DR image quality control method and system based on decoupling representation learning and multi-level feedback

Abstract

This invention provides a method and system for DR image quality control based on decoupled representation learning and multi-level feedback, belonging to the field of medical image processing technology. The method includes: acquiring DR images and DICOM metadata; automatically classifying areas if the examination site is missing; extracting mixed features using an adaptive anatomical perception network; obtaining anatomical, quality, and pathological features through orthogonal decoupling; guiding region-specific quality assessment based on anatomical features and inversely generating suggestions for adjusting imaging parameters. This invention achieves complete separation of lesion and quality information through orthogonal feature decoupling, avoiding interference from pathological features in quality assessment, while providing actionable parameter feedback, significantly improving the accuracy, robustness, and clinical applicability of DR image quality control.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, and in particular to a method and system for quality control of DR images based on decoupled representation learning and multi-level feedback. Background Technology

[0002] Digital radiography (DR) is the most fundamental imaging examination method in clinical radiology, and its image quality directly affects the accuracy of doctors' diagnoses and the effectiveness of subsequent AI-assisted diagnostic systems. Currently, clinical practice mainly relies on manual assessment of image quality based on experience or physical indicators (such as exposure index (EI) and signal-to-noise ratio (SNR). However, manual quality control suffers from problems such as strong subjectivity, inconsistent standards, and significant lag. In recent years, automated quality control methods based on deep learning have made some progress, but most existing technologies use global feature extraction, resulting in a high degree of coupling between pathological features (such as lesions) and imaging quality defects (such as noise and blurring) in the feature space. The model struggles to distinguish between "lesions" and "artifacts." For example, areas of consolidation in pneumonia are easily misjudged as underexposed, and small fracture lines are easily misjudged as noise, leading to false positives and seriously affecting the clinical usability of the system. In addition, existing models lack the ability to adapt to anatomical sites, and their feedback mechanisms are simplistic, making it difficult to output actionable suggestions for adjusting imaging parameters. Summary of the Invention

[0003] To address the technical problem of effectively decoupling pathological features from imaging quality defects in the aforementioned background technologies, this invention provides a DR image quality control method and system based on decoupling representation learning and multi-level feedback. By performing orthogonal decoupling processing of features, the mixed feature map is mapped to three orthogonal semantic subspaces to obtain anatomical features, quality features, and pathological features, respectively. This achieves complete separation of lesion information and quality information, avoiding interference from pathological features on quality evaluation.

[0004] To achieve the above objectives, the first aspect of the present invention provides a DR image quality control method based on decoupled representation learning and multi-level feedback, comprising: Acquire the DR image of the object to be inspected and its DICOM metadata; the DICOM metadata includes the inspection area and the current shooting parameters. If the examination area is missing, automatic area classification is performed based on the shallow features of the DR image to obtain the area classification result; The DR image is feature extracted using an adaptive anatomical perception feature extraction network. Based on the examination site or the site classification result, the site-specific weights in the network are dynamically activated to obtain a hybrid feature map containing anatomical, quality, and pathological information. The hybrid feature map containing anatomical, quality, and pathological information is subjected to feature orthogonal decoupling processing. Through cross-attention mechanism and orthogonal constraints, the hybrid feature map is mapped to three orthogonal semantic subspaces to obtain anatomical features, quality features, and pathological features, respectively. Based on the anatomical features, the quality features are used to perform region-specific quality assessment to obtain quality assessment results; the quality features and the current shooting parameters are input into an inverse parameter compensation network to generate shooting parameter adjustment suggestions; a quality control report is generated based on the quality assessment results and shooting parameter adjustment suggestions.

[0005] Furthermore, the adaptive anatomical perception feature extraction network is constructed based on the Swin-Transformer V2 backbone network and includes multiple cascaded feature extraction stages. The adaptive anatomical perception feature extraction network first extracts shallow features. If the examined area is missing, the automatic area classification is performed based on the shallow features of the DR image. Then, according to the examined area or the area classification result, a pre-trained area-specific LoRA adapter is dynamically activated, and the corresponding LoRA adapter is injected into the feature extraction stage at the corresponding depth to enhance the extraction of features in the corresponding frequency band. The enhanced features of each stage are fused to obtain the hybrid feature map. Among them, the high-frequency adapter is injected into the shallow high-frequency stage to enhance high-frequency features, the mid-frequency adapter is injected into the mid-medium-frequency stage to enhance mid-frequency features, and the low-frequency adapter is injected into the deep low-frequency stage to enhance low-frequency features.

[0006] Furthermore, the feature orthogonal decoupling process employs a decoupling cross-attention mechanism, utilizing three independent learnable query vector groups to perform cross-attention calculations with the hybrid feature map, and constraining the three subspace features of anatomical features, quality features, and pathological features to be pairwise orthogonal through an orthogonal loss function; the orthogonal loss function is: ; in, , , These respectively represent anatomical features, quality features, and pathological features.

[0007] Furthermore, the loss function used in the feature orthogonal decoupling process during the training phase includes the main task loss, consistency loss, and orthogonal loss; the consistency loss is the regression loss of the quality assessment task or the segmentation loss of the dissection and segmentation task. The consistency loss requires that the quality features extracted from the baseline image and the simulated lesion image be consistent, and its calculation formula is as follows: ; in, As the reference image, To simulate lesion images, GAP is a global average pooling operation.

[0008] Furthermore, the region-specific quality assessment includes: generating an anatomical mask using the anatomical features, and applying different evaluation operators to different regions of the quality features under the guidance of the anatomical mask; For the limbs / joints region, the edge spread function and modulation transfer function are calculated to evaluate image sharpness; For the chest / abdomen / pelvis region, grayscale cutoff rate and texture fractal dimension are calculated to evaluate contrast and penetration. For the spine region, the contrast-to-noise ratio is calculated to evaluate image quality.

[0009] Furthermore, the inverse parameter compensation network is a multilayer perceptron, which takes the quality features and the current shooting parameters as inputs and outputs the adjustment amount of the shooting parameters; the shooting parameters include tube voltage kVp, tube current-time product mAs and exposure time.

[0010] A second aspect of the present invention provides a DR image quality control system based on decoupled representation learning and multi-level feedback, comprising: The image acquisition module is used to acquire DR images of the object to be inspected and its DICOM metadata; the DICOM metadata includes the inspection area and the current shooting parameters; The part classification module is used to automatically classify the part based on the shallow features of the DR image if the examined part is missing, and obtain the part classification result. An adaptive anatomical perception feature extraction module is used to extract features from the DR image using an adaptive anatomical perception feature extraction network. Based on the examination site or the site classification result, the site-specific weights in the network are dynamically activated to obtain a hybrid feature map containing anatomical, quality, and pathological information. The feature orthogonal decoupling module is used to perform feature orthogonal decoupling processing on the hybrid feature map containing anatomical, quality, and pathological information. Through the cross attention mechanism and orthogonal constraints, the hybrid feature map is mapped to three orthogonal semantic subspaces to obtain anatomical features, quality features, and pathological features, respectively. The quality control result generation module is used to perform region-specific quality evaluation based on the anatomical features to obtain quality evaluation results; input the quality features and the current imaging parameters into the inverse parameter compensation network to generate imaging parameter adjustment suggestions; and generate a quality control report based on the quality evaluation results and imaging parameter adjustment suggestions.

[0011] A third aspect of the present invention provides an electronic device including a memory, a processor, and a program stored in the memory and running on the processor, wherein the processor executes the program to implement the steps of the DR image quality control method based on decoupled representation learning and multi-level feedback as described in the first aspect of the present invention.

[0012] A fourth aspect of the present invention provides a computer-readable storage medium having a program stored thereon that, when executed by a processor, implements the steps of the DR image quality control method based on decoupled representation learning and multi-level feedback as described in the first aspect of the present invention.

[0013] A fifth aspect of the present invention provides a computer program product comprising software code, wherein the program in the software code performs the steps of the DR image quality control method based on decoupled representation learning and multi-level feedback as described in the first aspect of the present invention.

[0014] Compared with existing technologies, the DR image quality control method and system based on decoupled representation learning and multi-level feedback provided by this invention has the following beneficial effects: (1) The orthogonal decoupling processing provided by the present invention maps the hybrid feature map to three orthogonal semantic subspaces through cross attention mechanism and orthogonal constraints, so that the anatomical, quality and pathological features are completely separated, and the lesion information and quality information are decoupled, avoiding the interference of lesions such as pneumonia consolidation and fracture on quality evaluation, and significantly reducing the false positive rate.

[0015] (2) Based on the adaptive anatomical perception feature extraction network, this invention strengthens different frequency band features (high frequency / medium frequency / low frequency) for different anatomical parts by dynamically activating the site-specific LoRA adapter, thereby achieving unified quality control for all parts of the body without the need to train a separate model for each part, thus reducing deployment costs.

[0016] (3) Based on the characteristics of the inverse parameter compensation network technology, this invention maps the quality features and the current shooting parameters to specific parameter adjustment amounts (ΔkVp, ΔmAs, etc.), realizing the inverse mapping from image features to physical shooting parameters, providing physicians with executable closed-loop guidance, and improving clinical applicability and teaching value. Attached Figure Description

[0017] The accompanying drawings, which form part of this disclosure, are used to provide a further understanding of this disclosure. The illustrative embodiments of this disclosure and their descriptions are used to explain this disclosure and do not constitute an undue limitation of this disclosure.

[0018] Figure 1 The flowchart shows the DR image quality control method based on decoupled representation learning and multi-level feedback provided in Embodiment 1 of the present invention. Figure 2This is a schematic diagram of the adaptive anatomical perception feature extraction network provided in Embodiment 1 of the present invention; Figure 3 This is a schematic diagram of the feature decoupling process provided in Embodiment 1 of the present invention; Figure 4 This is an architecture diagram of the DR image quality control system based on decoupled representation learning and multi-level feedback provided in Embodiment 2 of the present invention. Detailed Implementation

[0019] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of the invention. Unless otherwise specified, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0020] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, unless the context clearly indicates otherwise, the singular form is intended to include the plural form as well. Furthermore, it should be understood that the terms “comprising” and “having”, and any variations thereof, are intended to cover non-exclusive inclusion, for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0021] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0022] All data acquisition in this embodiment is carried out in accordance with laws and regulations and with user consent, and the data is used legally.

[0023] Example 1 like Figure 1 This embodiment provides a DR image quality control method based on decoupled representation learning and multi-level feedback, including: S1. Obtain the DR image of the object to be inspected and its DICOM metadata; the DICOM metadata includes the inspection area and the current shooting parameters.

[0024] Specifically, DR images and corresponding metadata are acquired from a PACS system or DR acquisition workstation via the DICOM protocol. The current imaging parameters include tube voltage (kVp), tube current-time product (mAs), and exposure time. The examination site is used to acquire preset anatomical information; parameters such as kVp and mAs are used for calculating subsequent parameter adjustment suggestions; and the DR images are used for the extraction and analysis of quality characteristics, anatomical features, and pathological features.

[0025] After acquisition, the DR images are preprocessed: uniformly resampled to a standard resolution (e.g., 1024×1024 pixels) and bit depth normalized to map pixel values ​​to the [0, 1] interval to eliminate the influence of differences in imaging resolution and bit depth between different devices.

[0026] S2. If the examined area is missing, automatic area classification is performed based on the shallow features of the DR image to obtain the area classification result.

[0027] Specifically, the Body Part Examined label in the DICOM metadata is checked. If this label is missing, a fallback mechanism is triggered: a lightweight classifier, built either based on ResNet-18 or directly from shallow features of the Swin-Transformer V2 backbone network followed by global pooling and fully connected layers, is used to classify the preprocessed image to determine anatomical locations (e.g., feet, chest, lumbar spine). This classification result will serve as the anatomical prior for subsequent adaptive feature extraction.

[0028] Based on the acquired anatomical location information, the system executes intelligent policy routing, matches the corresponding feature enhancement strategy, and prepares for injecting specific LoRA weights into the subsequent network.

[0029] S3. Use an adaptive anatomical perception feature extraction network to extract features from the DR image, and dynamically activate the location-specific weights in the network based on the examination site or the location classification result to obtain a hybrid feature map containing anatomical, quality, and pathological information.

[0030] like Figure 2 As shown, the adaptive anatomical perception feature extraction network is constructed based on the Swin-Transformer V2 backbone network and includes multiple cascaded feature extraction stages for hierarchical extraction of multi-scale features from images. To address the positional bias problem in large-size medical images, Log-CPB (Log-spaced Continuous Relative Position Bias) positional encoding is introduced into the network to improve the feature extraction accuracy and spatial modeling capability of large-size DR images.

[0031] The network hierarchy is designed as follows: Stage 1 (Patch Embedding): Downsample the input image (e.g., 1024×1024) by 4 times and extract the basic visual token.

[0032] Stage 2-4 (Swin Blocks): Log-CPB position coding is used to extract features from different frequency bands to solve the positional deviation problem in large-size medical images.

[0033] Stage 2 (L2 layer): Extracts high-frequency features such as subtle noise and trabecular texture; Stage 3 (L3 layer): Identifies mid-frequency features such as organ contours and overlapping regions; Stage 4 (L4 layer): Modeling low-frequency features such as global anatomical layout; The core of this network lies in a universal LoRA dynamic adaptation mechanism across the entire anatomical domain. Based on the anatomical prior obtained in step S2, the system first determines the strategy: if the location is a limb / joint (e.g., foot, hand), it is determined as strategy A; if it is a chest / abdomen / pelvis, it is determined as strategy B; if it is a spine (e.g., lumbar spine), it is determined as strategy C. Then, the pre-trained location-specific LoRA (low-rank adaptation) adapter is dynamically activated and injected into the feature extraction stage at the corresponding depth of the network to enhance the extraction of features in the corresponding frequency band. The specific strategy is as follows: High-frequency adapter (strategy A): When the anatomical location is the limbs / joints (such as the feet or hands), the high-frequency adapter is activated and injected into the superficial high-frequency stage (such as Stage 2) to enhance the extraction of high-frequency features such as bone trabeculae, fine fracture lines, and edge sharpness, thus obtaining L2 enhanced features.

[0034] Intermediate frequency adapter (strategy B): When the anatomical location is the chest / abdomen / pelvis, the intermediate frequency adapter is activated and injected into the intermediate frequency stage (such as Stage 3) to enhance the extraction of intermediate frequency features such as lung texture and organ contour, resulting in L3 enhanced features.

[0035] Low-frequency adapter (strategy C): When the anatomical location is the spine (such as the lumbar spine), the low-frequency adapter is activated and injected into the deep low-frequency stage (such as Stage 4) to enhance the extraction of low-frequency features such as overall contrast and penetration, resulting in L4 enhanced features.

[0036] Through the aforementioned layered enhancements, the LoRA adapter abandons the "one-size-fits-all" adaptation model, precisely intervening in feature layers of different network depths. It achieves targeted feature enhancement through residual fusion, ultimately outputting a hybrid feature map containing anatomical, quality, and pathological information. (Dimensions, for example, 1024×1024×512).

[0037] S4. Perform orthogonal decoupling processing on the hybrid feature map containing anatomical, quality, and pathological information. Map the hybrid feature map to three orthogonal semantic subspaces through a cross-attention mechanism and orthogonal constraints to obtain anatomical features. Quality characteristics and pathological features .

[0038] like Figure 3 As shown, the feature orthogonal decoupling process employs a decoupled cross-attention (DCA) mechanism, which is implemented through three parallel branches: pathological branch DCA-P, quality branch DCA-Q, and anatomical branch DCA-A.

[0039] The specific implementation details are as follows: (1) Query generation: The system maintains three learnable basis vectors (Learnable Queries). , respectively representing the a priori prototypes of anatomy, mass, and pathology.

[0040] (2) Attention Calculation: Using these three independent and orthogonal sets of learnable query vectors, attention is calculated for... Perform cross-attention calculations: ; in, Key (K) and Value (V) are both derived from The mapping is obtained. Through three parallel cross-attention branches, relevant components in the mixed features are "absorbed" into the corresponding subspaces, and pathological features, quality features, and anatomical features are output respectively.

[0041] (3) Example of effect: Even if the input image contains a large area of ​​pneumonia solidity, due to the orthogonal constraints during the training phase, The response of the corresponding real region in the vector is suppressed, and only the signal-to-noise ratio features of the image itself are preserved.

[0042] To ensure thorough decoupling, a training paradigm of "self-supervised + synthetic data" was adopted during the training phase, using baseline DR images. (Normal), add simulated lesions to obtain Adding physical artifacts to obtain And a double loss function is introduced for constraint: (1) Consistency loss Forced reference imagery (Normal images) and simulated lesion images Quality features extracted from (artificially added lesions on baseline images) Complete consistency is ensured, guaranteeing that lesion information does not interfere with quality characteristics, and completely eliminating the interference of pathological features such as "white lung" and "bone fracture" on quality evaluation. The calculation formula is as follows: ; in, As the reference image, To simulate lesion images, GAP is a global average pooling operation.

[0043] (2) Orthogonal loss Forced anatomical features Quality characteristics and pathological features The three subspace features are pairwise orthogonal, and their dot product approaches zero, ensuring that they are independent and uncoupled in the semantic space, thus achieving precise decoupling of features. The orthogonal loss function is: ; Therefore, the overall loss function used by the feature orthogonal decoupling module during the training phase is a combination of the main task loss and the regularization constraint term: ; in, The main task loss can be the regression loss for a quality scoring task or the segmentation loss for an anatomical structure. , Hyperparameters are used to balance the weights of each loss.

[0044] S5. Based on the anatomical features, guide the quality features to perform region-specific quality assessment and obtain quality assessment results; input the quality features and the current shooting parameters into the inverse parameter compensation network to generate shooting parameter adjustment suggestions; generate a quality control report based on the quality assessment results and shooting parameter adjustment suggestions.

[0045] This step further includes the following sub-steps: S5.1, Region-Specific Quality Assessment: Utilizing a decoder to assess anatomical features Recover the anatomical mask to guide quality features The system inputs a region-specific evaluation head and activates differential evaluation operators based on anatomical partitions. The generated mask, in Different evaluation operators are invoked within the space, as shown in the table below:

[0046] Based on this, the evaluation of different regions is further refined: Strategy A (Chest / Abdomen / Pelvis): In the lung field or mediastinal region, calculate the gray-level cutoff rate to evaluate penetration, and calculate the texture fractal dimension (FD) to evaluate low-contrast resolution. For example, if the FD in the lung field region is >1.8 but the gray-level histogram peak is skewed to the left (<50), it can be judged as "good contrast but underexposure"; calculate the gray-level cutoff rate in the mediastinal / posterior cardiac region to evaluate penetration.

[0047] Strategy B (Limbs / Joints): Within the skeletal mask area, extract the edges of the trabecular bone and calculate the Edge Spread Function (ESF) and Modulation Transfer Function (MTF). If the MTF value at 5 lp / mm decreases by more than 50%, it is judged as "motion blur"; if the ESF curve is flat, it is judged as insufficient sharpness.

[0048] Strategy C (Spine): Calculate the contrast-to-noise ratio (CNR) of adjacent vertebrae to detect the continuity of the vertebral sequence. If CNR < 3.0, it is judged as insufficient penetration or excessive noise.

[0049] S5.2 Intelligent parameter feedback generation: This involves generating the quality characteristics... and the current shooting parameters (kVp, mAs, Time) are input to the inverse parameter compensation network (PRN) to generate shooting parameter adjustment suggestions. This network is a 3-layer multilayer perceptron (MLP), and its model logic is a regression mapping function: .

[0050] The input consists of a quality defect feature vector (such as a "particle noise feature vector") and the current physical parameters, and the output consists of the specific parameter adjustment amount.

[0051] Example: If the current parameters are kVp=80, mAs=20, and It contains high-intensity quantum noise features (photon starvation), and PRN can output regression values ​​ΔkVp=+5, ΔmAs=+10, ΔTime=0.

[0052] Suggested generation: Mapped to text: "It is recommended to increase kVp by 5kV and increase mAs by 10mAs", and can be combined with rules to generate text prompts such as "It is recommended to use a filter grid in conjunction with this".

[0053] Principle: Simulating the experience of a radiology technician, when a "insufficient penetration" characteristic is detected, the PRN outputs a positive ΔkVp compensation suggestion.

[0054] S5.3 Quality Control Report Generation: Based on the quality evaluation results, the suggested adjustments to the imaging parameters, and optionally utilizing pathological features... Potential lesions are annotated to generate a complete structured quality control report.

[0055] In practical applications, a specific training example of the above network model is as follows: (1) Data collection: Collect more than 10,000 DR images from multiple devices and locations, and construct a triplet training set containing the original image, quality label, and suggested parameters.

[0056] (2) Phased training: Decoupling the pre-training phase: Freeze the backbone network parameters and train only the FDM module. Optimize the orthogonal loss using synthetic data (artificially superimposed lesions and noise). and consistency loss The learning rate is set to 5e-5, and the training run is repeated for 50 epochs.

[0057] Evaluation head fine-tuning phase: Unfreeze the shallow parts of the backbone network and jointly train the differential evaluation branch. Set the learning rate to 3e-5 and train for 30 epochs.

[0058] Feedback-regression training phase: Fine-tune the regression head of the PRN network. Set the learning rate to 1e-5 and train for 20 epochs.

[0059] Example 2 like Figure 4 As shown, this embodiment provides a DR image quality control system based on decoupled representation learning and multi-level feedback, including: The image acquisition module is used to acquire DR images of the object to be inspected and its DICOM metadata; the DICOM metadata includes the inspection area and the current shooting parameters; The part classification module is used to automatically classify the part based on the shallow features of the DR image if the examined part is missing, and obtain the part classification result. An adaptive anatomical perception feature extraction module is used to extract features from the DR image using an adaptive anatomical perception feature extraction network. Based on the examination site or the site classification result, the site-specific weights in the network are dynamically activated to obtain a hybrid feature map containing anatomical, quality, and pathological information. The feature orthogonal decoupling module is used to perform feature orthogonal decoupling processing on the hybrid feature map containing anatomical, quality, and pathological information. Through the cross attention mechanism and orthogonal constraints, the hybrid feature map is mapped to three orthogonal semantic subspaces to obtain anatomical features, quality features, and pathological features, respectively. The quality control result generation module is used to perform region-specific quality evaluation based on the anatomical features to obtain quality evaluation results; input the quality features and the current imaging parameters into the inverse parameter compensation network to generate imaging parameter adjustment suggestions; and generate a quality control report based on the quality evaluation results and imaging parameter adjustment suggestions.

[0060] Example 3 Embodiment 3 of the present invention provides an electronic device.

[0061] An electronic device includes a memory, a processor, and a program stored in the memory and running on the processor. The processor includes, but is not limited to, at least one of a central processing unit (CPU), a graphics processing unit (GPU), a neural network processor (NPU), a tensor processor (TPU), or an artificial intelligence acceleration chip. The program is used to implement the steps in the DR image quality control method based on decoupled representation learning and multi-level feedback as described in Embodiment 1 of the present invention when executing the program.

[0062] The detailed steps are the same as those of the DR image quality control method based on decoupled representation learning and multi-level feedback provided in Example 1, and will not be repeated here.

[0063] Example 4 Embodiment 4 of the present invention provides a computer-readable storage medium.

[0064] A computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements the steps of the DR image quality control method based on decoupled representation learning and multi-level feedback as described in Embodiment 1 of the present invention.

[0065] The detailed steps are the same as those of the DR image quality control method based on decoupled representation learning and multi-level feedback provided in Example 1, and will not be repeated here.

[0066] Example 5 Embodiment 5 of the present invention provides a computer program product.

[0067] A computer program product includes software code, wherein the program in the software code performs the steps of the DR image quality control method based on decoupled representation learning and multi-level feedback as described in Embodiment 1 of the present invention.

[0068] The detailed steps are the same as those of the DR image quality control method based on decoupled representation learning and multi-level feedback provided in Example 1, and will not be repeated here.

[0069] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The solutions in the embodiments of the present invention can be implemented using various computer languages. For example, in one implementation, the methods and systems can be developed based on deep learning frameworks (such as TensorFlow, PyTorch, etc.) and using the Python language. Those skilled in the art will understand that other suitable programming languages ​​or tools can also be used for implementation without departing from the core ideas of the present invention.

[0070] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0071] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0072] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0073] The above description is merely a preferred embodiment of this practice and is not intended to limit the scope of this practice. Various modifications and variations can be made to this practice by those skilled in the art. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of this practice should be included within the protection scope of this practice.

Claims

1. A DR image quality control method based on decoupled representation learning and multi-level feedback, characterized in that, include: Acquire the DR image of the object to be inspected and its DICOM metadata; the DICOM metadata includes the inspection area and the current shooting parameters. If the examined area is missing, automatic area classification is performed based on the shallow features of the DR image to obtain the area classification result; The DR image is feature extracted using an adaptive anatomical perception feature extraction network. Based on the examination site or the site classification result, the site-specific weights in the network are dynamically activated to obtain a hybrid feature map containing anatomical, quality, and pathological information. The hybrid feature map containing anatomical, quality, and pathological information is subjected to feature orthogonal decoupling processing. Through cross-attention mechanism and orthogonal constraints, the hybrid feature map is mapped to three orthogonal semantic subspaces to obtain anatomical features, quality features, and pathological features, respectively. Based on the anatomical features, the quality features are used to perform region-specific quality assessment to obtain quality assessment results; the quality features and the current imaging parameters are input into an inverse parameter compensation network to generate imaging parameter adjustment suggestions. A quality control report is generated based on the quality evaluation results and the shooting parameter adjustment suggestions.

2. The method as described in claim 1, characterized in that, The adaptive anatomical perception feature extraction network is constructed based on the Swin-Transformer V2 backbone network and includes multiple cascaded feature extraction stages. The adaptive anatomical perception feature extraction network first extracts shallow features. If the examined area is missing, the network performs automatic area classification based on the shallow features of the DR image. Then, according to the examined area or the area classification result, the network dynamically activates a pre-trained area-specific LoRA adapter and injects the corresponding LoRA adapter into the feature extraction stage at the corresponding depth to enhance the extraction of features in the corresponding frequency band. The enhanced features from each stage are then fused to obtain the hybrid feature map. Specifically, the high-frequency adapter is injected into the shallow high-frequency stage to enhance high-frequency features, the mid-frequency adapter is injected into the mid-medium-frequency stage to enhance mid-frequency features, and the low-frequency adapter is injected into the deep low-frequency stage to enhance low-frequency features.

3. The method as described in claim 1, characterized in that, The orthogonal decoupling process employs a decoupling cross-attention mechanism, utilizing three independent learnable query vector groups to perform cross-attention calculations with the hybrid feature map, and constraining the three subspace features of anatomical features, quality features, and pathological features to be pairwise orthogonal through an orthogonal loss function; the orthogonal loss function is: ； in, , , These respectively represent anatomical features, quality features, and pathological features.

4. The method as described in claim 1, characterized in that, The loss functions used in the orthogonal decoupling process during the training phase include the main task loss, consistency loss, and orthogonality loss; the consistency loss is either the regression loss for the quality assessment task or the segmentation loss for the dissection and segmentation task. The consistency loss requires that the quality features extracted from the baseline image and the simulated lesion image be consistent, and its calculation formula is as follows: ； in, As the reference image, To simulate lesion images, GAP is a global average pooling operation.

5. The method as described in claim 1, characterized in that, The region-specific quality assessment includes: generating an anatomical mask using the anatomical features, and applying different evaluation operators to different regions of the quality features under the guidance of the anatomical mask; For the limbs / joints region, the edge spread function and modulation transfer function are calculated to evaluate image sharpness; For the chest / abdomen / pelvis region, grayscale cutoff rate and texture fractal dimension are calculated to evaluate contrast and penetration. For the spine region, the contrast-to-noise ratio is calculated to evaluate image quality.

6. The method as described in claim 1, characterized in that, The inverse parameter compensation network is a multilayer sensor, which takes the quality characteristics and the current shooting parameters as inputs and outputs the adjustment amount of the shooting parameters; the shooting parameters include tube voltage kVp, tube current-time product mAs and exposure time.

7. A DR image quality control system based on decoupled representation learning and multi-level feedback, characterized in that, include: The image acquisition module is used to acquire DR images of the object to be inspected and its DICOM metadata; the DICOM metadata includes the inspection area and the current shooting parameters; The part classification module is used to automatically classify the part based on the shallow features of the DR image if the examined part is missing, and obtain the part classification result. An adaptive anatomical perception feature extraction module is used to extract features from the DR image using an adaptive anatomical perception feature extraction network. Based on the examination site or the site classification result, the site-specific weights in the network are dynamically activated to obtain a hybrid feature map containing anatomical, quality, and pathological information. The feature orthogonal decoupling module is used to perform feature orthogonal decoupling processing on the hybrid feature map containing anatomical, quality, and pathological information. Through the cross attention mechanism and orthogonal constraints, the hybrid feature map is mapped to three orthogonal semantic subspaces to obtain anatomical features, quality features, and pathological features, respectively. The quality control result generation module is used to guide the quality features to perform region-specific quality evaluation based on the anatomical features, and obtain the quality evaluation results; the quality features and the current imaging parameters are input into the inverse parameter compensation network to generate imaging parameter adjustment suggestions; A quality control report is generated based on the quality evaluation results and the shooting parameter adjustment suggestions.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the program, it implements the steps of the DR image quality control method based on decoupled representation learning and multi-level feedback as described in any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps of the DR image quality control method based on decoupled representation learning and multi-level feedback as described in any one of claims 1 to 6.

10. A computer program product, comprising software code, characterized in that, The program in the software code performs the steps of the DR image quality control method based on decoupled representation learning and multi-level feedback as described in any one of claims 1 to 6.

Images