CBCT image compression and progressive loading method based on hierarchical neural field

Abstract

The application discloses a CBCT image compression and progressive loading method based on a layered neural field, and relates to the technical field of medical image data processing.The application divides an implicit neural representation network into multiple levels by constructing a cascaded layered neural field network model.The low-level subnetwork fits the overall profile and low-frequency information of an image, and the high-level subnetwork further fits the high-frequency texture and details of the image by using the features transmitted by the previous layer.Through multi-level joint supervision training and layered storage of the trained network parameters, the progressive transmission of image data is realized.In actual application, the low-layer parameters with extremely small capacity are preferentially loaded to realize the first-frame rendering in seconds, and then the high-layer parameters are loaded in the background in a streaming manner, so that the output image is gradually and smoothly transitioned from a fuzzy overview to a high-fidelity full resolution, thereby significantly improving the real-time interaction experience of remote medical cooperation and mobile image diagnosis while greatly reducing the data storage and bandwidth costs.

Description

Technical Field

[0001] This invention relates to the field of medical image data processing technology, and in particular to a method for CBCT image compression and progressive loading based on a layered neural field. Background Technology

[0002] Cone-beam computed tomography (CBCT) has been widely used in oral and maxillofacial surgery, orthopedics, and interventional radiology due to its advantages such as high spatial resolution and small device size. However, the amount of three-dimensional volumetric data generated by CBCT equipment is usually very large. The CBCT data of a single patient often contains hundreds of high-resolution slices, and the data volume can reach hundreds of megabytes (MB) or even gigabytes (GB).

[0003] This massive amount of data presents serious clinical problems and engineering challenges for medical informatics. First, with the accumulation of cases, the sheer volume of data significantly increases hospital data storage costs, leading to an exponential increase in storage pressure. Second, data transmission efficiency hinders remote medical collaboration across hospital campuses and regions. Doctors often have to wait for lengthy data download processes when remotely reviewing images or conducting consultations, making real-time collaborative communication difficult. Third, in mobile healthcare scenarios, such as when patients are visiting at home or doctors are viewing images on mobile devices (phones, tablets), limitations in mobile network bandwidth and the storage and computing power of mobile devices prevent the smooth loading and viewing of high-precision 3D images, severely impacting the continuity and convenience of diagnosis and treatment.

[0004] Existing solutions mainly suffer from the following shortcomings:

[0005] 1. Raw data transmission mode: Direct transmission of raw DICOM files. Due to the large size of these files, the first frame loading delay is extremely high (Time-to-First-Frame). Users need to wait for the data to be completely downloaded before they can start watching the video, which cannot meet the requirement of "second-level interactive access".

[0006] 2. Server-Side Rendering: This method places the rendering load on the server side, with the client only receiving video streams or images. While it reduces the client's computing power requirements, frequent 3D interactions (rotation, scaling, slicing) require extremely high network bandwidth and extremely low network latency. Furthermore, the client cannot cache data, making it unusable offline.

[0007] 3. Traditional compression algorithms, such as JPEG2000, can reduce file size, but still consume a lot of memory after decompression and do not support content-based intelligent progressive loading.

[0008] In recent years, coordinate-based implicit neural representations (INR) or neural fields have provided new opportunities for data compression. Neural fields utilize neural networks (such as multilayer perceptrons, MLPs) to directly map spatial coordinates to physical properties (such as density and color), achieving extremely high compression ratios (typically 10-100 times). However, existing standard neural field methods usually require the entire network model to be fully loaded before inference rendering can begin. This means that users still have to wait for the entire model to download before seeing any frames, and the first-frame latency problem remains unresolved. Summary of the Invention

[0009] To address the problems of large data volume, slow transmission, high first-frame loading delay, and difficulty in adapting to different network environments in existing CBCT image technologies, this invention provides a CBCT image compression and progressive loading method based on hierarchical neural fields. This method implicitly represents 3D CBCT images by designing a hierarchical neural network structure. Unlike traditional monolithic networks, the network parameters of this invention are divided into multiple cascaded stages / levels. Each level has the ability to independently output images while simultaneously passing feature information to the next level. During transmission, a small number of parameters from lower levels are transmitted first, allowing users to quickly obtain low-resolution preview images; subsequently, higher-level parameters are continuously transmitted in the background, gradually improving image quality and achieving a progressive "view from the bottom up" experience. This is specifically achieved through the following technologies.

[0010] This invention provides a method for CBCT image compression and progressive loading based on layered neural fields, comprising the following steps:

[0011] Acquire raw CBCT images and perform normalization preprocessing;

[0012] A hierarchical neural field network model is constructed, comprising several cascaded sub-network layers; each sub-network layer outputs the predicted image intensity and feature vector of the current sub-network layer; the lowest-level sub-network layer inputs the original spatial coordinates; and each other sub-network layer inputs the original spatial coordinates and the feature vector output by the previous sub-network layer.

[0013] The hierarchical neural field network model is trained using the preprocessed original CBCT images until the error between the output image of each sub-network layer and the original CBCT image is minimized.

[0014] Upon receiving an image output request, the trained hierarchical neural network model progressively loads network parameters and renders CBCT images in hierarchical order until the output CBCT images reach full resolution.

[0015] Furthermore, the method for normalizing the original CBCT image includes: normalizing the original spatial coordinates and the original image intensity.

[0016] Furthermore, the method for normalizing the original spatial coordinates and original image intensity is as follows:

[0017] Convert the DICOM sequence of the original CBCT images into a three-dimensional voxel matrix;

[0018] The original image intensity of the original CBCT image is normalized to the range of [0,1].

[0019] Generate a corresponding three-dimensional coordinate grid and normalize the original spatial coordinates to a cubic space of [-1,1].

[0020] Furthermore, the hierarchical neural field network model is based on a multilayer perceptron and is defined as follows: It includes several sub-network layers.

[0021] Furthermore, the hierarchical neural field network model is trained using the preprocessed original CBCT images until the error between the output image of each sub-network layer and the original CBCT image is minimized.

[0022] Randomly initialize the network parameters of each sub-network layer and input them into the hierarchical neural field network model; the network parameters include weight parameters and bias parameters.

[0023] After receiving the network parameters of the first sub-network layer, the first sub-network layer is used to perform inference based on the original spatial coordinates, and the predicted image intensity and feature vector of the current level are generated to obtain the CBCT image.

[0024] Continue to input the original spatial coordinates and the feature vector output by the previous sub-network layer into each subsequent sub-network layer, perform incremental inference, render and generate the predicted image intensity and feature vector of the current level, refresh the rendering to obtain a new CBCT image, until finally obtain a CBCT image with full resolution.

[0025] Set and calculate the loss function used to train the hierarchical neural field network model, and select a training optimizer to iteratively train until the peak signal-to-noise ratio (PSNR) converges (i.e., simultaneously constrain the predicted image intensities I1, I2, and I3 to be as close as possible to I). GT ).

[0026] Furthermore, the loss function L used to train the hierarchical neural field network model Total The formula is:

[0027] ;

[0028] Among them, I k I represents the predicted image intensity of the sub-network layer at layer k. GT λ represents the original image intensity of the original CBCT image. k is the weight coefficient of the sub-network layer described in layer k, and Loss is the reconstruction loss function.

[0029] Furthermore, the reconstruction loss function Loss is a combination of one or more of the following: mean squared error, mean absolute error, and structural similarity loss.

[0030] Furthermore, after the hierarchical neural field network model has been trained, the network parameters of the trained hierarchical neural field network model are divided and stored according to the hierarchy.

[0031] Furthermore, the method for segmented storage is as follows:

[0032] Save the network parameters of each sub-network layer in the trained hierarchical neural field network model as their own independent network parameter data packets;

[0033] The weight parameters, originally represented by 32-bit floating-point numbers, are converted to 16-bit floating-point numbers or 8-bit integers, thereby quantizing the network parameter data packets of each sub-network layer.

[0034] Whether to segment and store the network parameters of the trained hierarchical neural network model depends on the data transmission format. For example:

[0035] (1) If data is transmitted in multiple requests using the HTTP protocol, and all requests are stateless, storing the weights as a whole would require the server to read the entire weight and then address it for each request, increasing the server load. Therefore, it is necessary to store the weights in segments.

[0036] (2) If the transmission is done via HTTP or using WebSocket or other stateful protocols, the storage can be stored without splitting.

[0037] It should be noted that the specific transmission protocol or whether storage is segmented does not affect the overall method.

[0038] Furthermore, the quantized network parameter data packets are compressed using a lossless compression method to generate the final compressed weight file;

[0039] Generate a metadata description file that matches the compressed weight file; the metadata description file includes network structure configuration, normalization parameters, image spatial parameters, and data verification information.

[0040] Specifically, the network structure configuration may include the number of layers in each sub-network layer, the number of neurons in each sub-network layer, and the activation function type; the normalization parameters include the normalization parameters after normalization preprocessing of the original CBCT images and the normalization range of the original spatial coordinates; the image spatial parameters include the image size and voxel resolution of the original CBCT images, and the data verification information includes the hash value of each network parameter data packet.

[0041] Furthermore, upon receiving an image output request, the trained hierarchical neural field network model progressively loads network parameters and renders CBCT images according to hierarchical order until the output CBCT images reach full resolution. The method is as follows:

[0042] Upon receiving a CBCT image generation request, the first-level sub-network layer is instantiated based on the metadata of the CBCT image and the network parameter data packet of the first-level sub-network layer. The predicted image intensity of the first level is calculated, and the corresponding image is displayed.

[0043] Download the network parameter data packets of each of the other sub-network layers in sequence, calculate the predicted image intensity of the current sub-network layer, and display the corresponding image;

[0044] Download the network parameter data packet of the highest-level sub-network layer, calculate the predicted image intensity of the current-level sub-network layer, and display the corresponding final high-definition image;

[0045] The present invention also provides a CBCT image compression and progressive loading system for performing the above-described CBCT image compression and progressive loading method; the CBCT image compression and progressive loading system includes:

[0046] The data preprocessing module is used to preprocess the original CBCT images;

[0047] The model building and training module is used to build and train the hierarchical neural field network model, and to divide and store the network parameters of the trained hierarchical neural field network model according to the hierarchy.

[0048] Data storage module, used to store network parameters of the hierarchical neural field network model;

[0049] The rendering output module is used to reconstruct and render CBCT images in real time based on the network parameters of the currently received hierarchical neural field network model.

[0050] The present invention also provides a computer-readable storage medium storing a computer program that performs the above-described CBCT image compression and progressive loading method.

[0051] The present invention also provides an electronic device, including a memory and at least one processor, wherein the memory stores at least one program, and the program executes the above-described CBCT image compression and progressive loading method through the processor.

[0052] Compared with the prior art, the advantages of the present invention are:

[0053] 1. Extremely low first-frame latency: By prioritizing the loading of a very small amount of underlying network data, users can see an overview of the image instantly after clicking, without having to wait for the complete data to download, which significantly improves the user experience.

[0054] 2. High compression ratio: By utilizing the implicit representation capability of neural fields, the data storage space and transmission bandwidth usage are significantly reduced.

[0055] 3. Adaptive Network Environment: The progressive loading mechanism naturally adapts to fluctuating network environments. In weak network environments, users can at least see the basic structure for initial diagnosis, while in strong network environments, they can quickly obtain high-definition details.

[0056] 4. Supports real-time rendering on the client side: The decompression process is the network inference process, which supports real-time reconstruction and volume rendering of arbitrary facets on client devices (including high-performance mobile phones and tablets) using GPUs, without continuously relying on server computing power. Attached Figure Description

[0057] Figure 1 This is a flowchart of a compression and loading method based on a hierarchical neural field according to the present invention.

[0058] Figure 2 This invention demonstrates the data interaction process between the client and the server.

[0059] Figure 3 This invention presents different levels of output effects. Target represents the original image, and Interm1-3 represent the visualization effects of the three-layer output. Detailed Implementation

[0060] The technical solution of the present invention will be clearly and completely described below. 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.

[0061] Example

[0062] The CBCT image compression and progressive loading method based on layered neural fields provided in this embodiment, taking clinical oral CBCT image data as an example, has the following specific steps:

[0063] Step 1: Acquire raw CBCT images and perform preprocessing.

[0064] CBCT image data of clinical oral cavity were selected, and the original format was DICOM sequence; the DICOM sequence was converted into a three-dimensional voxel matrix (VoxelGrid).

[0065] To adapt to the characteristics of the activation function in neural networks, the original image intensity (Hounsfield Unit, HU) is normalized to... scope.

[0066] Generate the corresponding three-dimensional coordinate grid and normalize the original spatial coordinates (x,y,z) to a cube space of [-1,1].

[0067] Step 2: Construct a hierarchical neural field network model

[0068] like Figure 2 As shown, the hierarchical neural field network model in this embodiment uses a multilayer perceptron (MLP) as the basic building block, defined as follows: A hierarchical neural field network model with 3 levels and a total of 14 sub-network layers was constructed.

[0069] In this system, the relatively low-level sub-network layers (such as the first level) have very few parameters and are used to capture the low-frequency profile of the image. As the sub-network layers become higher, the number of parameters gradually increases, and they are used to capture the high-frequency details and textures of the image.

[0070] 1. Level 1 (BaseStage)

[0071] The first layer consists of a small MLP, for example, two layers with 128 neurons each. The input to this sub-network layer is the original spatial coordinates (x, y, z). The output of this layer is a basic predicted image intensity I1 and a low-level feature vector f1.

[0072] Due to the limited number of parameters, this layer primarily learns the general outline and low-frequency information of the image, such as the basic boundary between bone and soft tissue.

[0073] 2. Second level (MiddleStage)

[0074] The input to this sub-network layer is the original spatial coordinates (x, y, z) and the feature vector output from the previous layer (in this embodiment, the feature vector f1 output by the first-level sub-network layer). This layer contains a medium-sized MLP, for example, 4 layers, with 128 neurons per layer. The output is the predicted image intensity I2 and the feature vector f2.

[0075] Using the contextual information provided by the feature vector f1, this layer focuses on refining the structure.

[0076] 3. Level 3 (FineStage)

[0077] The input to this sub-network layer is the coordinates (x, y, z) and the feature vector output from the previous layer (in this embodiment, the feature vector f2 output by the second-level sub-network layer). This layer is the largest, for example, with 8 layers and 256 neurons per layer. The final output is a high-quality predicted image intensity I3.

[0078] This layer is used to restore high-frequency textures and details of minute lesions in the image.

[0079] The cascaded structure of the hierarchical neural field network model with the above three sub-network layers ensures information reuse and enables each layer to learn in a "residual" or "incremental" manner based on the previous layer.

[0080] Step 3: Training the hierarchical neural field network model

[0081] Using the original CBCT images preprocessed in step one, the hierarchical neural field network model is overfitted until the error between the output image (predicted image intensity I1, I2, I3) of each sub-network layer and the original CBCT image is minimized. The specific method is as follows:

[0082] Randomly initialize the network parameters of each sub-network layer and input them into the hierarchical neural field network model (see https: / / zhuanlan.zhihu.com / p / 630487545); the network parameters include weight parameters and bias parameters;

[0083] The training of the hierarchical neural field network model employs single-scene overfitting. A separate network model is trained for each case's CBCT data.

[0084] The loss function is designed as a multi-stage joint loss, as shown in the following formula:

[0085] ;

[0086] In this embodiment, the following is selected: As a specific form of Loss, the above formula expands as follows:

[0087] ;

[0088] Among them, I GT These are actual voxel values. and Let λ represent the mean squared error (MSE), k = 0, 1, 2, 3, ... k These are the weighting coefficients of the loss function for each layer.

[0089] This step employs a multi-level joint supervision strategy, simultaneously constraining the predicted image intensities I1, I2, and I3 to be close to I. GT This embodiment mandates that even the first-level sub-network layers should strive to output meaningful images rather than incomprehensible intermediate features.

[0090] Therefore, λ1 ≤ λ2 ≤ λ3 is typically set to guide the network to focus on optimizing the final output quality. Specifically, λ1, λ2, and λ3 can all be set to 1, or they can be increased slightly in sequence. This ensures that even if only some sub-network layers are loaded, the hierarchical neural field network model can still output meaningful images, rather than noise.

[0091] The training optimizer can be Adam, with a learning rate of 3×10⁻⁶. -4 Iterate through the training until the PSNR (Peak Signal-to-Noise Ratio) metric converges.

[0092] Step 4: Segment and store the network parameters of the trained hierarchical neural field network model.

[0093] After training, the network parameters of the hierarchical neural field network model are extracted, segmented, and encapsulated to support the progressive loading described in step five. The specific implementation method is as follows:

[0094] 1. Weight extraction and hierarchical independence

[0095] Based on the hierarchical neural field network model constructed in step two, the weight parameters and bias parameters after training convergence are physically segmented according to the sub-network layers.

[0096] Specifically, the network parameters of the first layer (BaseStage) sub-network layer are saved as an independent network parameter data packet P1, the network parameters of the second layer (MiddleStage) sub-network layer are saved as an independent network parameter data packet P2, and the network parameters of the third layer (FineStage) sub-network layer are saved as an independent network parameter data packet P3.

[0097] This physical segmentation ensures the independence of network parameter data packets for each subnet layer in terms of storage and transmission. The client can instantiate and run the first-layer subnet layer by simply obtaining the network parameter data packet P1, without relying on the network parameters of subsequent layers, thus achieving fast first-frame rendering under low bandwidth.

[0098] 2. Weight Quantization and Compression

[0099] To further reduce storage size and transmission latency, network parameter data packets at each sub-network layer are quantized.

[0100] The weight parameters, originally represented by 32-bit floating-point numbers (Float 32), are converted to 16-bit floating-point numbers (Float 16). The resulting network parameter data packets are then compressed using lossless compression algorithms (such as gzip or LZ4) to generate the final compressed weight file.

[0101] 3. Encapsulation of metadata and configuration information

[0102] Generate a metadata file to accompany the compressed weight file. The metadata file can be in JSON or Protocol Buffers format. The metadata file can be used to restore the image spatial size, remap grayscale ranges, etc.

[0103] The metadata description file contains the following key information:

[0104] (1) Network structure configuration

[0105] Record the number of layers, the number of neurons per layer, and the type of activation function for each sub-network layer so that the client can dynamically construct the computation graph.

[0106] (2) Normalized parameters

[0107] Record the normalization parameters (such as Max / Min of HU value) and the normalization range of the original spatial coordinates used in step one; this is used by the client to denormalize the output value to restore the true HU value after inference.

[0108] (3) Image spatial parameters

[0109] Record the image size (e.g., 512×512×512) and voxel resolution (e.g., 0.3 mm×0.3 mm×0.3 mm) of the original CBCT images; this is used by the client to map normalized coordinates back to the real physical space and to correctly calculate the sampling step size and field of view during rendering;

[0110] (4) Data verification information

[0111] It contains the hash value of each network parameter data packet; used for transmission integrity verification.

[0112] Step 5: Progressive Loading and Rendering Process

[0113] like Figure 1 and Figure 3 As shown, when requesting the generation of CBCT images, network parameters are progressively transmitted and loaded in hierarchical order. Lower-level network parameters are loaded first for low-resolution initial frame rendering, followed by higher-level network parameters to gradually improve image reconstruction quality until full resolution is achieved. The specific process is as follows:

[0114] 1. Handshake and metadata data transmission

[0115] The client initiates a CBCT image generation request, and the server returns the image's metadata (i.e., "metadata description file", including network structure configuration, normalization parameters, image spatial parameters, data verification information, etc.) and the network parameter data packet P1 of the first sub-network layer (e.g., only 50 kB).

[0116] It should be noted that the metadata to be returned in this application is not limited to a specific type, and not every type is mandatory to be returned; rather, it depends on the front-end and back-end protocols. For example:

[0117] If the network structure is fixed and all images use the same one, then there is no need to return the transmission network parameters;

[0118] If data integrity verification is not performed, there is no need to return a hash value;

[0119] If the image size is fixed, the image spatial parameters do not need to be returned;

[0120] If other patient information is needed, such as name, date of imaging, etc., additional metadata can be added.

[0121] 2. Fast rendering of the first frame

[0122] After receiving the network parameter data packet P1 from the first-level sub-network layer, the client immediately instantiates the first-level sub-network layer. At this time, the user rotates the view, and the client samples the coordinates on the view plane in real time. The first-level sub-network layer is then input, the model undergoes forward propagation, the predicted image intensity I1 is calculated, and the image is displayed. The specific calculation process is as follows:

[0123] ;

[0124] This process requires only millisecond-level latency under a normal 4G network, achieving first-frame loading within seconds. At this point, the image is a rough preview, relatively smooth and blurry, but the skeletal structure is visible.

[0125] 3. Background transmission and seamless upgrades

[0126] While the user is browsing, the client silently downloads the network parameter data packet P2 (e.g., 500 kB) of the second-level sub-network layer in the background.

[0127] Switching the rendering pipeline to "Layer 1 + Layer 2" mode will instantly improve image clarity for users. The specific calculation process is as follows:

[0128] ;

[0129] 4. Full quality ready

[0130] Finally, the network parameter data packet P3 (e.g., 10 MB) for the third-level sub-network layer is downloaded. After loading, the rendering pipeline switches to full-function mode, calculates the predicted image intensity I3, and displays high-resolution images. This ultimately yields high-resolution images that are almost indistinguishable from the original CBCT images in a DICOM sequence, thus meeting clinical diagnostic needs. The specific calculation process is as follows:

[0131] .

[0132] In addition to the above embodiments, the sub-network layers in the image rendering process provided by the present invention are not limited to three layers. Therefore, apart from the first-level sub-network layer, the network parameter data packets of the k-th level sub-network layer can be downloaded subsequently, and the predicted image intensity I of the current level sub-network layer can be calculated according to the following formula. k It then displays the corresponding final high-definition image.

[0133] .

[0134] Through the above methods, the present invention successfully achieves a balance between compression rate and real-time performance, and is particularly suitable for remote medical consultation and mobile image access scenarios.

[0135] The above detailed embodiments describe the implementation of the present invention; however, the present invention is not limited to the specific details described in the above embodiments. Within the scope of the claims and technical concept of the present invention, various simple modifications and changes can be made to the technical solution of the present invention, and these simple modifications all fall within the protection scope of the present invention.

Claims

1. A method for CBCT image compression and progressive loading based on layered neural fields, characterized in that, Includes the following steps: Acquire raw CBCT images and perform normalization preprocessing; A hierarchical neural field network model is constructed, comprising several cascaded sub-network layers; each sub-network layer outputs the predicted image intensity and feature vector of the current sub-network layer; the lowest-level sub-network layer inputs the original spatial coordinates; and each other sub-network layer inputs the original spatial coordinates and the feature vector output by the previous sub-network layer. The hierarchical neural field network model is trained using the preprocessed original CBCT images until the error between the output image of each sub-network layer and the original CBCT image is minimized. Upon receiving an image output request, the trained hierarchical neural network model progressively loads network parameters and renders CBCT images in hierarchical order until the output CBCT images reach full resolution.

2. The CBCT image compression and progressive loading method based on layered neural fields according to claim 1, characterized in that, The method for normalizing the original CBCT images includes: normalizing the original spatial coordinates and the original image intensity; Furthermore, the method for normalizing the original spatial coordinates and original image intensity is as follows: Convert the DICOM sequence of the original CBCT images into a three-dimensional voxel matrix; The original image intensity of the original CBCT image is normalized to the range of [0,1]. Generate a corresponding three-dimensional coordinate grid and normalize the original spatial coordinates to a cubic space of [-1,1].

3. The CBCT image compression and progressive loading method based on layered neural fields according to claim 1, characterized in that, The hierarchical neural field network model is based on a multilayer perceptron and is defined as follows: It includes several sub-network layers.

4. The CBCT image compression and progressive loading method based on layered neural fields according to claim 1, characterized in that, The method for training the hierarchical neural field network model using the preprocessed original CBCT images until the error between the output image of each sub-network layer and the original CBCT image is minimized is as follows: Randomly initialize the network parameters of each sub-network layer and input them into the hierarchical neural field network model; the network parameters include weight parameters and bias parameters. After receiving the network parameters of the first sub-network layer, the first sub-network layer is used to perform inference based on the original spatial coordinates, and the predicted image intensity and feature vector of the current level are generated to obtain the CBCT image. Continue to input the original spatial coordinates and the feature vector output by the previous sub-network layer into each subsequent sub-network layer, perform incremental inference, render and generate the predicted image intensity and feature vector of the current level, refresh the rendering to obtain a new CBCT image, until finally obtain a CBCT image with full resolution. Set and calculate the loss function for training the hierarchical neural field network model, and use a training optimizer to iterate the training until the peak signal-to-noise ratio converges.

5. The CBCT image compression and progressive loading method based on layered neural fields according to claim 4, characterized in that, Loss function L used for training hierarchical neural field network models Total The formula is: ； Among them, I k I represents the predicted image intensity of the sub-network layer at layer k. GT λ represents the original image intensity of the original CBCT image. k The weights of the sub-network layer at layer k are the weight coefficients, and Loss is the reconstruction loss function. Furthermore, the reconstruction loss function Loss is a combination of one or more of the following: mean squared error, mean absolute error, and structural similarity loss.

6. The CBCT image compression and progressive loading method based on layered neural fields according to claim 1, characterized in that, After the hierarchical neural field network model is trained, the network parameters of the trained hierarchical neural field network model are also divided and stored according to the hierarchy. Furthermore, the method for segmented storage is as follows: Save the network parameters of each sub-network layer in the trained hierarchical neural field network model as their own independent network parameter data packets; The weight parameters, originally represented by 32-bit floating-point numbers, are converted to 16-bit floating-point numbers or 8-bit integers, thereby quantizing the network parameter data packets of each sub-network layer. Furthermore, the quantized network parameter data packets are compressed using a lossless compression method to generate a compressed weight file; Generate a metadata description file that matches the compressed weight file; the metadata description file includes network structure configuration, normalization parameters, image spatial parameters, and data verification information; Furthermore, the network structure configuration includes the number of layers in each sub-network layer, the number of neurons in each sub-network layer, and the activation function type; the normalization parameters include the normalization parameters after normalization preprocessing of the original CBCT images and the normalization range of the original spatial coordinates; the image spatial parameters include the image size and voxel resolution of the original CBCT images, and the data verification information includes the hash value of each network parameter data packet.

7. The CBCT image compression and progressive loading method based on layered neural fields according to claim 1, characterized in that, Upon receiving an image output request, the trained hierarchical neural network model progressively loads network parameters and renders CBCT images according to hierarchical order until the output CBCT images reach full resolution. The method is as follows: Upon receiving a CBCT image generation request, the first-level sub-network layer is instantiated based on the metadata of the CBCT image and the network parameter data packet of the first-level sub-network layer. The predicted image intensity of the first-level sub-network layer is calculated, and the corresponding image is displayed. Download the network parameter data packets of each of the other sub-network layers in sequence, calculate the predicted image intensity of the current sub-network layer, and display the corresponding image; Download the network parameter data packet of the highest-level sub-network layer, calculate the predicted image intensity I1 of the current-level sub-network layer, and display the corresponding final high-definition image.

8. A CBCT image compression and progressive loading system, characterized in that, Used to perform the CBCT image compression and progressive loading method according to any one of claims 1-7; The CBCT image compression and progressive loading system includes: The data preprocessing module is used to preprocess the original CBCT images; The model building and training module is used to build and train the hierarchical neural field network model, and to divide and store the network parameters of the trained hierarchical neural field network model according to the hierarchy. Data storage module, used to store network parameters of the hierarchical neural field network model; The rendering output module is used to reconstruct and render CBCT images in real time based on the network parameters of the currently received hierarchical neural field network model.

9. A computer-readable storage medium, characterized in that, The computer program stores the method for performing CBCT image compression and progressive loading as described in any one of claims 1-7.

10. An electronic device, characterized in that, It includes a memory and at least one processor, the memory storing at least one program, the program executing the CBCT image compression and progressive loading method according to any one of claims 1-7 through the processor.

Images