorting and removing regions from the contour points, simulating different missing patterns, and mapping the two-dimensional incomplete contours to three-dimensional space to form incomplete meridian data.

5. The method according to claim 4, characterized in that, During the generation of the incomplete meridian dataset, the contour points in the ultrasound image are sorted counterclockwise. Based on the top point, the point set is batch-removed according to the preset incomplete interval to simulate the contour loss caused by probe obstruction, tissue interference or equipment limitations in clinical practice.

6. The method according to claim 1, characterized in that, In step S4, the multi-path point cloud completion model includes at least one of the following completion paths: Path 1: For high-quality image datasets, input the relatively complete data of the 12 meridians into the deep learning model and directly output dense point cloud data; Path 2: For low-quality image set data, the incomplete 12 meridian data are first repaired into complete 12 meridian data through image segmentation or traditional geometric algorithms, and then input into the deep learning model to output dense point cloud data.

7. The method according to claim 6, characterized in that, The point cloud completion model adopts an architecture based on Transformer or convolutional neural network and uses chamfer distance as the loss function for training.

8. The method according to claim 1, characterized in that, In step S5, the Crust algorithm includes: Delaunay triangulation is performed on dense point clouds. The concave and convex structures are identified by judging the positional relationship between the center of the circumcircle of adjacent triangles and the common chord. Triangular facets of concave regions are retained, invalid facets are removed, and a closed triangular mesh of the bladder surface is constructed. The total volume of the bladder can be calculated by constructing a tetrahedron from the surface triangular facets and the internal reference points and summing their volumes, or by directly summing the volumes of the internal tetrahedrons.

9. The method according to claim 1, characterized in that, It also includes using phantoms to calibrate ultrasound-CT data before model training to estimate systematic errors and perform background correction.

10. A system for performing the method according to any one of claims 1-9, characterized in that, include: The data acquisition module is used to acquire CT three-dimensional image data and ultrasound image data of the bladder; The point cloud generation module is used to generate dense point cloud data of the bladder based on CT 3D image data; The non-rigid deformation registration module is used to preprocess dense point cloud data to simulate probe pressure deformation. The meridian dataset generation module is used to extract a meridian dataset simulating ultrasound scanning from the dense point cloud data based on the principle of ultrasound imaging. The ultrasound data quality grading module is used to grade actual ultrasound images into high-quality / low-quality image sets and perform corresponding preprocessing to generate partial point clouds. The point cloud completion module includes a deep learning model used to complete partial point clouds into dense point cloud data. The 3D reconstruction and volume calculation module uses the Crust algorithm to reconstruct the 3D surface of the completed dense point cloud data and calculate the bladder volume.

Images