A flat panel ct projection domain conditional super-resolution reconstruction method and system

Abstract

The application relates to the technical field of medical image reconstruction, and discloses a flat-plate CT image reconstruction method and system. The method acquires low-resolution projections collected by adopting a specific binning mode and an angle sampling scheme, constructs a condition vector containing binning parameters, angle parameters and system geometry parameters, inputs the condition vector and adjacent angle projections into a projection domain conditional super-resolution network, then performs data consistency correction based on a binning degradation operator and frequency domain compensation based on a flat-plate detector MTF model, and finally performs cone beam or fan beam reconstruction to obtain a high-resolution flat-plate CT image. The application strongly binds the super-resolution process with the acquisition physical process, and can improve the spatial resolution and edge fidelity under the conditions of low dose, sparse angle or large binning scanning.

Description

Technical Field

[0001] This invention relates to the field of medical image reconstruction and computer-aided imaging technology, and in particular to a method, system, device and storage medium for super-resolution reconstruction under projection domain conditions for flat panel CT, cone-beam CT or medical CT equipment with a flat panel CT structure. Background Technology

[0002] Flat-panel CT typically employs a two-dimensional flat-panel detector positioned opposite the X-ray source. It acquires projection data from multiple angles through rotational scanning, and then uses cone-beam or fan-beam reconstruction algorithms to obtain volumetric data. Compared to traditional multi-slice spiral CT, flat-panel CT offers advantages such as flexible equipment structure, wider field of view, higher spatial resolution, and easier integration with treatment equipment in scenarios including oral and maxillofacial surgery, orthopedic joint surgery, intraoperative navigation, interventional therapy localization, and radiotherapy image guidance. Therefore, it has become one of the important technical approaches in clinical imaging.

[0003] However, in actual clinical use, flat-panel CT often faces two conflicting demands. The first conflict is between dose and resolution. To reduce patient dose, shorten exposure time, or reduce motion artifacts, the equipment often needs to reduce tube current, shorten single-frame exposure time, and adopt a larger binning mode on the detector side, such as merging 2×2 or 4×4 detector pixels for readout. Binning can improve the signal-to-noise ratio of a single readout and reduce data throughput, but it directly reduces the projection sampling resolution, causing high-frequency details to be lost prematurely in the projection domain, resulting in blunted edges in the reconstructed image, blurred fine cortical bone structures, and poor visualization of small instruments. The second conflict is between scan time and angular sampling density. For intraoperative navigation, chest or abdominal examinations with poor respiratory coordination, and rapid rescanning scenarios in interventional treatments, the system often uses a larger angular interval for sparse angular sampling to shorten the scan time. However, this can lead to problems such as stripe artifacts, undersampling artifacts, and local structural discontinuities in the reconstructed image.

[0004] To address the aforementioned issues, common approaches in existing technologies include: first, performing image-domain super-resolution processing on CT images after reconstruction; second, establishing a multi-round iterative re-irradiation process between projection reconstruction and image enhancement; and third, using generative networks oriented towards natural images to directly map details to the image-domain results. While these solutions can improve visual clarity on certain datasets, they still have significant shortcomings in flat-panel CT applications: on the one hand, if the super-resolution process does not establish a clear relationship with the binning mode, angle sampling scheme, and system geometric parameters during actual acquisition, the recovered high-frequency information is difficult to maintain physical consistency with the measured projection, easily leading to texture "illusions," edge position drift, and grayscale quantitative deviations; on the other hand, if enhancement is only performed in the image domain, the network cannot fully utilize the geometric constraints, viewing angle constraints, and detector transfer characteristics still retained in the projection domain, easily amplifying metal artifacts, scattering residues, and reconstruction artifacts; furthermore, the control process for multi-round iterative re-irradiation is lengthy and computationally burdensome, making system implementation complex and unsuitable for intraoperative or online clinical workflow deployment.

[0005] The imaging quality of flat-panel CT is affected not only by the reconstruction algorithm but also by factors such as the detector's intrinsic modulation transfer function (MTF), scintillator diffusion characteristics, focal spot size, source-detector distance, source-object distance, pixel pitch, binning rules, and angle sampling strategies. In particular, changes in the binning mode alter the detector's effective point spread response and frequency domain response; changes in the angle sampling scheme also change the complementarity and propagable information range between adjacent projected views. If super-resolution methods do not explicitly incorporate these parameters into the network input or post-processing constraints, relying solely on "clearer-looking" visual targets for learning often fails to meet the requirements of edge location accuracy, structural continuity, and reconstruction reliability in medical imaging scenarios.

[0006] Therefore, there is an urgent need to propose an image reconstruction technology solution for flat panel CT that can closely bind the super-resolution process of the projection domain with the actual physical process of acquisition, so that it can improve spatial resolution and edge fidelity in scenarios such as low dose, large binning, sparse angle or fast scanning, while maintaining consistency with the measured projection, and taking into account the system's feasibility and real-time clinical use. Summary of the Invention

[0007] The purpose of this invention is to provide a method and system for super-resolution reconstruction under projection domain conditions in flat-panel CT, so as to at least solve the problems in the prior art such as insufficient coupling between the super-resolution scheme and the acquisition parameters of flat-panel CT, the easy appearance of pseudo-details that do not conform to physical constraints in the reconstruction results, and the poor stability of image edge recovery under large binning and sparse angle acquisition conditions.

[0008] To achieve the above objectives, the present invention provides the following technical solution: A method for reconstructing flat-panel CT images includes the following steps: Acquire measured projection data collected by a flat panel CT system under a preset scanning protocol, wherein the preset scanning protocol includes at least binning mode, angle sampling scheme and system geometric parameters; The measured projection data is subjected to dark field correction, flat field correction, bad pixel correction, logarithmic transformation and geometric normalization to obtain a low-resolution projection sequence; Based on the binning mode, angle sampling scheme and system geometric parameters, a conditional vector is constructed, and the low-resolution projection of the current angle, the projection of the acquired angle adjacent to the current angle and the conditional vector are input into a pre-trained projection domain conditional super-resolution network to obtain the first high-resolution projection. Data consistency correction is performed on the first high-resolution projection based on the binning degradation operator corresponding to the binning mode, so that the pseudo low-resolution projection processed by the binning degradation operator matches the low-resolution projection of the current angle, and the fusion range and fusion weight of adjacent angle features are constrained based on the angle sampling scheme to obtain the second high-resolution projection. Based on the modulation transfer function (MTF) model of the flat panel detector in the binning mode, frequency domain compensation and noise suppression coupling correction are performed on the second high-resolution projection to obtain the target high-resolution projection sequence; according to the target high-resolution projection sequence and the system geometric parameters, cone-beam reconstruction or fan-beam reconstruction is performed to output the flat panel CT reconstruction image.

[0009] Furthermore, the condition vector includes at least: binning factor, detector pixel pitch, distance from X-ray source to flat panel detector, distance from X-ray source to rotation center, current projection angle index, adjacent sampling angle interval, X-ray tube voltage, X-ray tube current, exposure time, and MTF identifier parameter corresponding to the flat panel detector.

[0010] Furthermore, the projection domain conditional super-resolution network includes a current view coding branch, a neighboring view coding branch, a conditional embedding branch, a feature fusion branch, and a residual decoding branch; the conditional features output by the conditional embedding branch are injected into the current view coding branch and the neighboring view coding branch through a feature linear modulation layer, a conditional normalization layer, or an attention gating layer to output the first high-resolution projection.

[0011] Furthermore, the data consistency correction includes: The first high-resolution projection is degraded using the same binning window and aggregation weights as during acquisition to obtain a pseudo-low-resolution projection. Calculate the projection residual between the pseudo-low-resolution projection and the measured low-resolution projection; The projection residual is written back to the high-resolution grid according to the inverse mapping relationship of the binning window, and the first high-resolution projection is corrected accordingly to obtain the second high-resolution projection.

[0012] Furthermore, the constraint on the fusion range and fusion weight of adjacent angle features based on the angle sampling scheme includes: A view reachable mask is constructed based on the angle difference between the current projection angle and the adjacent acquired projection angles; Block adjacent viewpoint features that exceed the preset angle window; The fusion weights of unmasked adjacent viewpoint features are determined based on the angular difference and the system's geometric magnification.

[0013] Furthermore, the frequency domain compensation and noise suppression coupling correction based on the modulation transfer function (MTF) model includes: Perform frequency domain transformation on the second high-resolution projection; Construct a target compensation filter based on the MTF curve corresponding to the current binning mode; The high-frequency components are enhanced in a restricted manner using the target compensation filter and the noise power spectrum suppression window function; and the enhanced frequency domain projection is subjected to an inverse transformation to obtain the target high-resolution projection sequence.

[0014] Furthermore, the projection domain conditional super-resolution network is trained in the following manner: Acquire high-resolution true projections of flat-panel CT systems in small binning mode and high angle sampling density; Binning degradation, angle undersampling degradation, noise injection, and MTF blurring are applied to the high-resolution ground truth projection to obtain the training input; The parameters of the projection domain conditional super-resolution network are optimized using the weighted sum of binning consistency loss, frequency domain MTF loss, projection gradient loss, and reconstruction domain structural similarity loss as the objective function.

[0015] Furthermore, including: The scanning control module is used to control the X-ray source, flat panel detector and rotating mechanism to acquire measured projection data according to the preset scanning protocol, and record the binning mode, angle sampling scheme and system geometric parameters; The preprocessing module is used to perform dark field correction, flat field correction, bad pixel correction, logarithmic transformation and geometric normalization on the measured projection data to obtain a low-resolution projection sequence. The condition construction module is used to construct a condition vector based on the binning mode, angle sampling scheme, and system geometric parameters; The projection domain conditional super-resolution module is used to input the low-resolution projection of the current angle, the projections of adjacent acquired angles, and the conditional vector into a pre-trained projection domain conditional super-resolution network to obtain a first high-resolution projection. The data consistency correction module is used to perform data consistency correction on the first high-resolution projection according to the binning degradation operator corresponding to the binning mode and the angle sampling scheme to obtain the second high-resolution projection. The MTF constraint module is used to perform frequency domain compensation and noise suppression coupled correction on the second high-resolution projection based on the modulation transfer function (MTF) model of the flat panel detector in the binning mode, so as to obtain the target high-resolution projection sequence. The reconstruction module is used to perform cone-beam reconstruction or fan-beam reconstruction based on the target high-resolution projection sequence and the system geometric parameters, and output a flat-panel CT reconstruction image.

[0016] Furthermore, the X-ray source and the flat panel detector are respectively located at opposite ends of the rotating gantry or C-arm, and are positioned relative to each other around the rotation center of the subject; the table passes through the rotation center; the flat panel detector is connected to the acquisition controller via the detector controller, the acquisition controller is communicatively connected to the X-ray source controller, the rotation drive mechanism and the reconstruction server, and the reconstruction server is connected to the clinical workstation.

[0017] Furthermore, a computer-readable storage medium storing a computer program thereon, characterized in that, when the computer program is executed by a processor, it implements the flat panel CT image reconstruction method according to any one of claims 1 to 7.

[0018] Compared with the prior art, the present invention has at least the following beneficial effects: First, this invention places super-resolution processing in the projection domain and uses binning mode, angle sampling scheme and system geometric parameters as conditional inputs, so that the restoration process is directly oriented towards the real acquisition link of flat panel CT, rather than performing posterior enhancement on the reconstructed image out of the acquisition context. Therefore, it is more conducive to maintaining the accuracy of structural position and grayscale consistency.

[0019] Second, this invention adds projection data consistency correction based on the binning degradation operator after the network output, which enables the recovered high-resolution projection to be consistent with the measured projection after degrading back to the low-resolution space, thereby significantly suppressing false details and unreliable textures caused by the network's free generation and improving interpretability in medical scenarios.

[0020] Third, this invention explicitly introduces the flat panel detector MTF model into the super-resolution link, limits the enhancement range at the frequency domain level and combines it with noise suppression, so that high-frequency recovery is no longer unconstrained sharpening, but rather constrained compensation combined with the detector's transmission characteristics, which is more suitable for the physical characteristics of flat panel CT in different readout modes.

[0021] Fourth, this invention constrains cross-view feature propagation through an angle sampling scheme, enabling the network to adopt different view fusion strategies under sparse angle and dense sampling protocols, thereby improving the compatibility between different clinical scanning protocols. It is especially suitable for speed and dose-sensitive scenarios such as intraoperative rescanning, oral CBCT, orthopedic joint CBCT, interventional navigation, and image-guided radiotherapy.

[0022] Fifth, this invention can achieve high-quality reconstruction without relying on pure generative adversarial mapping in the image domain or multiple rounds of reconstructed image re-irradiation, by utilizing single projection domain forward recovery, data consistency correction, and MTF-limited compensation. The process is shorter, the control logic is clearer, and it is easier to implement in engineering and deploy in clinical settings. Attached Figure Description

[0023] Figure 1 This is a schematic diagram of the overall structure of the flat-panel CT image reconstruction system in an embodiment of the present invention; Figure 2 This is a schematic flowchart of the flat-panel CT image reconstruction method in an embodiment of the present invention; Figure 3 This is a schematic diagram of the projection domain conditional super-resolution network in an embodiment of the present invention; Figure 4 This is a schematic diagram of the data consistency correction and MTF constraint processing flow in an embodiment of the present invention; Figure 5 This is a schematic diagram illustrating the process of constructing network training data and optimizing the loss function in an embodiment of the present invention.

[0024] 11. X-ray source; 12. Flat panel detector; 13. Bed table; 14. Driver; 15. Acquisition controller; 20. Acquisition controller (image acquisition / data receiving unit); 30. Reconstruction server; 40. Clinical workstation; 50. Storage unit. Detailed Implementation

[0025] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments. It should be understood that the following embodiments are for illustrative purposes only and are not intended to limit the scope of protection of the present invention. For those skilled in the art, any equivalent substitutions, combinations, or sequential adjustments to the technical features described in the various embodiments without departing from the concept of the present invention should fall within the scope of protection of the present invention.

[0026] I. Terminology and Scope of Applicable Equipment In this application, "flat panel CT" preferably refers to a medical imaging device that uses a two-dimensional flat panel detector to perform projection acquisition and CT reconstruction, including but not limited to C-arm cone-beam CT, gantry flat panel CT, oral and maxillofacial CBCT, orthopedic joint CBCT, intraoperative navigation flat panel CT, and flat panel CT modules in radiotherapy image guidance systems. The flat panel detector can be an amorphous silicon flat panel detector, a CMOS flat panel detector, or other digital detectors capable of outputting two-dimensional projection images.

[0027] In this application, "binning mode" refers to the method by which a flat panel detector merges adjacent pixels during readout. Taking 2×2 binning as an example, the detector controller performs charge merging or digital aggregation on adjacent 2 rows and 2 columns of pixels in the high-resolution grid, and then outputs a low-resolution pixel value. Different binning modes not only affect the number of pixels, but also change the effective pixel pitch, dot spread response, MTF curve, readout noise performance, and subsequent data throughput. The "strong binding with binning" in this invention does not merely use the binning factor as a label, but incorporates the degradation operators, conditional parameters, and MTF differences related to binning into the reconstruction link simultaneously.

[0028] In this application, "angle sampling scheme" refers to a combination of parameters during the scanning process, including the angular coverage area of ​​the projection, the starting angle, the ending angle, the angle interval, the total number of views, and whether there are skipped angles or irregular sampling. For example, in some intraoperative rapid rescanning scenarios, only 300 frames of projection within a 200° angle range may be acquired; in some oral CBCT scenarios, 400 or 600 frames of projection within a 360° range may be acquired; and in some low-dose schemes, the angle interval may be deliberately increased to reduce the total number of exposure frames. The "strong binding with angle sampling" of this invention means that when the network utilizes cross-view information, it must control feature propagation according to the actual sampling interval and the reachable range of the view, and not forcibly apply cross-angle relationships that do not exist to the sparse sampling protocol.

[0029] II. Overall System Structure, Component Positional Relationships, and Connections like Figure 1 As shown, the flat panel CT image reconstruction system of this embodiment includes: a rotating gantry 10, an X-ray source 11, a flat panel detector 12, a bed 13, a source controller 14, a detector controller 15, an acquisition controller 20, a reconstruction server 30, a clinical workstation 40, and a storage unit 50.

[0030] The rotating gantry 10 can be a ring-shaped gantry structure or a C-arm structure. Regardless of the structure, the X-ray source 11 and the flat panel detector 12 are located on the same rotating component and are positioned opposite each other, with the line connecting them passing through the imaging rotation center O. In a typical structure, the X-ray source 11 is located at one end of the rotating gantry 10, and the flat panel detector 12 is located at the other end, with their central rays pointing towards the rotation center O. The table 13 is arranged along the patient's body axis, with its centerline passing through the rotation center O, so that the patient's examination site can be located near the scanning isocenter. For dental CBCT equipment, the table 13 can be replaced by a sitting positioning support frame or a headrest assembly; for orthopedic standing positioning equipment, the table 13 can be replaced by a standing platform, but neither affects the applicability of the present invention to the source, detector, and control logic.

[0031] X-ray source 11 is electrically connected to source controller 14. Source controller 14 controls tube voltage (kV), tube current (mA), pulse width, focus mode, and exposure trigger timing. Flat panel detector 12 is electrically connected to detector controller 15 or via fiber optic connection. Detector controller 15 controls gain mode, readout mode, binning mode, integration time, bad pixel mapping table access, and image frame buffering. Detector controller 15 is further connected to acquisition controller 20 via a high-speed data link. The high-speed data link can be fiber optic Ethernet, Camera Link, CoaXPress, PCIe extension link, or other industrial buses that meet bandwidth requirements.

[0032] The acquisition controller 20 is connected to the source controller 14, the detector controller 15, and the drive unit of the rotating frame 10. The drive unit of the rotating frame 10 includes at least a rotary motor, an angle encoder, and a position feedback unit. The acquisition controller 20 synchronously controls the source exposure, detector readout, and frame angle position according to the scanning protocol, so that each frame projection is precisely correlated with the corresponding angle value. The acquisition controller 20 is also connected to the motion control unit of the bed 13 for performing axial bed feeding, positioning calibration, or repositioning in some embodiments.

[0033] The reconstruction server 30 is communicatively connected to the acquisition controller 20. The reconstruction server 30 can be a GPU server, a CPU / GPU hybrid workstation, or a dedicated edge inference box. The reconstruction server 30 is responsible for performing projection preprocessing, conditional vector construction, projection domain conditional super-resolution inference, binning consistency correction, MTF-limited compensation, and cone-beam reconstruction. The reconstruction server 30 is connected to the clinical workstation 40, which is used to display cross-sectional, coronal, sagittal, and surface reconstructions, volume rendering, maximum density projection, or device navigation overlay results. The storage unit 50 is connected to the reconstruction server 30 and is used to store the original projection, protocol parameters, reconstructed images, log information, and model parameter files.

[0034] In a preferred embodiment, the distance from the X-ray source 11 to the flat panel detector 12 is denoted as SDD, the distance from the X-ray source 11 to the rotation center O is denoted as SOD, the detector pixel pitch is denoted as p, and the current binning factor is denoted as b. For different binning modes of 1×1, 2×2, or 4×4, the system records the corresponding readout rules in the detector controller 15 and reads the corresponding degradation operator parameters and MTF index in the reconstruction server 30. In other words, the key point of the control logic in this invention is that it does not only transmit the rough information of "how much the image size has been reduced" to the network, but also transforms the physical degradation, frequency domain changes, and acquisition protocol changes caused by binning into a computable, traceable, and constrainable model input.

[0035] III. Scanning Protocol Planning and Acquisition Control Logic like Figure 2 As shown, the reconstruction method in this embodiment can be divided into seven main stages: scan protocol planning, projection acquisition, projection preprocessing, conditional super-resolution, data consistency correction, MTF-limited compensation, and volumetric data reconstruction. To meet the different requirements of clinical scenarios for dose, time, and resolution, the acquisition controller 20 first receives information such as the examination site, patient position, target voxel resolution, dose level, and motion risk level from the clinical workstation 40 before the scan begins, and selects or generates a scan protocol accordingly.

[0036] For example, in intraoperative spinal screw localization rescanning scenarios, the system can select the "low-dose rapid rescanning protocol": tube voltage 80kV~100kV, short single-frame exposure time, 2×2 binning detector, total sampling angle range 200°~220°, large angle interval, and fewer views. For oral and maxillofacial imaging scenarios, the "high spatial resolution protocol" can be selected: tube voltage around 90kV, 1×1 or 2×2 binning detector, 360° angle coverage, and more views. For orthopedic joint imaging scenarios, different binning and angle strategies can be determined based on factors such as standing or supine position, presence of metal implants, and whether navigation overlay is required. The above protocols themselves do not limit the invention, but the advantage of this invention lies in its ability to uniformly perform conditional reconstruction under these different protocols.

[0037] According to the protocol, the data acquisition controller 20 sends the following control parameters: Source control parameters Detector control parameters Geometric control parameters Angle control parameters Where b is Factor, gain refers to the gain mode. This determines whether to enable local readout regions. The controller writes these parameters into the metadata of each frame's projection, enabling subsequent reconstruction to track the actual acquisition conditions frame by frame.

[0038] During the acquisition process, the acquisition controller 20 triggers the X-ray source 11 to emit pulses based on the current gantry angle θi fed back by the angle encoder, and simultaneously drives the flat panel detector 12 to complete one frame acquisition in the current binning mode. The detector controller 15 outputs the raw projection frame Iraw(i) and transmits it, along with the dark field frame D, the flat field frame F, the angle value θi, the frame number i, and the protocol parameters, to the acquisition controller 20 or directly caches it in the reconstruction server 30. If the device supports real-time streaming reconstruction, the reconstruction server 30 can process the data while acquiring it; if the device uses offline reconstruction, it can process the data in batches after the scan is completed. For clinical workflows requiring online navigation, the system preferably adopts a method of acquiring and preprocessing simultaneously, and performing super-resolution and reconstruction immediately after the scan is completed, in order to shorten the image output time.

[0039] It should be noted that this invention does not require the acquisition controller 20 to perform complex image enhancement calculations during the scanning phase. Its main function is to keep the source, detector, and rack control in a time-synchronized state and to fully record the acquisition conditions matching each frame of projection. Because each frame of projection carries real binning patterns, angle information, and geometric information, the subsequent conditional super-resolution network is not "blind enhancement" but rather reasoning constrained by physical conditions with a clear context.

[0040] IV. Projection Preprocessing Process After receiving the measured projection data, the reconstruction server 30 first performs preprocessing. Preprocessing can be performed in the following order: dark field correction, flat field correction, bad pixel repair, nonlinear correction, logarithmic transformation, geometric normalization, and optional scattering correction. Preferably, for each frame of original projection Iraw(i), the preprocessing is performed according to the formula... Normalization is performed, where ε is a minimal constant to prevent the denominator from being zero; then, obviously abnormal bad points or failure columns are repaired using neighborhood median, bilateral interpolation, or a pre-calibrated bad point mapping table; subsequently, nonlinear response correction is performed based on the system calibration curve; finally, a logarithmic transformation is performed to obtain the line integral projection. .

[0041] For devices employing offset detectors or detectors with lateral offset relative to the rotation center, geometric normalization and coordinate alignment of the projection can be performed during the preprocessing stage, mapping projections from different protocols to a unified reconstruction coordinate system. For scenarios with significant scattering effects, scattering correction steps based on main beam occlusion calibration, first-order scattering estimation, or convolution approximation can be added during the preprocessing stage. These steps are optional conventional preprocessing steps in flat-panel CT reconstruction. This invention is not limited to any particular preprocessing method, but preferably requires that the preprocessed projection reflects the true line integral information and maintains consistent annotation with the scanning protocol.

[0042] After preprocessing, the low-resolution projection sequence plr = plr(i) is obtained. It should be noted that "low resolution" here is relative to the target recovery resolution. If 2×2 binning is used during acquisition, and the target recovery resolution corresponds to a 1×1 grid, then plr(i) belongs to low-resolution projection; if 1×1 mode is used during acquisition and this invention is mainly used for robust detail recovery under sparse angle conditions, then although plr(i) is already the highest number of pixels at the detector's raw readout level, it can still be regarded as an input that needs conditional enhancement processing in the context of adjacent viewpoints and MTF compensation.

[0043] V. Logic for Constructing Condition Vectors To ensure the super-resolution network's output is strongly bound to the acquisition protocol, the reconstruction server 30 further constructs a conditional vector ci after generating Plr. For the i-th frame projection, the preferred conditional vector ci includes at least: the current binning factor b, the effective pixel pitch p×b, the source-detector distance SDD, the source-object distance SOD, the current gantry angle θi, the adjacent sampling angle interval Δθ, the tube voltage kV, the tube current mA, the single-frame exposure time texp, the focal mode focal_mode, the detector gain mode gain, and the MTF index mid corresponding to the current detector. Additional conditions such as patient size level, examination site identification, collimator settings, filter grid status, or reconstructed voxel target size can also be added as needed.

[0044] In the preferred implementation, the reconstruction server 30 first normalizes the aforementioned values. For example, the angle θi can be mapped to a dual-channel sinθi and cosθi to avoid periodic angle jumps; the binning factor can be a one-hot vector or a combination of numerical encoding and learnable embedding; continuous parameters such as kV, mA, and texp can be linearly normalized to between 0 and 1 within a preset range; the MTF index can correspond to a set of frequency response features in a lookup table. Subsequently, the system concatenates these parameters into an original conditional vector and generates conditional features fc(i) through a lightweight fully connected network or a one-dimensional convolutional embedding module.

[0045] The significance of constructing conditional vectors lies in enabling the network to explicitly understand "which acquisition scenario it is facing" when reconstructing projections. For example, under protocols with 2×2 binning and large angular intervals, the network should prioritize data consistency and limited high-frequency compensation, reducing aggressive detail generation; while under 1×1 or more densely spaced angular protocols, the network can make greater use of complementary information from adjacent viewpoints to recover details. In other words, this invention explicitly injects "protocol knowledge" into the network through conditional vectors, rather than attempting to make the network implicitly guess protocol differences from all training samples.

[0046] VI. Projection Domain Conditional Super-Resolution Network Structure and Control Logic like Figure 3 As shown, the projection domain conditional super-resolution network 100 in this embodiment includes at least a current view coding branch 110, a neighboring view coding branch 120, a conditional embedding branch 130, a feature fusion branch 140, and a residual decoding branch 150.

[0047] The current view encoding branch 110 is used to extract features from the low-resolution projection prl(i) of the current angle. This branch can employ several two-dimensional convolutional layers, residual blocks, dilated convolutional blocks, or lightweight Transformer blocks. Its function is to extract information such as edge responses, texture gradients, metallic shadow distribution, and low-frequency attenuation patterns in the current view. The neighboring view encoding branch 120 is used to encode several acquired view projections adjacent to the current angle. In a preferred embodiment, the adjacent projection views can be taken from K frames before and after the current angle, where K can be dynamically adjusted according to the angle interval Δθ. For example, when Δθ is large, K=1 or 2, and when Δθ is small, K=2 or 3. The purpose of this is to make full use of view complementarity when the angles are dense, and to avoid error propagation caused by excessively distant views when the angles are sparse.

[0048] Conditional embedding branch 130 receives the conditional vector ci and generates conditional features fc(i). The conditional features fc(i) can be injected into multiple levels in the current viewpoint coding branch 110 and the neighboring viewpoint coding branch 120, respectively. The injection method can be: performing conditional normalization after the convolution output, i.e., using fc(i) to generate scaling parameter γ and offset parameter β, and then applying this normalization to the feature map. It can also be linear modulation using the FiLM method; or it can be using fc(i) as an additional query vector when calculating attention weights. Regardless of the method used, the core idea is to make the network aware of the current acquisition protocol at each level of feature extraction, rather than simply concatenating a label at the input.

[0049] Feature fusion branch 140 is used to fuse current viewpoint features, neighboring viewpoint features, and conditional features. Preferably, feature fusion branch 140 includes a viewpoint reachability mask unit 141, an angle position encoding unit 142, and an adaptive fusion unit 143. The viewpoint reachability mask unit 141 generates a mask Mij based on the angle difference Δθij between the current angle and neighboring angles. When Δθij exceeds a preset window, the corresponding neighboring features are suppressed or set to zero. The angle position encoding unit 142 is used to map the angle differences between different neighboring viewpoints relative to the current viewpoint into learnable position codes. The adaptive fusion unit 143 weights the neighboring features according to the angle difference, geometric magnification, and local structural similarity. In this way, the network does not simply average multi-view information, but limits the multi-view propagation within a reasonable range, achieving a strong binding with the angle sampling scheme.

[0050] The residual decoding branch 150 is responsible for outputting the high-resolution residual ri of the current angle, and adding it to the basis projection u(plr(i)) obtained by geometrically consistent upsampling of the low-resolution projection to obtain the first high-resolution projection. Here, u(·) can be bicubic interpolation, zero-interpolation with convolutional thinning, a learnable upsampling module, or an anti-aggregation initial value mapping that matches the binning window. Preferably, u(·) adopts an initial value unrolling method consistent with the actual binning layout, so that the residual decoding branch 150 is mainly responsible for recovering the missing high-frequency information, rather than simultaneously undertaking the large-scale geometric alignment task. This reduces the learning difficulty of the network and makes the output more stable.

[0051] In terms of control logic, after acquisition, reconstruction server 30 calls network 100 sequentially or in parallel according to the projection sequence number. For the i-th frame projection, the server reads prl(i) and the neighborhood projection set N(i) from the cache, reads the condition vector ci, generates features, and completes forward inference. If the system is in real-time mode, a sliding window approach can also be used to update the neighborhood cache when a new projection arrives and output phr1(i) immediately. The characteristic of this processing flow is that it does not require obtaining the reconstructed image first and then feeding it back into the network, but completes the main detail recovery in one forward pass in the projection domain, which is convenient for online deployment.

[0052] VII. Data consistency correction mechanism strongly bound to binning like Figure 4As shown, although the first high-resolution projection phr1(i) output by the network already possesses good detail representation capabilities, it may still have local high-frequency biases. Therefore, this invention adds a data consistency correction step strongly bound to the binning mode after the network. The key to this step is not simply comparing high-resolution and low-resolution images, but rather matching the network output, degraded back to a low-resolution space, with the measured projection item by item according to the rule of "how to transform from a high-resolution grid to a low-resolution readout during actual acquisition."

[0053] Let Bb be the binning degradation operator of the current acquisition protocol. For 2×2 binning, Bb can be understood as aggregating high-resolution pixels using a 2×2 window, also considering pixel fill rate, scintillator spread, and readout weights when necessary; for 4×4 binning, a corresponding aggregation is performed using a 4×4 window. The system first calculates the pseudo-low-resolution projection. Subsequently, it was compared with the measured low-resolution projection plc(i) to obtain the projection residual. .

[0054] Next, the system writes the residual ei back to the high-resolution grid according to the inverse mapping relationship Ub corresponding to Bb. For example, in the simplest uniform inverse allocation strategy, the residual of each low-resolution pixel of ei can be evenly allocated back to its corresponding high-resolution window; in a more preferred implementation, ei can be non-uniformly allocated to high-resolution sub-pixels according to the gradient distribution of phr1(i) in the current window, the local edge direction, or prior weights, to avoid excessive edge smoothing. The corrected second high-resolution projection can be expressed as... , where α is the correction coefficient, which can be 1 or set to a value less than 1 based on the local noise estimation.

[0055] Through the above mechanism, the system can guarantee the following physical consistency: if the final high-resolution projection is degraded again according to the current binning rule, the original low-resolution measured projection should be recovered as much as possible. This is especially important for medical imaging, because clinical needs are not just about "looking sharp," but about "maintaining consistency with the measurement when degrading back to the original acquisition resolution." Compared with pure learning-based super-resolution methods that only use L1 or perceptual loss, this projection domain data consistency correction can significantly reduce the risk of introducing artifacts and false details.

[0056] It should be noted that data consistency correction does not require the reconstruction server 30 to undergo multiple large-scale iterations. This invention preferably achieves stable improvement with a single or few residual write-backs, thus avoiding the problems of lengthy processes, complex parameter coupling, and poor real-time performance found in traditional iterative backfeedback schemes.

[0057] 8. Limited Frequency Domain Compensation Mechanism Strongly Bound to MTF After ensuring binning consistency, the system further utilizes the MTF model of the flat panel detector under the current protocol to perform restricted frequency domain compensation. This is because the upper limit of the detector's high-frequency response changes with the binning mode. Directly sharpening without considering the MTF can easily amplify noise as detail; conversely, overly conservative approaches fail to recover edges. This invention employs a restricted compensation filter constructed in the frequency domain based on the actual detector response curve, achieving a balance between "maximizing recovery within the physically permissible bandwidth and suppressing overshoot in high-noise regions."

[0058] Specifically, a two-dimensional Fourier transform is performed on the second high-resolution projection phr2(i) to obtain F(phr2(i)). The system retrieves the MTF curve MTFb(fx,fy) corresponding to the current binning mode b and the current detector state from the MTF lookup table. If the system has calibrated the target MTF curve MTFref(fx,fy) in 1×1 mode, a target compensation filter can be constructed. Here, W(fx,fy) is the noise suppression window function, which can be in the form of a Wiener window, a cosine roll-off window, a Butterworth window, or a stabilization weight related to the noise power spectrum NPS. By introducing W(fx,fy), the filter will not gain infinitely near the cutoff frequency, but will gradually converge according to the noise level.

[0059] Subsequently, the system calculates Then, an inverse Fourier transform is performed to obtain the target high-resolution projection phr3(i). In engineering implementation, the MTF lookup table can be obtained during the equipment factory calibration stage using the hypotenuse method, line-pair phantom method, or point response method, and the frequency response curves in 1×1, 2×2, and 4×4 modes can be stored respectively. The reconstruction server 30 only needs to read the corresponding parameters based on mid during runtime to complete rapid compensation.

[0060] The combined effect of MTF-limited compensation and binning consistency correction forms a key feature that distinguishes this invention from ordinary projection super-resolution methods: the former ensures that "recovery cannot violate the actual achievable bandwidth and noise characteristics of the detector," while the latter ensures that "recovery cannot violate the original measured projection readout results." The combination of these two ensures that the final high-resolution projection possesses both good edge sharpness and is less prone to generating unreliable pseudo-high frequencies.

[0061] IX. Reconstruction Module and Clinical Output Control Logic High-resolution projection sequence of targets from all angles After generation, the reconstruction server 30 calls the reconstruction module to perform volumetric data reconstruction. For C-arm cone-beam CT equipment, the reconstruction module can use FDK (Feldkamp-Davis-Kress) reconstruction; for some quasi-fan-beam geometry equipment, fan-beam filtered backprojection or model-based reconstruction can also be used. The system preferably uses weighted FDK, that is, introducing weights determined by data consistency residuals and local noise estimations during the filtering and backprojection stages, so that projection areas with better consistency contribute more to the reconstruction.

[0062] Specifically, the reconstruction module first performs geometric correction based on SDD, SOD, detector offset, rotation center offset, and angle sequence. Then, it performs reconstruction filtering on phr3(i) and performs back-projection accumulation according to geometric relationships to obtain three-dimensional volume data V(x,y,z). In some embodiments, the system can also add local artifact suppression, edge smoothing, or voxel consistency optimization for metal implants after reconstruction, but this is not a necessary limitation of the present invention.

[0063] After reconstruction, the clinical workstation 40 automatically loads the corresponding display protocol based on the examined area. For example, in orthopedic scenarios, the workstation preferably displays cross-sections, coronal planes, and 3D volume renderings of the bone window; in oral scenarios, it preferably displays arch surface reconstruction and magnified views of the periapical region; in interventional scenarios, the reconstructed volume data can be overlaid with intraoperative instrument positions, navigation paths, or previous surgical plans. Because this invention achieves high-resolution reconstruction under physical constraints in the projection domain, it typically provides clearer and more stable display results than traditional low-dose large-binning reconstructions for clinically significant structures such as cortical bone margins, periapical regions of teeth, small frameworks, guidewires, and trabeculae.

[0064] 10. Network Training Process and Parameter Optimization like Figure 5 As shown, although claim 1 mainly defines the online reconstruction process, the network training process is further described below to enable those skilled in the art to implement the present invention. Preferably, network training is performed in an offline environment, and the model weights are deployed to the reconstruction server 30 after training is completed. The training process includes four stages: training sample construction, input-output pairing, loss function design, and parameter update.

[0065] (a) Training the construction of the true value projection Preferably, the system selects the original projection of a high-quality flat-panel CT scan acquired under 1×1 binning, small angular interval, and high view density conditions as the ground truth projection Pgt. Ground truth data can come from phantom scans, animal experiments, anonymous clinical cases, or simulated projections. To enhance the model's generalization ability, the training set should cover different examination sites and body types, and include typical structures such as bone tissue, soft tissue, air interfaces, and metal instruments. For clinical data, it is preferable to extract pure projections and protocol parameters without compromising patient privacy protection requirements.

[0066] (ii) Training Input Generation Apply degradation consistent with the actual acquisition process to the ground truth projection Pgt to generate training input. The degradation process preferably includes the following sub-steps: 1. Perform pixel aggregation according to the target binning factor b, resulting in binning degradation; 2. Deleting or extracting parts of the view according to the target angle sampling scheme results in angle undersampling degradation; 3. Add Poisson noise and readout Gaussian noise according to the set dose level; 4. Apply corresponding MTF blurring according to the target binning mode and detector state; 5. As needed, add a small amount of scattering residue, gain fluctuation or bad pixel simulation to improve robustness.

[0067] The training input obtained in this way is closer to real low-dose, large-binning, sparse-angle clinical data, rather than just idealized samples constructed through simple downsampling.

[0068] (III) Network Input Organization For the i-th training sample, the system extracts the low-resolution projection prl_train(i) of the current viewpoint, the set of previous and next neighboring viewpoints Ntrain(i), and the conditional vector ci_train, and uses the ground truth high-resolution projection pgt(i) as the supervision target. If batch training is used, samples with different binning modes and different angle schemes can be mixed in a batch, and the conditional vector helps the network distinguish the protocols.

[0069] (iv) Loss Function Design The preferred total loss function can be expressed as:

[0070] Wherein, Ldc is the binning consistency loss, which can be expressed as the sum of the L1 distances between Bb(psr(i)) and plr_train(i); Lmtf is the frequency domain MTF loss, which can be expressed as the L1 or L2 distance after weighting in the frequency domain, used to constrain the frequency distribution of psr(i) to match the ground truth; Lgrad is the projection gradient loss, used to emphasize the consistency of edge positions and orientations; Lrec is the reconstruction domain loss, which is to quickly reconstruct the volume data from the projection of the network output and then compare it with the ground truth volume data, and can use L1, SSIM, or a combination of both; Lang is the angular consistency loss, used to constrain the structural continuity between adjacent viewpoints and the stability of cross-angle feature propagation. The weights λ1 to λ5 can be set according to task requirements. For example, in low-dose, large-binning scenarios, the weights of Ldc and Lmtf can be appropriately increased to enhance physical consistency.

[0071] It is important to note that this invention preferably does not use adversarial loss as the core loss term, nor does it rely on pure image domain generation of adversarial mappings as the primary source of detail. This is because medical projection data emphasizes the authenticity and physical consistency of structural locations, rather than the "visual realism" of natural images. Therefore, data consistency, frequency domain consistency, and structural similarity in the reconstruction domain are preferred as the primary objectives. Even when introducing discriminative auxiliary terms with extremely low weights, it should be done without violating physical constraints.

[0072] (v) Parameter update and deployment The training server updates network parameters using the backpropagation algorithm, employing the Adam, AdamW, or RMSProp optimizer. After training, the optimal weights that balance Ldc, Lrec, and clinical structural visibility on the validation set are selected and exported as a deployment model. The deployment model can be converted to ONNX, TensorRT, or other inference formats and loaded into the reconstruction server 30. Because the online workflow of this invention does not require multiple rounds of image re-evaluation, the inference latency after deployment is more easily controlled, making it suitable for intraoperative or bedside applications.

[0073] XI. Example 1: Intraoperative spinal re-scan scenario The following example, using rapid follow-up scanning after intraoperative spinal screw placement, illustrates the application process of this invention. The patient lies supine on the bed 13, with the vertebral body region to be observed located at the rotation center O. The rotating gantry 10 performs a scan covering an angle of approximately 200°. To minimize surgical interruption time and patient dose, the system selects a 2×2 binning mode, acquiring approximately 300 projection frames. The source controller 14 sets a shorter pulse exposure time, the detector controller 15 adopts the corresponding gain mode, and the acquisition controller 20 records the angle value and source / detector status for each frame.

[0074] After scanning, the reconstruction server 30 first performs dark field / flat field correction and logarithmic transformation to obtain a low-resolution projection sequence. Then, it constructs a condition vector based on information such as b=2, the current angular interval Δθ, SDD, SOD, kV, mA, and texp. For the i-th frame projection, the system selects one or two neighboring frames from before and after it, and inputs them together with the current projection into the network 100 to generate the first high-resolution projection. Subsequently, the system uses a 2×2 binning degradation operator to degrade the high-resolution projection back to the low-resolution space, compares it with the measured projection to obtain the residual, and then writes the residual back into the high-resolution grid to form the second high-resolution projection.

[0075] After completing data consistency correction, the system retrieves the pre-calibrated MTF curve of the device in 2×2 binning mode and performs restricted frequency domain compensation on the second high-resolution projection. The compensated projection sequence is then reconstructed using weighted FDK to obtain three-dimensional volume data. The clinical workstation 40 displays the transverse section of the bone window and multiplanar reconstruction images. For the surgeon, the most important aspect is the relationship between the screw channel and the pedicle cortex boundary. In this scenario, the present invention can improve the sharpness of the cortical bone edge, reduce the ambiguity caused by large binning, and help the surgeon determine whether the screw has broken through the cortex. At the same time, because the reconstruction result is consistent with the measured projection, it is less likely to introduce false bony structures.

[0076] 12. Example 2: Oral and Maxillofacial CBCT Scene In maxillofacial scenarios, patients are typically positioned in a sitting or standing position with their heads fixed by a headrest. An X-ray source 11 and a flat panel detector 12 perform a 360° or partial circular scan around the maxillofacial region. To complete imaging quickly and control the dosage, some devices employ 2×2 binning in standard mode, resulting in poor visualization of alveolar bone details, periapical contours, and periodontal ligament spaces.

[0077] When using this invention, the acquisition controller 20 records the binning mode, angle value, and exposure parameters of each frame of projection according to the oral protocol. After receiving the projection, the reconstruction server 30, in addition to conventional preprocessing, can also appropriately expand the neighborhood view window, taking into account the small field of view and high geometric stability of the oral device, for example, by taking 2 or 3 frames of neighborhood projection before and after as the network context. The conditional embedding branch 130 uses the oral mode parameters and MTF index of the current device to generate conditional features, guiding the network to adopt a recovery strategy more suitable for small-volume, high-contrast structures when recovering tooth boundaries and local details of the periapical region.

[0078] For the periapical region and alveolar bone margins, due to the inherent high gradient of these structures, the high-resolution projection output by the network, after binning consistency correction, still maintains consistency with the measured projection, avoiding unrealistic double edges caused by pure sharpening. Subsequently, MTF-restricted compensation further restores high frequencies, making the dentition boundaries in the reconstructed surface image clearer. Clinicians can obtain more stable boundary information when observing periapical lesions, periodontal bone resorption, or peri-implant bone walls.

[0079] 13. Example 3: Low-dose orthopedic joint standing flat panel CT scenario In weight-bearing examinations of joints such as the knee and ankle, the equipment sometimes employs a standing configuration. The patient stands on a platform, the joint to be examined is located near the center of rotation, and the source and detector rotate around the joint. Because the risk of motion is high when the patient remains standing, the scanning protocol is typically biased towards shorter time intervals, potentially resulting in larger angular intervals or fewer projection frames. For such scenarios, the angle sampling constraint mechanism of this invention offers significant advantages.

[0080] Specifically, when the angular interval Δθ in the conditional vector is large, the viewing angle can reach the mask unit 141, which automatically shrinks the neighborhood window, allowing only closer viewing angles to participate in the fusion, in order to avoid misalignment of joint edges caused by the propagation of long-distance viewing angles. At the same time, data consistency correction continues to ensure that the restored projection of each frame matches the measured data when degenerated back to the original acquisition resolution. Finally, MTF-limited compensation helps restore the articular surfaces and cortical bone edges. For standing joint weight-bearing axis analysis and osteophyte edge observation, this invention can provide a more reliable image basis under shorter scan protocols.

[0081] XIV. Online and Offline Coordination Methods in Control Logic This invention can be deployed as a fully offline reconstruction scheme or as a collaborative scheme of "online acquisition + near real-time reconstruction". In the fully offline scheme, after scanning, the acquisition controller 20 sends all projections and protocol parameters to the reconstruction server 30 at once, which then executes all processing steps sequentially. In the near real-time scheme, each time the acquisition controller 20 receives a new projection frame, it sends the frame and its metadata to the reconstruction server 30. The server first performs preprocessing and writes the data into a sliding window buffer; when a frame has sufficient neighboring views, the network forward computation is triggered. In this way, when scanning is complete, most of the projection preprocessing and some super-resolution work have been completed, and only the remaining frame processing, data consistency correction, and reconstruction are needed to generate the image.

[0082] In engineering implementation, both data consistency correction and MTF compensation can be designed as GPU-accelerated operators. Because the geometric correspondence between Bb and Ub is fixed under the same binning mode, the system can pre-generate an index mapping table; the MTF compensation filter G(fx,fy) can also be pre-cached based on mid. Therefore, although this invention introduces multiple physical constraint steps, it does not significantly increase control complexity. On the contrary, by avoiding multiple rounds of re-irrigation of the reconstructed image, the overall process is more direct.

[0083] XV. Explanation of Optional Variations 1. Regarding the binning degradation operator Bb: In some devices, binning is analog charge combining; in others, binning is digital averaging or summing. In this invention, Bb can be implemented using a summing model, an averaging model, or a weighted convolution model, depending on the specific device. As long as the operator reflects the relationship between the actual low-resolution readout and the high-resolution pixels, it can be used in this invention.

[0084] 2. Regarding the MTF model: The MTF can be obtained by looking up a table based on measured curves or by using analytical approximation, such as a comprehensive model obtained by convolving the pixel aperture function, scintillator spread function, and focus blur function. For this invention, the key is that the MTF is bound to the current binning mode and device state, and plays a role in limiting the enhancement range in frequency domain compensation.

[0085] 3. Regarding network structure: Network 100 can employ a convolutional neural network, a Transformer network, or a hybrid network of convolutional and Transformer networks. The fusion of the current viewpoint and neighboring viewpoints can also employ attention mechanisms, gated recurrent units, or cross-view correlation matching. This invention can be implemented as long as the network can receive protocol conditions in the projection domain and output a high-resolution projection.

[0086] 4. Regarding the reconstruction algorithm: Although FDK or fan-beam filtering back projection algorithms are preferred, when higher quantitative accuracy is required, the target high-resolution projection output by this invention can also be used as the initial projection input for the iterative reconstruction algorithm. This should be understood as follows: the core of this invention does not lie in prohibiting any subsequent iterative reconstruction, but rather in the fact that the super-resolution main link does not rely on the multi-round enhancement mode of "reconstructed image feedback to the previous super-resolution network," and that super-resolution, data consistency, and MTF constraints have formed a complete closed loop in the projection domain.

[0087] 5. Regarding clinical output: This invention can output two-dimensional slices, three-dimensional volume renderings, navigation overlay maps, or volume data for subsequent segmentation and measurement. The display method may differ for different scenarios such as orthopedics, dentistry, interventional radiotherapy, and radiotherapy, but this does not change the essence of the invention.

[0088] XVI. Feasibility and Clinical Suitability Explanation From the perspective of the implementation conditions of this invention, flat-panel CT equipment typically already possesses the hardware foundation such as a source, detector, gantry, and acquisition controller, as well as the software conditions for performing conventional FDK reconstruction. The key additions of this invention are: first, the complete transmission of protocol parameters to the reconstruction end; second, the deployment of a projection domain conditional super-resolution model in the reconstruction server 30; third, the addition of data consistency correction corresponding to the binning mode; and fourth, the addition of restricted frequency domain compensation corresponding to the MTF model. All of the above-mentioned new functions can be implemented through software upgrades and a small amount of calibration data expansion, without requiring replacement of the host hardware, thus possessing good engineering feasibility.

[0089] From a clinical suitability perspective, this invention is particularly suitable for the following scenarios: Firstly, in scenarios where scanning time must be shortened, such as intraoperative navigation, interventional re-scanning, and rapid scanning of patients with poor respiratory coordination; Secondly, in situations where the dosage must be reduced, such as in children, patients requiring repeated follow-up, and patients requiring multiple intraoperative scans; Third, scenarios sensitive to marginal structures, such as cortical bone, root apex of teeth, microwires, scaffolds, screws, and observation of fine fracture lines; Fourth, scenarios where different scanning protocols are frequently switched, such as performing both routine high-resolution examinations and low-dose rapid examinations on the same platform.

[0090] Because this invention uses conditional vectors to explicitly represent scanning protocols, the same model can switch between multiple protocols without having to design completely different enhancement modules for each protocol.

[0091] In summary, this invention provides an implementable, interpretable, and clinically deployable technical solution for image reconstruction of flat panel CT under low-dose, large-binning, and sparse-angle conditions through a combination of "projection domain conditional super-resolution network + data consistency correction strongly bound to binning + restricted frequency domain compensation strongly bound to MTF + cross-view fusion constraint strongly bound to angle sampling".

Claims

1. A method for reconstructing flat-panel CT images, characterized in that, Includes the following steps: Acquire measured projection data collected by a flat panel CT system under a preset scanning protocol, wherein the preset scanning protocol includes at least binning mode, angle sampling scheme and system geometric parameters; The measured projection data is subjected to dark field correction, flat field correction, bad pixel correction, logarithmic transformation and geometric normalization to obtain a low-resolution projection sequence; Based on the binning mode, angle sampling scheme and system geometric parameters, a conditional vector is constructed, and the low-resolution projection of the current angle, the projection of the acquired angle adjacent to the current angle and the conditional vector are input into a pre-trained projection domain conditional super-resolution network to obtain the first high-resolution projection. Data consistency correction is performed on the first high-resolution projection based on the binning degradation operator corresponding to the binning mode, so that the pseudo low-resolution projection processed by the binning degradation operator matches the low-resolution projection of the current angle, and the fusion range and fusion weight of adjacent angle features are constrained based on the angle sampling scheme to obtain the second high-resolution projection. Based on the modulation transfer function (MTF) model of the flat panel detector in the binning mode, frequency domain compensation and noise suppression coupling correction are performed on the second high-resolution projection to obtain the target high-resolution projection sequence; according to the target high-resolution projection sequence and the system geometric parameters, cone-beam reconstruction or fan-beam reconstruction is performed to output the flat panel CT reconstruction image.

2. The flat panel CT image reconstruction method of claim 1, wherein, The condition vector includes at least: binning factor, detector pixel pitch, distance from X-ray source to flat panel detector, distance from X-ray source to rotation center, current projection angle index, adjacent sampling angle interval, X-ray tube voltage, X-ray tube current, exposure time, and MTF identifier parameter corresponding to the flat panel detector.

3. The flat panel CT image reconstruction method of claim 2, wherein, The projection domain conditional super-resolution network includes a current view coding branch, a neighboring view coding branch, a conditional embedding branch, a feature fusion branch, and a residual decoding branch. The conditional features output by the conditional embedding branch are injected into the current view coding branch and the neighboring view coding branch through a feature linear modulation layer, a conditional normalization layer, or an attention gating layer to output the first high-resolution projection.

4. The flat panel CT image reconstruction method of claim 3, wherein, The data consistency correction includes: The first high-resolution projection is degraded using the same binning window and aggregation weights as during acquisition to obtain a pseudo-low-resolution projection. Calculate the projection residual between the pseudo-low-resolution projection and the measured low-resolution projection; The projection residual is written back to the high-resolution grid according to the inverse mapping relationship of the binning window, and the first high-resolution projection is corrected accordingly to obtain the second high-resolution projection.

5. The flat panel CT image reconstruction method of claim 4, wherein, The constraint on the fusion range and fusion weight of adjacent angle features based on the angle sampling scheme includes: A view reachable mask is constructed based on the angle difference between the current projection angle and the adjacent acquired projection angles; Block adjacent viewpoint features that exceed the preset angle window; The fusion weights of unmasked adjacent viewpoint features are determined based on the angular difference and the system's geometric magnification.

6. The flat panel CT image reconstruction method of claim 5, wherein, The frequency domain compensation and noise suppression coupling correction based on the modulation transfer function (MTF) model includes: Perform frequency domain transformation on the second high-resolution projection; Construct a target compensation filter based on the MTF curve corresponding to the current binning mode; The high-frequency components are enhanced in a restricted manner using the target compensation filter and the noise power spectrum suppression window function; and the enhanced frequency domain projection is subjected to an inverse transformation to obtain the target high-resolution projection sequence.

7. The flat panel CT image reconstruction method of claim 3, wherein, The projection domain conditional super-resolution network is trained in the following manner: Acquire high-resolution true projections of flat-panel CT systems in small binning mode and high angle sampling density; Binning degradation, angle undersampling degradation, noise injection, and MTF blurring are applied to the high-resolution ground truth projection to obtain the training input; The parameters of the projection domain conditional super-resolution network are optimized using the weighted sum of binning consistency loss, frequency domain MTF loss, projection gradient loss, and reconstruction domain structural similarity loss as the objective function.

8. A flat-panel CT image reconstruction system, characterized in that, include: The scanning control module is used to control the X-ray source, flat panel detector and rotating mechanism to acquire measured projection data according to the preset scanning protocol, and record the binning mode, angle sampling scheme and system geometric parameters; The preprocessing module is used to perform dark field correction, flat field correction, bad pixel correction, logarithmic transformation and geometric normalization on the measured projection data to obtain a low-resolution projection sequence. The condition construction module is used to construct a condition vector based on the binning mode, angle sampling scheme, and system geometric parameters; The projection domain conditional super-resolution module is used to input the low-resolution projection of the current angle, the projections of adjacent acquired angles, and the conditional vector into a pre-trained projection domain conditional super-resolution network to obtain a first high-resolution projection. The data consistency correction module is used to perform data consistency correction on the first high-resolution projection according to the binning degradation operator corresponding to the binning mode and the angle sampling scheme to obtain the second high-resolution projection. The MTF constraint module is used to perform frequency domain compensation and noise suppression coupled correction on the second high-resolution projection based on the modulation transfer function (MTF) model of the flat panel detector in the binning mode, so as to obtain the target high-resolution projection sequence. The reconstruction module is used to perform cone-beam reconstruction or fan-beam reconstruction based on the target high-resolution projection sequence and the system geometric parameters, and output a flat-panel CT reconstruction image.

9. The flat-panel CT image reconstruction system according to claim 8, characterized in that, The X-ray source and the flat panel detector are respectively located at opposite ends of the rotating gantry or C-arm and are positioned relative to each other around the rotation center of the subject; the bed passes through the rotation center; the flat panel detector is connected to the acquisition controller via the detector controller, the acquisition controller is communicatively connected to the X-ray source controller, the rotation drive mechanism and the reconstruction server, and the reconstruction server is connected to the clinical workstation.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, When the computer program is executed by the processor, it implements the flat panel CT image reconstruction method according to any one of claims 1 to 7.

Images