4DCBCT image reconstruction method and device based on diaphragm segmentation, equipment and storage medium

Abstract

The invention discloses a diaphragm segmentation-based 4DCBCT image reconstruction method, apparatus and device, and a storage medium. The method comprises the steps of collecting a time sequence projection image sequence of a patient; inputting each frame of projection image into a pre-trained high-throughput diaphragm segmentation neural network model, predicting a diaphragm region in the projection image, and extracting a diaphragm top position coordinate; arranging the diaphragm top position coordinates corresponding to all the projection images according to a time sequence to generate a diaphragm height curve; determining breathing phases based on the diaphragm height curve, and constructing a phase projection set of each breathing phase; and performing three-dimensional reconstruction on the phase projection set of each breathing phase to obtain a three-dimensional image corresponding to each breathing phase, and generating a 4D-CBCT image set. According to the method, external equipment is not needed, the respiration signals are extracted completely based on the image content, and accurate and efficient 4D-CBCT reconstruction is achieved by constructing a high-throughput diaphragm segmentation neural network model and an image-driven phase division mechanism.

Description

Technical Field

[0001] This invention belongs to the field of CT image processing technology, specifically relating to a 4DCBCT image reconstruction method, apparatus, device, and storage medium based on diaphragm segmentation. Background Technology

[0002] 4D cone-beam computed tomography (4D-CBCT) technology is widely used in lung cancer and adaptive radiotherapy, with its key being the acquisition of CBCT images at different respiratory phases. However, traditional 4D-CBCT imaging procedures rely on additional respiratory monitoring equipment, such as pressure bands, infrared markers, and laser masks. These devices are used to simultaneously record respiratory signals and divide the acquired projected images into different respiratory phases. However, this additional hardware increases clinical costs and operational complexity, and is prone to problems such as signal drift and timing mismatch, affecting the accuracy of phase division.

[0003] In addition, some studies have attempted to extract respiratory information from the images themselves, such as analyzing image brightness, structural changes, or edge contour changes of target areas. Existing technology CN101743568A generates time-series images from compressed projection images to identify respiratory cycles, forming 4D sequence images. Furthermore, Mask R-CNN is used for semantic segmentation of the diaphragm, and the position of the diaphragm apex (highest point) is extracted for respiratory phase determination.

[0004] However, such methods have obvious limitations:

[0005] Methods based on brightness, edge, or handcrafted features struggle to accurately extract respiratory signals, are sensitive to noise, and rely on handcrafted features or simple image processing algorithms, lacking robustness and adaptability.

[0006] In terms of automatic segmentation of the diaphragm, most current methods use traditional image processing techniques, such as gradient enhancement, edge detection, and morphological manipulation. These methods are difficult to adapt to the anatomical differences between different patients and the artifacts and occlusion problems in the projected images, and cannot meet the needs of high-throughput clinical practice. Summary of the Invention

[0007] To address the aforementioned technical problems, this invention proposes a 4DCBCT image reconstruction method, apparatus, device, and storage medium based on diaphragm segmentation.

[0008] To achieve the above objectives, the technical solution of the present invention is as follows:

[0009] In a first aspect, the present invention discloses a 4DCBCT image reconstruction method based on diaphragm segmentation, comprising:

[0010] Step S1: Acquire a sequence of time-series projection images of the patient during a 360° scan while the patient is at rest and breathing.

[0011] Step S2: Input each frame of the projected image into the pre-trained high-throughput diaphragm segmentation neural network model to predict the diaphragm region in the projected image and extract the coordinates of the diaphragm top position;

[0012] The high-throughput diaphragm segmentation neural network model includes an encoder and a decoder;

[0013] The encoder uses the convolutional or Transformer encoder structure in the open-source X-ray image basic model as the initial weights, and the encoder network weights are not updated during training. The decoder uses a lightweight upsampling module and jump connections to fuse multi-scale information, and outputs a single-channel probability map with the same size as the input projected image, representing the distribution of the diaphragm region.

[0014] Step S3: Arrange the coordinates of the diaphragm top position corresponding to all the projected images in chronological order to generate a diaphragm height curve, and use it as a respiratory signal;

[0015] Step S4: Determine the respiratory phase based on the diaphragm height curve, and filter the projection images corresponding to each respiratory phase to construct a phase projection set for each respiratory phase;

[0016] Step S5: Perform three-dimensional reconstruction on the phase projection set of each respiratory phase to obtain a three-dimensional image corresponding to each respiratory phase, and generate a 4D-CBCT image set.

[0017] Based on the above technical solution, the following improvements can be made:

[0018] As a preferred approach, the training process of the high-throughput diaphragm segmentation neural network model includes:

[0019] Step A: Use a set of X-ray projection images with annotated diaphragm regions as the training dataset;

[0020] Step B: Perform data augmentation on the training dataset. Data augmentation includes one or more of the following: random rotation, affine transformation, and contrast perturbation.

[0021] Step C: Model fine-tuning training;

[0022] Fixed network weights of the pre-trained encoder;

[0023] The decoder training is supervised by a composite loss function, which includes a weighted combination of Dice loss and binary cross-entropy, and the maximum diaphragm position error is used as an auxiliary optimization index.

[0024] As a preferred embodiment, step S4 includes:

[0025] Step S4.1: Smooth and interpolate the diaphragm height curve to determine the periodic respiratory waveform;

[0026] Step S4.2: Determine the respiratory cycle based on the periodic respiratory waveform, and divide each respiratory cycle into m respiratory phases;

[0027] Step S4.3: Select several projection images belonging to each respiratory phase to construct a phase projection set.

[0028] As a preferred option, step S4.3 filters the projected image corresponding to each breathing phase using a time index or similarity matching method.

[0029] Secondly, the present invention discloses a 4DCBCT image reconstruction device based on diaphragm segmentation, comprising:

[0030] The image acquisition module is used to acquire a sequence of time-projected images of a patient in a 360° scan while at rest and breathing.

[0031] The coordinate prediction module is used to input each frame of the projected image into a pre-trained high-throughput diaphragm segmentation neural network model, predict the diaphragm region in the projected image, and extract the coordinates of the diaphragm top position.

[0032] The high-throughput diaphragm segmentation neural network model includes an encoder and a decoder;

[0033] The encoder uses the convolutional or Transformer encoder structure in the open-source X-ray image basic model as the initial weights, and the encoder network weights are not updated during training. The decoder uses a lightweight upsampling module and jump connections to fuse multi-scale information, and outputs a single-channel probability map with the same size as the input projected image, representing the distribution of the diaphragm region.

[0034] The curve generation module is used to arrange the coordinates of the diaphragm top position corresponding to all projected images in chronological order, generate a diaphragm height curve, and use it as a respiratory signal.

[0035] The projection set construction module is used to determine the respiratory phase based on the diaphragm height curve, and to filter the projection images corresponding to each respiratory phase to construct the phase projection set for each respiratory phase.

[0036] The reconstruction module is used to perform three-dimensional reconstruction of the phase projection set for each respiratory phase, obtain the three-dimensional image corresponding to each respiratory phase, and generate a 4D-CBCT image set.

[0037] As a preferred embodiment, the 4DCBCT image reconstruction device further includes: a training module for training a high-throughput diaphragm segmentation neural network model, including:

[0038] The training set construction unit is used to employ a set of X-ray projection images of the annotated diaphragm region as the training dataset.

[0039] The data augmentation unit is used to perform data augmentation processing on the training dataset. Data augmentation includes one or more of random rotation, affine transformation, and contrast perturbation.

[0040] The model training unit is used for fine-tuning and training the model.

[0041] Fixed network weights of the pre-trained encoder;

[0042] The decoder training is supervised by a composite loss function, which includes a weighted combination of Dice loss and binary cross-entropy, and the maximum diaphragm position error is used as an auxiliary optimization index.

[0043] As a preferred solution, the projection set construction module includes:

[0044] The smoothing interpolation unit is used to smooth and interpolate the diaphragm height curve to determine the periodic respiratory waveform;

[0045] The phase division unit is used to determine the respiratory cycle based on the periodic respiratory waveform and divide each respiratory cycle into m respiratory phases;

[0046] The filtering unit is used to filter several projection images belonging to each respiratory phase to construct a phase projection set.

[0047] As a preferred approach, the filtering unit filters the projected images corresponding to each respiratory phase using a time index or similarity matching method.

[0048] Thirdly, the present invention discloses a computing device, comprising:

[0049] One or more processors;

[0050] Memory;

[0051] And one or more programs, wherein the one or more programs are stored in memory and configured to be executed by one or more processors, and the one or more programs include any of the above-described diaphragm segmentation-based 4DCBCT image reconstruction methods.

[0052] Fourthly, the present invention discloses a storage medium storing one or more computer-readable programs, the one or more programs including instructions adapted to be loaded by a memory and executed any of the above-described diaphragm-based 4DCBCT image reconstruction methods.

[0053] This invention discloses a 4D-CBCT image reconstruction method, device, equipment, and storage medium based on diaphragm segmentation. It requires no external equipment and extracts respiratory signals entirely from the image's own content. By constructing a high-throughput diaphragm segmentation neural network model and an image-driven phase division mechanism, it achieves accurate and efficient 4D-CBCT reconstruction, significantly improving image quality and clinical operability. Its beneficial effects include:

[0054] First, this invention eliminates the need for external equipment, reduces costs, does not rely on external respiratory monitoring equipment, and reduces clinical intervention.

[0055] Second, this invention utilizes the structural features of the image itself (diaphragm movement) to generate a diaphragm height curve, and uses it as a respiratory signal to achieve highly robust respiratory signal extraction.

[0056] Third, the high-throughput diaphragm segmentation neural network model of this invention adopts transfer learning and a lightweight network structure, which can quickly process thousands of projection images and has high-throughput automatic processing capabilities.

[0057] Fourth, this invention automatically identifies the diaphragm contour using deep learning methods, adapting to different body types and pathological features, and achieving high diaphragm positioning accuracy.

[0058] Fifth, this invention can be integrated into existing CBCT systems, exhibiting good compatibility without requiring hardware modifications to the original scanning system. The invention boasts a stable process, simple operation, and adaptability to routine clinical scanning modes, facilitating large-scale application. Attached Figure Description

[0059] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0060] Figure 1 A flowchart of the 4DCBCT image reconstruction method provided in an embodiment of the present invention.

[0061] Figure 2 This is a schematic diagram of a high-throughput diaphragm segmentation neural network model provided in an embodiment of the present invention. Detailed Implementation

[0062] The preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0063] 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.

[0064] The expression “includes” is an “open-ended” expression, which means that there is a corresponding component or step, and should not be interpreted as excluding additional components or steps.

[0065] To achieve the objectives of this invention, some embodiments of the 4DCBCT image reconstruction method based on diaphragm segmentation, such as Figure 1 As shown, the 4DCBCT image reconstruction method includes:

[0066] Step S101: Acquire a sequence of time-series projection images of the patient during a 360° scan while the patient is at rest and breathing.

[0067] Step S102: Input each frame of the projected image into the pre-trained high-throughput diaphragm segmentation neural network model to predict the diaphragm region in the projected image and extract the coordinates of the diaphragm top position;

[0068] The high-throughput diaphragm segmentation neural network model includes an encoder and a decoder;

[0069] The encoder uses the convolutional or Transformer encoder structure in the open-source X-ray image basic model as the initial weights, and the encoder network weights are not updated during training. The decoder uses a lightweight upsampling module and jump connections to fuse multi-scale information, and outputs a single-channel probability map with the same size as the input projected image, representing the distribution of the diaphragm region.

[0070] Step S103: Arrange the coordinates of the diaphragm top position corresponding to all the projected images in chronological order to generate a diaphragm height curve and use it as a respiratory signal;

[0071] Step S104: Determine the respiratory phase based on the diaphragm height curve, and filter the projection images corresponding to each respiratory phase to construct a phase projection set for each respiratory phase;

[0072] Step S105: Perform three-dimensional reconstruction on the phase projection set of each respiratory phase to obtain a three-dimensional image corresponding to each respiratory phase, and generate a 4D-CBCT image set.

[0073] Each of the above steps will be explained in detail below.

[0074] Step S101 is the projection image acquisition step.

[0075] The patient undergoes a 360° circular orbit CBCT scan while at rest, resulting in several time-series X-ray projection images. This step can be performed using any X-ray system that supports CBCT acquisition.

[0076] Step S102 is a crucial step in extracting the location of the diaphragm.

[0077] Each projected image is input into the high-throughput diaphragm segmentation neural network model proposed in this invention, which automatically predicts the diaphragm region in the projected image and extracts the coordinates of the diaphragm top (i.e., the coordinates of the highest point of the diaphragm).

[0078] The high-throughput diaphragm segmentation neural network model is based on a lightweight convolutional network architecture and is optimized for projection image resolution and noise characteristics, enabling the localization and height prediction of the diaphragm region in milliseconds.

[0079] Compared to existing methods based on external devices (such as pressure bands and marker tracking), this method extracts physiological signals from the image itself without the need for additional hardware.

[0080] For diaphragm location extraction in projection images, a high-throughput diaphragm segmentation neural network model is characterized by its speed, stability, and high accuracy. This model is not trained from scratch; instead, it transfers the encoder structure of a large-scale X-ray pre-trained model, adds a lightweight decoder adapted to this task, and fine-tunes it for diaphragm segmentation.

[0081] like Figure 2 As shown, the high-throughput diaphragm segmentation neural network model includes an encoder and a decoder.

[0082] The encoder uses a convolutional or Transformer encoder structure disclosed in the open-source X-ray image base model as the initial encoder weights of the model; the encoder has the ability to pre-train on large-scale chest X-ray data and has good generalization ability to projected image texture features (such as diaphragm boundaries); the encoder network weights are not updated during training.

[0083] The decoder employs a lightweight upsampling module (such as a UNet-style decoder) and combines it with skip connections to fuse multi-scale information; the output is a single-channel probability map with the same size as the input projected image, representing the distribution of the diaphragm region.

[0084] This invention uses an X-ray pre-trained model as an encoder to adapt to the structure of projected images; it constructs a lightweight decoder to achieve efficient segmentation of the diaphragm region.

[0085] The training process of the high-throughput diaphragm segmentation neural network model includes:

[0086] Step A: Use a set of X-ray projection images (which can be manually annotated or synthetic images) labeled with the diaphragm region as the training dataset;

[0087] Based on the original resolution and pixel resolution of the input image in the training dataset, interpolate or downsample to a 512×512 projected image with a pixel resolution of 0.224mm*0.224mm.

[0088] Step B: Perform data augmentation on the training dataset. Data augmentation includes random rotation, affine transformation, contrast perturbation, etc., to improve generalization performance.

[0089] Step C: Model fine-tuning training;

[0090] Fixed network weights of the pre-trained encoder;

[0091] The decoder training is supervised by a composite loss function, which includes a weighted combination of Dice loss and binary cross entropy (BCE), and the maximum diaphragm position error is used as an auxiliary optimization index.

[0092] A weighted combination of Dice loss and binary cross-entropy is used to ensure boundary fineness and overall region overlap.

[0093] During the training phase, the "maximum diaphragm position error" was introduced as an auxiliary indicator to optimize positioning accuracy.

[0094] The single-frame prediction time of the high-throughput diaphragm segmentation neural network model can be controlled within 30 to 50 milliseconds, making it suitable for processing thousands of projection sequences generated from clinical medium-dose scans.

[0095] Step S103 arranges the diaphragm height values ​​corresponding to all projected images in chronological order to obtain a continuous diaphragm height curve, which serves as an image-driven respiratory signal to replace the traditional respiratory monitoring signal.

[0096] Step S104 is the phase division step.

[0097] Furthermore, step S104 includes:

[0098] Step S104.1: Smooth and interpolate the diaphragm height curve to determine the periodic respiratory waveform;

[0099] Step S104.2: Determine the respiratory cycle based on the periodic respiratory waveform, and divide each respiratory cycle into m respiratory phases (by m = 8 or 10);

[0100] Step S104.3: For each respiratory phase, select several projected images belonging to that respiratory phase by means of time indexing or similarity matching to construct a phase projection set.

[0101] Steps S103 and S104 achieve device-free phase registration, use the diaphragm height curve to replace the traditional respiratory monitoring signal, and automatically construct the respiratory cycle only through the image sequence, thus solving the problem of traditional signal drift or loss.

[0102] For each respiratory phase, its corresponding phase projection set is automatically determined, which can adapt to unstable respiratory rate conditions.

[0103] Step S105 is the 4DCBCT image reconstruction step based on phase projection.

[0104] CBCT reconstruction is performed on the phase projection set of each respiratory phase to obtain the three-dimensional image corresponding to each phase, and finally a 4D-CBCT dataset containing m time frames is formed.

[0105] It is worth noting that the reconstruction algorithm can be, but is not limited to, FDK, SART, OSEM, etc.

[0106] This invention discloses a 4DCBCT image reconstruction method based on diaphragm segmentation. Compared with the prior art, this invention introduces a large model transfer strategy that is more in line with the characteristics of CBCT, and discloses a complete image-driven workflow: segmentation → signal generation → phase division → reconstruction.

[0107] The present invention has the following effects:

[0108] First, it completely eliminates the need for external respiratory monitoring devices such as pressure bands and infrared markers that traditional 4D-CBCT relies on, requiring no additional hardware investment. This not only reduces the cost of clinical equipment procurement and maintenance but also simplifies the scanning process, reduces patient discomfort and operational complexity caused by wearing the device, and improves clinical work efficiency.

[0109] Secondly, by generating respiratory signals based on the physiological structural characteristics of diaphragmatic movement within the image itself, the problem of drift and loss caused by external device signals due to positional changes and electromagnetic interference is avoided. The diaphragmatic height curve is directly derived from the content of the projected image and is highly synchronized with organ movement, ensuring the accuracy and stability of respiratory phase segmentation, which is particularly suitable for imaging scenarios of patients with irregular breathing.

[0110] Third, by adopting a transfer learning strategy and a lightweight neural network architecture, the encoder of the large-scale X-ray pre-trained model is reused to quickly adapt to the characteristics of CBCT projection images; the lightweight decoder design achieves diaphragm segmentation and localization in 30-50 milliseconds per frame image, which can efficiently process thousands of projection sequences, meet the needs of high-throughput and automated clinical processing, and eliminate the need for manual intervention.

[0111] Fourth, leveraging the deep feature mining capabilities of deep learning, the model automatically identifies the diaphragm contour and extracts the diaphragm dome position, overcoming the sensitivity of traditional image processing methods to individual anatomical differences, projection artifacts, and occlusion. Through data augmentation and composite loss function optimization, the model can adapt to the diaphragm morphology of patients with different body types and pathological conditions (such as emphysema and pleural effusion), and its localization accuracy is significantly better than manual features or traditional algorithms.

[0112] Fifth, relying solely on the projection images acquired by the CBCT system itself, no hardware modifications to existing scanning equipment are required, allowing direct integration into existing CBCT systems. The complete image-driven workflow (segmentation-signal generation-phase division-reconstruction) is stable, easy to operate, and adaptable to routine clinical scanning modes, facilitating large-scale deployment and providing reliable 4D imaging support for scenarios such as adaptive radiotherapy for lung tumors.

[0113] In other embodiments, the present invention discloses a 4DCBCT image reconstruction device based on diaphragm segmentation, comprising:

[0114] The image acquisition module is used to acquire a sequence of time-projected images of a patient in a 360° scan while at rest and breathing.

[0115] The coordinate prediction module is used to input each frame of the projected image into a pre-trained high-throughput diaphragm segmentation neural network model, predict the diaphragm region in the projected image, and extract the coordinates of the diaphragm top position.

[0116] The high-throughput diaphragm segmentation neural network model includes an encoder and a decoder;

[0117] The encoder uses the convolutional or Transformer encoder structure in the open-source X-ray image basic model as the initial weights, and the encoder network weights are not updated during training. The decoder uses a lightweight upsampling module and jump connections to fuse multi-scale information, and outputs a single-channel probability map with the same size as the input projected image, representing the distribution of the diaphragm region.

[0118] The curve generation module is used to arrange the coordinates of the diaphragm top position corresponding to all projected images in chronological order, generate a diaphragm height curve, and use it as a respiratory signal.

[0119] The projection set construction module is used to determine the respiratory phase based on the diaphragm height curve, and to filter the projection images corresponding to each respiratory phase to construct the phase projection set for each respiratory phase.

[0120] The reconstruction module is used to perform three-dimensional reconstruction of the phase projection set for each respiratory phase, obtain the three-dimensional image corresponding to each respiratory phase, and generate a 4D-CBCT image set.

[0121] Furthermore, based on the above embodiments, the 4DCBCT image reconstruction device further includes: a training module for training a high-throughput diaphragm segmentation neural network model, including:

[0122] The training set construction unit is used to employ a set of X-ray projection images of the annotated diaphragm region as the training dataset.

[0123] The data augmentation unit is used to perform data augmentation processing on the training dataset. Data augmentation includes one or more of random rotation, affine transformation, and contrast perturbation.

[0124] The model training unit is used for fine-tuning and training the model.

[0125] Fixed network weights of the pre-trained encoder;

[0126] The decoder training is supervised by a composite loss function, which includes a weighted combination of Dice loss and binary cross-entropy, and the maximum diaphragm position error is used as an auxiliary optimization index.

[0127] Furthermore, based on the above embodiments, the projection set construction module includes:

[0128] The smoothing interpolation unit is used to smooth and interpolate the diaphragm height curve to determine the periodic respiratory waveform;

[0129] The phase division unit is used to determine the respiratory cycle based on the periodic respiratory waveform and divide each respiratory cycle into m respiratory phases;

[0130] The filtering unit is used to filter several projection images belonging to each respiratory phase to construct a phase projection set.

[0131] Furthermore, based on the above embodiments, the filtering unit filters the projected images corresponding to each respiratory phase using a time index or similarity matching method.

[0132] Furthermore, it should be noted that: when determining the dominant vertex, the 4DCBCT image reconstruction device based on diaphragm segmentation provided in the above embodiments is only illustrated by the division of the above functional modules. In actual applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the 4DCBCT image reconstruction device based on diaphragm segmentation can be divided into different functional modules to complete all or part of the functions described above.

[0133] Furthermore, the embodiments of the 4DCBCT image reconstruction device based on diaphragm segmentation and the 4DCBCT image reconstruction method based on diaphragm segmentation provided above belong to the same concept. The specific implementation process can be found in the method embodiments, and will not be repeated here.

[0134] In another embodiment, the present invention discloses a computing device, comprising:

[0135] One or more processors;

[0136] Memory;

[0137] And one or more programs, wherein the one or more programs are stored in memory and configured to be executed by one or more processors, the one or more programs including the 4DCBCT image reconstruction method based on diaphragm segmentation disclosed in any of the above embodiments.

[0138] In another embodiment, the present invention discloses a storage medium storing one or more computer-readable programs, the one or more programs including instructions adapted to be loaded by memory and executed the diaphragm segmentation-based 4DCBCT image reconstruction method disclosed in any of the above embodiments.

[0139] This invention discloses a 4D-CBCT image reconstruction method, device, equipment, and storage medium based on diaphragm segmentation. It requires no external equipment and extracts respiratory signals entirely from the image's own content. By constructing a high-throughput diaphragm segmentation neural network model and an image-driven phase division mechanism, it achieves accurate and efficient 4D-CBCT reconstruction, significantly improving image quality and clinical operability. Its beneficial effects include:

[0140] First, this invention eliminates the need for external equipment, reduces costs, does not rely on external respiratory monitoring equipment, and reduces clinical intervention.

[0141] Second, this invention utilizes the structural features of the image itself (diaphragm movement) to generate a diaphragm height curve, and uses it as a respiratory signal to achieve highly robust respiratory signal extraction.

[0142] Third, the high-throughput diaphragm segmentation neural network model of this invention adopts transfer learning and a lightweight network structure, which can quickly process thousands of projection images and has high-throughput automatic processing capabilities.

[0143] Fourth, this invention automatically identifies the diaphragm contour using deep learning methods, adapting to different body types and pathological features, and achieving high diaphragm positioning accuracy.

[0144] Fifth, this invention can be integrated into existing CBCT systems, exhibiting good compatibility without requiring hardware modifications to the original scanning system. The invention boasts a stable process, simple operation, and adaptability to routine clinical scanning modes, facilitating large-scale application.

[0145] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the present invention. Various changes and modifications can be made to the present invention without departing from its spirit and scope. All such changes and modifications fall within the scope of the present invention as claimed, which is defined by the appended claims and their equivalents.

Claims

1. A 4D CBCT image reconstruction method based on diaphragm segmentation, characterized in that, The method comprises the following steps: Step S1: collecting a time sequence projection image sequence of a patient in a resting breathing state in 360° scanning; Step S2: inputting each frame of projection image into a pre-trained high-throughput diaphragm segmentation neural network model, predicting the diaphragm region in the projection image, and extracting the diaphragm top position coordinates; The high-throughput diaphragm segmentation neural network model comprises an encoder and a decoder; The encoder adopts the convolution or Transformer encoder structure in the open source X-ray image base model as the initial weight, and the encoder network weight is not updated in the training process, and the decoder adopts a light-weight upsampling module and cooperates with a jump connection to fuse multi-scale information, and outputs a single-channel probability map consistent with the size of the input projection image, representing the distribution of the diaphragm region; Step S3: arranging the diaphragm top position coordinates corresponding to all projection images in time sequence to generate a diaphragm height curve, and taking the diaphragm height curve as a breathing signal; Step S4: determining the breathing phase based on the diaphragm height curve, and screening the projection images corresponding to each breathing phase to construct a phase projection set of each breathing phase; Step S5: performing three-dimensional reconstruction on the phase projection set of each breathing phase to obtain a three-dimensional image corresponding to each breathing phase, and generating a 4D-CBCT image set.

2. The 4D CBCT image reconstruction method of claim 1, wherein, The training process of the high-throughput diaphragm segmentation neural network model comprises the following steps: Step A: using an X-ray projection image set labeled with a diaphragm region as a training data set; Step B: performing data enhancement processing on the training data set, wherein the data enhancement comprises one or more of random rotation, affine transformation, and contrast disturbance; Step C: model fine-tuning training; Fixing the network weight of the pre-trained encoder; Using a composite loss function to supervise the decoder training, comprising: a weighted combination of Dice loss and binary cross entropy, and taking the maximum diaphragm top position error as an auxiliary optimization index.

3. The 4D CBCT image reconstruction method of claim 1, wherein, The step S4 comprises: Step S4.1: smoothing and interpolating the diaphragm height curve to determine a periodic breathing waveform; Step S4.2: determining a breathing period according to the periodic breathing waveform, and dividing each breathing period into m breathing phases; Step S4.3: screening a plurality of projection images belonging to each breathing phase to construct a phase projection set.

4. The 4D CBCT image reconstruction method of claim 3, wherein, The step S4.3 screens the projection images corresponding to each breathing phase by time index or similarity matching method.

5. Apparatus for 4D CBCT image reconstruction based on diaphragm segmentation, characterized in that, The method comprises the following steps: An image acquisition module for acquiring a time sequence projection image sequence of a patient in a resting breathing state in 360° scanning; A coordinate prediction module for inputting each frame of projection image into a pre-trained high-throughput diaphragm segmentation neural network model, predicting the diaphragm region in the projection image, and extracting the diaphragm top position coordinates; The high-throughput diaphragm segmentation neural network model comprises an encoder and a decoder; The encoder adopts the convolution or Transformer encoder structure in the open source X-ray image base model as the initial weight, and the encoder network weight is not updated in the training process, and the decoder adopts a light-weight upsampling module and cooperates with a jump connection to fuse multi-scale information, and outputs a single-channel probability map consistent with the size of the input projection image, representing the distribution of the diaphragm region; The curve generation module is configured to arrange diaphragm top position coordinates corresponding to all projection images in time sequence, generate a diaphragm height curve, and take the diaphragm height curve as a breathing signal; The projection set construction module is configured to determine a breathing phase based on the diaphragm height curve, and screen projection images corresponding to each breathing phase to construct a phase projection set of each breathing phase; The reconstruction module is configured to perform three-dimensional reconstruction on the phase projection set of each breathing phase to obtain a three-dimensional image corresponding to each breathing phase, and generate a 4D-CBCT image set.

6. The 4D CBCT image reconstruction apparatus of claim 5, wherein, The 4D CBCT image reconstruction device further comprises a training module configured to perform a training process on a high-throughput diaphragm segmentation neural network model, including: A training set construction unit configured to take an X-ray projection image set with labeled diaphragm regions as a training data set; A data enhancement unit configured to perform data enhancement processing on the training data set, the data enhancement including one or more of random rotation, affine transformation, and contrast disturbance; A model training unit configured to perform model fine-tuning training; Fixing network weights of a pre-trained encoder; Supervising decoder training using a composite loss function, including a weighted combination of Dice loss and binary cross entropy, and taking maximum diaphragm top position error as an auxiliary optimization indicator.

7. The 4D CBCT image reconstruction apparatus of claim 5, wherein, The projection set construction module comprises: A smoothing and interpolation unit configured to perform smoothing and interpolation processing on the diaphragm height curve to determine a periodic breathing waveform; A phase division unit configured to determine a breathing period according to the periodic breathing waveform, and divide each breathing period into m breathing phases; A screening unit configured to screen a number of projection images belonging to each breathing phase to construct a phase projection set.

8. The 4D CBCT image reconstruction apparatus of claim 7, wherein, The screening unit screens projection images corresponding to each breathing phase by time index or similarity matching method.

9. A computing device, characterized by Comprise: One or more processors; Memory; And one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by one or more processors, one or more programs include the diaphragm segmentation based 4D CBCT image reconstruction method of any one of the above claims 1-4.

10. Storage medium, characterized in that The storage medium stores one or more computer readable programs, and the one or more programs include instructions suitable for being loaded by the memory and executing the diaphragm segmentation based 4D CBCT image reconstruction method of any one of the above claims 1-4.

Images