Digital x-ray radiography real-time image quality assessment control system, device and medium

Abstract

The application relates to the technical field of image quality evaluation, and discloses a digital X-ray radiography real-time image quality evaluation control system, equipment and a medium, which comprises a collection end used for collecting X-ray image data of a digital radiological imaging system (DROC) interface in real time, a service end provided with a storage module, a classification model, a segmentation model, a detection model and a quality control analysis module, an optimized classification model used for identifying a detection part of the X-ray image data transmitted by the collection end, an optimized segmentation model used for segmenting key anatomical parts of the detection part in the X-ray image data to obtain a segmentation result, an optimized detection model used for high-precision detection of a region of interest in the segmentation result to obtain a detection result of the detection part about a quality control index, a quality control analysis module used for quality control analysis of the detection result to obtain a quality control analysis result, and a client used for automatically generating a quality control report. The application improves the efficiency and accuracy of quality control.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image quality evaluation, in particular to a digital X-ray radiography real-time image quality evaluation control system, equipment and medium. BACKGROUND

[0002] Digital radiography (DR) has become a core imaging diagnostic method in modern medicine and is widely used in the examination of various parts of the body. Compared with traditional film imaging, DR has the advantages of fast acquisition, large amount of image information, and reusability, and is gradually popularized in large hospitals and medical institutions. However, due to the differences in positioning level and operation standardization of image technicians in each hospital, the quality of DR images is uneven, and the image examination results between hospitals are difficult to achieve mutual recognition. This mutual recognition problem is particularly significant when patients seek medical treatment across institutions, which leads to waste of medical resources and reduces the patient's medical experience.

[0003] To solve this problem, hospitals at all levels generally perform quality control on X-ray images to improve the consistency and quality of images. However, traditional image quality control methods mainly rely on manual interpretation, requiring technicians to subjectively evaluate image quality, positioning, and exposure parameters. This approach not only requires a lot of experience, but is also subject to the skill level of the operator, resulting in unstable quality control results. In addition, manual quality control is time-consuming and labor-intensive, making it difficult to handle large-scale image data and increasing the burden on medical resources.

[0004] In recent years, some quality control systems based on artificial intelligence have gradually entered this field. For example, some quality control systems have attempted to use deep learning models for image quality evaluation to automatically detect positioning-related indicators. However, these systems still have significant shortcomings in practical application, including low real-time and accuracy of quality evaluation, lack of comprehensiveness for different anatomical sites, limited evaluation indicators, and difficulty in meeting the standard requirements of clinical quality control. SUMMARY

[0005] The present application provides a digital X-ray radiography real-time image quality evaluation control system, which aims to realize real-time and accurate quality evaluation of X-ray films through automation technology and deep learning models, to solve the problems of low efficiency, unstable evaluation, and high cost of existing manual quality control.

[0006] The present application is achieved by the following technical solutions:

[0007] A digital X-ray radiography real-time image quality evaluation control system, comprising:

[0008] An acquisition end for real-time acquisition of X-ray image data of a digital radiography system (DROC) interface;

[0009] The server is provided with a storage module, a classification model, a segmentation model, a detection model and a quality control analysis module, the storage module stores sample images and expert consensus quality control standards, the classification model, the segmentation model and the detection model are respectively trained and optimized through the sample images, wherein the optimized classification model is used to identify the detection part of the X-ray image data transmitted by the acquisition end, the optimized segmentation model is used to segment the key anatomical part of the detection part in the X-ray image data to obtain a segmentation result; the optimized detection model is used to detect the region of interest in the segmentation result with high precision to obtain the detection result of the detection part on the quality control indicators; the quality control analysis module performs quality control analysis on the detection result based on the expert consensus quality control standard to obtain a quality control analysis result;

[0010] The client receives and parses the X-ray image data, the segmentation result, the detection result and the quality control analysis result to automatically generate a quality control report and display the segmentation result, the detection result and the analysis result in real time.

[0011] As an optimization, the classification model is a combined model of ResNetXt-50 model and OCR recognition model, wherein the specific process of training the classification model is as follows:

[0012] The sample image is input into the ResNetXt-50 model, the predicted label is output through the ResNetXt-50 model, and the text information in the sample image is identified as the actual label through the OCR recognition model, wherein the sample image comes from the X-ray image data of the DROC interface, and the sample image is labeled with the body part name corresponding to the X-ray image data;

[0013] The predicted label and the actual label are input into a preset loss function, and the parameters of the ResNet50-32x4d model are optimized according to the loss function.

[0014] As optimization, the segmentation model is an MSU-Unet model, the MSU-Unet model comprising an encoder and a decoder, wherein the encoder and the decoder each comprise a plurality of convolutional layers, and the convolutional layers in the encoder are symmetrically arranged with the convolutional layers in the decoder, and the outputs of the convolutional layers in the encoder are connected to the inputs of the symmetrically arranged convolutional layers in the decoder by a jump connection; each of the convolutional layers comprises a first convolutional group composed of two 3*3 convolutional first convolutional blocks arranged in series, a second convolutional group composed of two 7*7 convolutional second convolutional blocks arranged in series, and a 1*1 convolutional third convolutional block, the image features of the input of the convolutional layer are obtained as first intermediate features and second intermediate features by the first convolutional group and the second convolutional group respectively, and the first intermediate features and the second intermediate features are combined after channel splicing and then output image features are obtained by the third convolutional block.

[0015] As optimization, in the encoder, the image features output by the convolutional layers are down-sampled by a max-pooling operation.

[0016] As optimization, in the decoder, the image features output by the convolutional layers are up-sampled by deconvolution or bilinear.

[0017] As optimization, the sample pictures used for training the segmentation model and the classification model are subjected to micro data enhancement, and the operation process of the micro data enhancement specifically comprises one or more of the following operations: random brightness adjustment, random saturation adjustment, random contrast adjustment, random translation adjustment, and random cropping adjustment.

[0018] As optimization, the client comprises:

[0019] A receiving and storing module is configured to receive and store the X-ray image data, the segmentation result, the detection result, and the quality control analysis result transmitted by the server.

[0020] A cropping and displaying module is configured to automatically crop the area in the segmentation result that needs to be controlled according to the quality control analysis result and adaptively display the area, and if the analysis result is abnormal, an alarm is sent in the client to prompt an operator to timely adjust the shooting parameters.

[0021] A correction module is configured to manually switch the information of the detection part when it is detected that the detection part recognized by the server does not match the actual part, and feed back the modified information to the server for correction.

[0022] An index displaying module is configured to display all quality control indexes related to the detection part, and if a certain quality control index is abnormal, an abnormality prompt is given.

[0023] An automatic evaluation report generation module is configured to mark the detection result corresponding to the quality control index according to the quality control analysis result, mark the detection result as normal when the detection result is within the expert consensus quality control standard range, mark the detection result as abnormal when the detection result is out of the expert consensus quality control standard range, and output to a report.

[0024] A historical query management module is configured to convert the report into a document format, save the report into a storage, and query and track in the storage.

[0025] As an optimization, the server and the client are both implemented based on ReviewX software.

[0026] The application further discloses a device, including a memory, a processor, and a computer program stored on the memory and executable on the processor, and the processor implements the functions of the digital X-ray radiography real-time image quality evaluation control system when executing the computer program.

[0027] The application further discloses a computer readable storage medium, which stores computer instructions, and the computer instructions implement the functions of the digital X-ray radiography real-time image quality evaluation control system when executed by a processor.

[0028] Compared with the prior art, the application has the following advantages and beneficial effects:

[0029] The digital X-ray radiography real-time image quality evaluation control system has the following technical effects:

[0030] 1. Automatic image quality evaluation: the server automatically analyzes the quality of the image through classification, segmentation and detection models, accurately identifies the key anatomical structures in the image, and ensures that the X-ray images of different parts meet the quality control standards. The classification model (ResNetXt-50) ensures the accuracy of image part recognition; the segmentation model (MSU_UNET) provides clear anatomical region segmentation, ensuring the integrity of key part details; and the detection model (YoloV8) monitors the quality indicators of the image in real time. In this way, through the automatic image analysis process, the time and energy demand of manual intervention are reduced, and the efficiency and accuracy of quality control are significantly improved.

[0031] 2. Real-time and efficient feedback: The real-time quality control capability of the system ensures that the image can be immediately evaluated and feedbacked after acquisition. The client receives and displays the quality control results pushed by the server. If the image has abnormal conditions such as inaccurate positioning, overexposure or underexposure, the system will issue a prompt on the client interface for the technician to adjust in real time. The real-time feedback mechanism reduces the number of image retakes, reduces radiation exposure, and improves patient diagnosis and treatment efficiency while ensuring image quality.

[0032] 3. Standardized and consistent quality control: The system uses hospital expert consensus quality control standards. The server algorithm compares the deep learning model analysis results with the standard to automatically determine whether the image meets the standard and generate a report. This standardized quality control process can effectively reduce the result deviation caused by operator experience differences, achieve consistency in image quality control, support image mutual recognition between multiple hospitals, and help improve the accuracy of diagnosis and treatment.

[0033] 4. Convenient image data storage and management: The client automatically generates image quality control reports and supports exporting the reports in PDF format for long-term storage and follow-up. After completing quality control, the technician can directly call historical quality control data through the client, effectively improving the convenience and systematicness of image quality control management, and providing data support for the quality control process.

[0034] 5. System scalability and adaptability: Through the flexibility of the server multi-model architecture, the system can be applied to DR image quality control of multiple parts such as chest, abdomen, cervical spine, lumbar spine, and knee joint. The server model can be adjusted and optimized, and the client data analysis function ensures the adaptability of the system in different medical environments, making the system not only suitable for large-scale medical institutions, but also suitable for image quality control in primary hospitals. BRIEF DESCRIPTION OF DRAWINGS

[0035] The accompanying drawings, which are included to provide a further understanding of the embodiments of the application and are incorporated in and constitute a part of this application, illustrate embodiments of the application and, together with the description, serve to explain the principles of the application. In the drawings:

[0036] Figure 1 is a software architecture diagram of ReviewX;

[0037] Figure 2 is an example of an interface displayed by the client of the system of the present application;

[0038] Figure 3 is Figure 2 is an enlarged schematic view of the lower part;

[0039] Figure 4 is Figure 2 is an enlarged schematic view of the upper part. DETAILED DESCRIPTION

[0040] In order to make the objectives, technical solutions and advantages of the present application clearer, further detailed description will be made to the present application in combination with embodiments and drawings, the illustrative embodiments of the present application and the description thereof are only used to explain the present application, and do not limit the present application.

[0041] The embodiment 1 provides a digital X-ray radiography real-time image quality evaluation control system, as shown in the figure, comprising: Figure 1

[0042] 1, the acquisition end is used for real-time acquisition of X-ray image data of the DROC interface of the digital radiographic imaging system;

[0043] The acquisition end acquires the X-ray image data of the DROC interface in real time through a high-resolution acquisition device supporting 2k / 4k, 30hz or 60hz, so as to ensure the accuracy of quality control. During the acquisition process, the acquisition end does not process the original data, and directly transmits the acquired picture to the service end through the USB3.0 or Type-C interface.

[0044] The specific operation includes:

[0045] HDMI IN: connect the DROC device that needs to be quality controlled, and acquire X-ray images in real time.

[0046] Output interface: transmit the original image without processing to the service end through the USB3.0 or Type-C interface for further processing and analysis.

[0047] 2, the service end is provided with a storage module, a classification model, a segmentation model, a detection model and a quality control analysis module, the storage module stores sample images and expert consensus quality control standards, the classification model, the segmentation model and the detection model are respectively trained and optimized through the sample images, wherein the optimized classification model is used to identify the detection part of the X-ray image data transmitted by the acquisition end, the optimized segmentation model is used to segment the key anatomical part of the detection part in the X-ray image data to obtain a segmentation result; the optimized detection model is used to detect the region of interest in the segmentation result with high precision, so as to obtain the detection result of the detection part about the quality control index; the quality control analysis module performs quality control analysis on the detection result based on the expert consensus quality control standard, so as to obtain a quality control analysis result;

[0048] The service end undertakes the storage, processing and quality control analysis of image data, realizes the automatic classification, segmentation and detection of X-ray images by integrating various deep learning models and quality control standards. The main components and processes of the service end are as follows: ​

[0049] Data collection: The server collects X-ray images of various parts such as the abdomen, cervical spine, knee joint, chest, lumbar spine, etc., totaling 100,000 cases. The data comes from hospitals and is used to train and optimize the performance of subsequent models.

[0050] Classification model: ResNetXt-50(32x4d) model is used (optimized). For X-ray image classification of different parts, especially for difficult-to-distinguish abdominal and lumbar spine images, the server introduces segmentation results as posterior references, uses OCR to identify anatomical part information in the DROC interface in real time to improve the accuracy and stability of classification.

[0051] Segmentation model: The optimized MSU_UNET model is used to generate realistic medical image data through micro-data augmentation, ensuring accurate segmentation of key anatomical parts in the image. The segmentation results are compared with quality control standards to ensure compliance with medical image quality requirements.

[0052] Detection model: YoloV8 model is used (also optimized) for high-precision detection of regions of interest in the image. The server uses data generation augmentation methods to expand training samples and improve the detection accuracy and reliability of the model.

[0053] Quality control algorithm: Based on the quality control standards developed by hospital experts, combined with the results of classification, segmentation, and detection model reasoning, the server calculates image quality indicators in real time, quantitatively analyzes the image quality of 9 different parts, and ensures compliance with medical image quality requirements. It should be noted that subsequent medical image requirements can be quantitatively analyzed for other parts.

[0054] 3. Client, for receiving and parsing the X-ray image data, segmentation results, detection results, and quality control analysis results, thereby automatically generating a quality control report and displaying the segmentation results, detection results, and analysis results in real time. As shown in Figures 2-4 It should be noted that the data in Figure 2 and Figure 4 are hypothetical example data.

[0055] The client is responsible for receiving and displaying quality control results, as well as managing quality control reports and historical data. The main functions include data parsing, image display, automatic evaluation report generation, and historical query. The specific function modules include:

[0056] Receiving and storing module, realizing data receiving and parsing, used for receiving and storing the X-ray image data, segmentation results, detection results, and quality control analysis results transmitted by the server; specifically, receiving image data transmitted by the server and storing it locally to ensure consistency and real-time performance of display data.

[0057] The cropping display module is used for adaptive display of the segmentation result, and if the segmentation result has an abnormal condition, an alarm is sent out in the client to prompt an operator to timely adjust a shooting parameter. Here, the client displays the X-ray original image and the corresponding model segmentation processing result after adaptive cropping transmitted from the server in the image area of the client for adaptive display. The adaptive cropping work is completed in the communication process between the acquisition end device and the server, and the purpose is to crop unnecessary parts in the X-ray image displayed in the DROC (for example, dicom information or other irrelevant areas, and only the X-ray foreground area is reserved). The image displayed by the client is the X-ray original image after adaptive cropping transmitted from the server and the corresponding segmentation result (for example, the segmentation result of the lung field area, so that the lung field area can be drawn in the original X-ray image in the displayed image). The client only needs to perform an initial cropping once during the first installation, that is, the cropping area (the area for displaying the image by the DROC) is set. Subsequent use of the client does not need to perform any cropping operation, and the cropping is completely automatically completed in the server. Of course, the cropping function is also reserved, and the client can also be cropped if necessary.

[0058] The correction module is used for the client to support a technician to manually switch the information of the detection part when it is detected that the detection part identified by the server does not match the actual part, and the modified information is fed back to the server for correction.

[0059] The index display module is used for displaying all quality control indexes related to the detection part, and if a certain quality control index is abnormal, an abnormal prompt is performed; for example, if a certain quality control index is abnormal, a flashing prompt is performed.

[0060] The automatic evaluation report generation module is used for marking the detection result corresponding to the quality control index according to the quality control analysis result, marking the detection result as normal when the detection result is within the expert consensus quality control standard range, and marking the detection result as abnormal when the detection result is out of the expert consensus quality control standard range, and outputting to a report.

[0061] The client automatically generates a quality control evaluation report according to the quality control standard and the index data (detection result) transmitted from the server, and compares the report with the standard (expert consensus quality control standard). The data within the quality control standard range is marked as normal, the data out of the standard is marked as abnormal, and is output to a report.

[0062] The historical query management module is used to convert the report into a document format and save it in the memory, and the query and tracking can be performed in the memory. For example, while saving the quality control evaluation report, the system automatically extracts the patient information in the image for confirmation, and converts the report into a PDF format and saves it to the local hard disk, supporting the technician to query and track the historical report in the client interface.

[0063] In the embodiment, the server and the client are both implemented based on ReviewX software.

[0064] ReviewX is a software for X-ray photography imaging quality evaluation, which consists of two parts: a server (ReviewX-S) and a client (ReviewX-C). The running core of ReviewX-S is based on NVIDIA Jetson AGX Orin, which can provide real-time computing power for the running of ReviewX-C. ReviewX-S has four module nodes, which are image collector node (ImageCollector Node), image processor node (ImageProcessor Node), controller node (Controller Node) and device gateway node (DeviceGateway Node). ReviewX-C has four functional components, which are video / information display component, system setting component, image cropping component and report saving / query component.

[0065] For ReviewX-S, when the main control node receives a start signal (Signal), that is, the running core is ready for the self-start program, the Signal is transmitted to the main mode (Main Mode), and then the Main Mode responds to the ImageCollectorNode and the ImageProcessor Node, respectively.

[0066] When responding to the ImageCollector Node, the Main Mode calls the image acquisition card software program (HDMI SDK) in the node to capture the digital X-ray image acquisition and processing system (DROC) screen interface of the host connected to the current HDMI device in real time by responding to the camera (Camera). The real-time picture obtained is transmitted to the Controller Node through the Main Mode, and the result is packaged and published to ReviewX-C based on the ZMQ communication protocol through the DeviceGateway Node.

[0067] When responding to the ImageProcessor Node, the Main Mode calls the image processing software program (Processor SDK) in the node to process the screen obtained by the previous Main Mode and ImageCollector Node response in real time. The processed image result is responded to the Controller Node again through the Main Mode, and the result is packaged and published to ReviewX-C through the DeviceGateway Node based on the ZMQ communication protocol.

[0068] For ReviewX-C, all the contents of ReviewX-S are obtained by subscribing to the contents packaged and published by ReviewX-S. The parsed HDMI data is transmitted to the video / information display component and the image cropping component for UI interface display. The video / information display component displays the HDMI data transmitted by ReviewX-S in real time, and the image cropping component is initialized and cropped using the HDMI data transmitted by ReviewX-S. On the one hand, the cropped image is used for the UI master display unit to display the current processed image, and ReviewX-C feeds back the cropping coordinates to ReviewX-S for guiding MainMode to adjust the processing result of ImageProcessor Node. On the other hand, according to the specific requirements of the report saving component for saving DICOM text information, the specific area of the DICOM text information display in the current HDMI data is cropped for the normal operation of the text recognition function of the report component. The image cropping component is only used for the initialization step, and ideally the user only needs to use it once. If it needs to be adjusted according to personal reading habits or other factors, repeat the initialization operation. For the system setting component, the user only needs to set the communication IP between the software and the device in this environment, which is the default Ipv4. For the report saving / query component, the quality evaluation information of the current evaluation image and each evaluation detail image are saved, and according to the result of the text recognition function of the image cropping component, the DICOM information of the current processed image is saved for responsibility tracing. The report is saved locally in ReviewX-C, and historical queries can be made in ReviewX-C according to the report name in the report saving / query component.

[0069] Next, the structure of each model on the server side is specifically introduced.

[0070] The classification model is a combined model of ResNetXt-50 model and OCR recognition model, and the specific process of training the classification model is as follows:

[0071] inputting the sample image into the ResNetXt-50 model, outputting a predicted label by the ResNetXt-50 model, and identifying, as an actual label, text information in the sample image by the OCR recognition model, wherein the sample image is from X-ray image data of a DROC interface, and the sample image is labeled with a body part name corresponding to the X-ray image data;

[0072] inputting the predicted label and the actual label into a preset loss function, and optimizing parameters of the ResNetXt-50 model according to the loss function.

[0073] ResNetXt-50 is a deep convolutional neural network based on the ResNet50 structure, but it uses a specific configuration method - 32x4d, which is mainly related to the channel number configuration of its bottleneck layer. The following is a detailed description of the model architecture and principles:

[0074] I. Basic architecture of ResNet50:

[0075] ResNet50 is a residual network that solves the gradient vanishing problem in deep networks by introducing residual blocks. The main idea is to pass the input directly to the later layers through residual connections (skip connections), so that even if the gradient vanishes in the middle layer, the gradient can still be backpropagated through the residual path.

[0076] ResNet50 has 50 layers, mainly composed of convolutional layers and batch normalization layers. A key module in ResNet is the bottleneck block. In ResNet50, the residual block uses three convolutional layers:

[0077] 1x1 convolution to reduce the number of channels, 3x3 convolution for feature extraction, and 1x1 convolution to increase the number of channels. Residual connections directly add input to output, forming a similar skip connection to help the network learn deeper features.

[0078] The meaning of 32x4d: In ResNetXt-50, 32x4d is a modification of the convolutional layer channel configuration in the residual block. 32: represents the number of convolution groups in each group convolution. 4d: represents the base number in each group, meaning that the 3x3 convolution in each residual bottleneck block uses a configuration of 4 basic channels (or feature map dimensions). Here d represents the channel (dimension).

[0079] Therefore, ResNetXt-50 uses a method called Group Convolution, which introduces multiple groups in the convolution layer to handle different input channels. This approach helps reduce computational complexity while increasing the model's feature expression ability. This is different from the standard ResNet, which does not use group convolution or uses a smaller number of groups. Specifically: In the traditional ResNet50, the 3x3 convolution does not use group convolution. In ResNetXt-50, the 3x3 convolution of the bottleneck block is split into 32 groups, and each group's convolution kernel only processes 4 channels of input.

[0080] II. Main levels of ResNetXt-50 architecture

[0081] Convolutional and pooling layers: The network first reduces the size of the input and increases the number of channels through a 7x7 convolutional layer and a max pooling layer.

[0082] Residual blocks: ResNet50 is mainly composed of 4 stages, each containing multiple residual blocks. In ResNetXt-50, these residual blocks are adjusted in terms of channel number and convolution layer grouping, with 32 groups performing 3x3 convolution operations.

[0083] Global pooling and fully connected layers: After the last convolution layer, the network compresses all feature maps into a feature vector through global average pooling. Finally, the feature vector is output through a fully connected layer, usually used for classification tasks.

[0084] III. Principles:

[0085] The design of ResNetXt-50 mainly follows the following two principles:

[0086] 1. Residual learning: Through residual connections, it helps deep networks learn, allowing the network to continue effective training when increasing the number of layers, avoiding the problem of gradient vanishing.

[0087] 2. Group convolution: Use group convolution to reduce computational complexity while maintaining high feature expression ability. Group convolution can reduce the number of parameters for each convolution operation, thereby reducing computational complexity, and enrich features by grouping different channels.

[0088] In some embodiments, the segmentation model is an MSU-Unet model, the MSU-Unet model comprising an encoder and a decoder, wherein the encoder and the decoder each comprise a plurality of convolutional layers, and the convolutional layers in the encoder are symmetrically arranged with the convolutional layers in the decoder, and the outputs of the convolutional layers in the encoder are connected to the inputs of the symmetrically arranged convolutional layers in the decoder by a skip connection; each of the convolutional layers comprises a first convolutional group composed of a first convolutional block of two 3*3 convolutions arranged in series, a second convolutional group composed of a second convolutional block of two 7*7 convolutions arranged in series, and a third convolutional block of a 1*1 convolution, and the image features of the input of the convolutional layer are obtained as first intermediate features and second intermediate features by the first convolutional group and the second convolutional group, respectively, and the first intermediate features and the second intermediate features are combined after channel splicing and then output image features are obtained by the third convolutional block.

[0089] In the encoder, the image features output by the convolutional layer are down-sampled by a max-pooling operation.

[0090] In the decoder, the image features output by the convolutional layer are up-sampled by deconvolution or bilinear.

[0091] MSU-Unet (Multi-Scale U-Net) is an advanced medical image segmentation model based on the classic U-Net architecture. It improves the U-Net by integrating multi-scale convolutional blocks (using convolution kernels of different receptive fields). This design allows better feature extraction at different image scales, making it more effective for dense segmentation tasks such as detecting small or detailed structures. The multi-scale approach enhances the network's ability to capture various semantic features, thereby improving the segmentation performance of medical images.

[0092] Architecture levels:

[0093] The overall architecture of MSU_UNet and U-Net is similar, both composed of an encoder (down-sampling path) and a decoder (up-sampling path), forming a U-shaped network structure.

[0094] The encoder is responsible for feature extraction, gradually reducing spatial dimensions and increasing feature channels; the decoder gradually recovers spatial dimensions through up-sampling while fusing the skip connections from the encoder.

[0095] Main components:

[0096] Encoder part:

[0097] The encoder is composed of multiple convolutional blocks, each using the conv_3_1 module. conv_3_1 is the core module of MSU_UNet, which combines different size kernels like 3x3 and 7x7 to capture more contextual information. Specifically, it uses two convolutional blocks: conv_block_3: 3x3 convolution. conv_block_7: 7x7 convolution.

[0098] The results of these convolutional operations are merged through channel concatenation (torch.cat) and then passed through a 1x1 convolution to reduce the number of channels. The encoder hierarchy starts with shallow features and progressively downsamples the image through Maxpool operations, processing feature maps with 32, 64, 128, 256, and 512 channels in sequence.

[0099] Decoder section:

[0100] The decoder section restores the feature maps to the original resolution through transposed convolutions (deconvolution) or bilinear upsampling. After each decoding step, the feature maps from the encoder are concatenated (torch.cat) with the decoder's output through a skip connection, enabling multi-level feature fusion. In each upsampling convolution operation, the conv_3_1 module is used to maintain the ability to extract multi-scale features. The number of feature channels decreases gradually after upsampling, and the final result is output through a 1x1 convolution.

[0101] Working principle:

[0102] Multi-scale convolution: By using different size kernels (like 3x3 and 7x7) in each convolutional block, MSU_UNet can effectively capture image features of different sizes and improve segmentation accuracy by fusing these multi-scale information.

[0103] Skip connection: The skip connection concatenates the feature maps of each layer in the encoder with the output of the corresponding layer in the decoder, which helps to preserve more low-level detail information and helps the network better recover the spatial structure of the image.

[0104] Upsampling: Using deconvolution or bilinear interpolation to gradually restore the resolution and combine the features from the encoder to gradually refine the segmentation results.

[0105] In this embodiment, the sample pictures used to train the segmentation model and the classification model are subjected to micro data enhancement. The specific operation process of micro data enhancement includes: performing one or more of the following operations on the sample pictures: random brightness adjustment, random saturation adjustment, random contrast adjustment, random translation adjustment, and random cropping adjustment.

[0106] The micro data enhancement method is to replace each m(x) with m(T(x)), wherein x can be an original image Img1 or a preprocessed image Img2, m is a segmentation model or a classification model, and T is a micro data enhancement operation process; the micro data enhancement operation process T includes adjustment of at least one of the following dimensions:

[0107] 1). Random brightness adjustment: adding a random brightness x to the input tensor;

[0108] 2). Random saturation adjustment: x adjusts the saturation of the input tensor by randomly scaling the color channels of the input tensor;

[0109] 3). Random contrast adjustment: x adjusts the contrast of the input tensor by randomly scaling the pixel values;

[0110] 4). Random translation adjustment: x applies a random translation to the input tensor at a specified ratio of 0.125;

[0111] 5). Random crop adjustment: random cropping is applied by zeroing a rectangular region within the input tensor, and the size and position of the cutout are random.

[0112] In this embodiment, the quality control indicators involved in the expert consensus quality control standard include:

[0113] "Chest frontal position":

[0114] Clavicle angle, space centering, scapula spin, exposure range, inspiration detection, chest rotation;

[0115] "Chest lateral position":

[0116] Exposure range, inspiration detection, midaxillary line angle, arms raised, chest rotation;

[0117] "Abdomen":

[0118] Exposure range;

[0119] "Knee joint frontal position":

[0120] Fibular head overlap, exposure range, position centering without deviation;

[0121] "Knee joint lateral position":

[0122] Fibular head overlap, knee joint flexion, medial and lateral condyle overlap;

[0123] "Cervical spine frontal position":

[0124] Exposure range, mandible and occipital bone overlap detection, space detection, mandible left-right asymmetry detection;

[0125] "Cervical spine frontal functional position":

[0126] Exposure range, mandible and occipital bone overlap detection, centering detection, mandible left-right asymmetry detection;

[0127] Cervical spine lateral view:

[0128] Exposure detection, gonial angle overlap detection, cervical spine and mandible overlap detection;

[0129] Lumbar spine frontal view:

[0130] Top and bottom exposure range, pelvis centering and left and right exposure range, spinal column symmetry detection, centerline detection;

[0131] Lumbar spine lateral view:

[0132] Exposure range, lumbar spine centering detection.

[0133] It should be noted that the system adopts a three-layer architecture of acquisition end, service end and client end to realize full-process quality control of X-ray images. The architecture includes a full-automatic workflow from image acquisition to processing, analysis, quality control feedback and storage, and is especially suitable for real-time quality control requirements in medical scenarios. The system architecture design has wide adaptability and can be extended to different medical devices and medical scenarios, which is one of the key innovations of the present application.

[0134] The present application integrates multiple deep learning models on the service end for automatic classification, segmentation and detection of X-ray images of different parts. These include the optimized ResNet50-32x4d classification model, MSU_UNET segmentation model and YoloV8 detection model to ensure accurate analysis of image quality. These models combined with data enhancement and OCR recognition technology help to achieve efficient and stable quality control. However, since some models are open source projects, they may not be covered within the scope of patent protection, but their unique model optimization and application configuration are an important part of the invention.

[0135] The present application adopts quality control standards developed by expert consensus, combined with quality control data (detection results) output by deep learning models, to realize multi-index quantitative evaluation of image quality through algorithms. This post-processing algorithm is based on expert consensus standards and realizes real-time and automated image quality evaluation through integrated analysis of classification, segmentation and detection results. This algorithm not only enables accurate quality control of current images, but also can be applied as a standardized tool for quality control indicators in other types of DR systems, with the potential for cross-system transplantation.

[0136] The client automatically generates a quality control evaluation report by real-time analysis of the quality control data of the server, and supports PDF format export and storage. At the same time, the report history management function of the client provides a convenient storage and query way for subsequent quality control tracking and data review. This report management system can be applied to other medical image quality control systems and has high technical protection value.

[0137] Next, the system of the application is described in detail through specific cases.

[0138] Case 1, system architecture and data flow processing:

[0139] The acquisition end connects the DROC (digital radiological imaging system) device through the HDMI interface to collect X-ray images in real time at a resolution of 2k / 4k, and transmits the images to the server using a USB 3.0 or Type-C interface.

[0140] The server: After the image data is transmitted to the server, the server first identifies the shooting part of the image through the classification model, and transmits the image to the segmentation and detection module to obtain the segmentation result and detection result of the key anatomical structure and quality control indicators in the image. The quality control algorithm of the server calculates the quality control indicators in real time according to the expert consensus standard, and transmits the results to the client.

[0141] The client: The client receives and analyzes the quality control data of the server, automatically generates a quality control report, and displays the quality control results in real time. In abnormal situations (such as overexposure or inaccurate positioning), the client will issue a warning to allow the operator to adjust the shooting parameters in a timely manner.

[0142] Case 2, quality control evaluation of chest frontal X-ray images:

[0143] Acquisition process: The acquisition end captures the chest frontal image and transmits it to the server.

[0144] Classification model: The server identifies the chest frontal image through the classification model.

[0145] Quality control evaluation: The server performs segmentation and detection on the image, calculates quality control indicators including clavicle angle, scapula rotation, exposure range, and inspiration detection. If a certain indicator exceeds the quality control standard range (for example, the clavicle angle exceeds 30°), the server algorithm will generate a quality control warning for the abnormal indicator.

[0146] Report generation: The client receives the quality control data and automatically generates a report, prompting the technician to adjust the posture to ensure that the image quality meets the requirements.

[0147] Case 3, quality control evaluation of lumbar spine frontal X-ray images:

[0148] Acquisition process: The acquisition end collects the lumbar spine frontal image and transmits it to the server.

[0149] Segmentation model: The server-side segmentation model identifies key anatomical regions of the lumbar spine, such as the spinal centerline and bilateral pelvis.

[0150] Quality control evaluation: The server-side detects based on quality control standards such as the symmetry of the spine and pelvis, calculates key quality control parameters such as the upper and lower exposure ranges, pelvis symmetry, centerline symmetry, etc. If a parameter (such as whether the pelvis is aligned) does not meet the standard, the server will generate an abnormal warning.

[0151] Report display: The client generates a detailed quality control report containing the calculation results of each quality control indicator detected, and displays a prompt symbol next to the abnormal item to remind the technician to make necessary adjustments.

[0152] Case 4, generation and management of quality control report:

[0153] Real-time report generation: After each image quality control, the client automatically generates a report containing various quality control indicators, and displays the qualified and abnormal items separately.

[0154] Historical report management: The quality control report is stored in PDF format on the client's hard disk, and the technician can view and manage the saved quality control report through the historical query function of the client, which is convenient for long-term tracking and analysis of quality control data.

[0155] Embodiment 2 also discloses a device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein the processor executes the computer program to realize the functions of the digital X-ray radiography real-time image quality evaluation control system of embodiment 1.

[0156] Embodiment 3 also discloses a computer readable storage medium storing computer instructions, wherein the computer instructions are executed by a processor to realize the functions of the digital X-ray radiography real-time image quality evaluation control system of embodiment 1.

[0157] The above specific embodiments further illustrate the purpose, technical solutions and beneficial effects of the present application. It should be understood that the above description is only a specific embodiment of the present application and is not intended to limit the protection scope of the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A digital radiography real-time image quality assessment control system, characterized by, The application relates to a digital radiology image quality control system. The system comprises a collection terminal, a service terminal and a client terminal. The collection terminal is used for collecting X-ray image data of a DROC interface in real time. The service terminal is provided with a storage module, a classification model, a segmentation model, a detection model and a quality control analysis module. The storage module stores sample images and expert consensus quality control standards. The classification model, the segmentation model and the detection model are trained and optimized through the sample images. The optimized classification model is used for identifying a detection part of the X-ray image data transmitted by the collection terminal. The optimized segmentation model is used for segmenting key anatomical parts of the detection part in the X-ray image data to obtain a segmentation result. The optimized detection model is used for high-precision detection of a region of interest in the segmentation result to obtain a detection result of the detection part with respect to a quality control index. The quality control analysis module performs quality control analysis on the detection result based on the expert consensus quality control standards to obtain a quality control analysis result. The client terminal is used for receiving and analyzing the X-ray image data, the segmentation result, the detection result and the quality control analysis result to automatically generate a quality control report and display the segmentation result, the detection result and the analysis result in real time. If the X-ray image data has an abnormal condition, which is a detection result not meeting the expert consensus quality control standards, an alarm is sent on a client terminal interface to prompt a technician to timely adjust patient positioning and collimation. The quality control index involved in the expert consensus quality control standards comprises: For a chest frontal position: Angle of clavicle, centering of occupying, scapula rotation, exposure range, inspiration detection and chest rotation; For a chest lateral position: Exposure range, inspiration detection, angle of midaxillary line, arms raised, chest rotation; For an abdomen: Exposure range; For a knee joint frontal position: Overlapping degree of fibular head, exposure range and centering of positioning without deviation; For a knee joint lateral position: Overlapping of fibular head, knee joint bending and internal and external condyle overlapping; For a cervical spine frontal position: Exposure range, overlapping detection of mandible and occipital bone, centering detection and left-right asymmetry detection of mandible; For a cervical spine frontal functional position: Exposure range, overlapping detection of mandible and occipital bone, centering detection and left-right asymmetry detection of mandible; For a cervical spine lateral position: Exposure detection, overlapping detection of mandibular angle and overlapping detection of cervical spine and mandible; For a lumbar spine frontal position: Upper and lower exposure ranges, pelvis positioning and left and right exposure ranges, spinal symmetry detection and center line detection; For a lumbar spine lateral position: Exposure range and lumbar spine positioning detection; The classification model is a combined model of a ResNetXt-50 model and an OCR recognition model. The specific process of training the classification model is as follows: The sample images are input into the ResNetXt-50 model to output predicted labels. The OCR recognition model is used to identify text information in the sample images as actual labels. The sample images come from X-ray image data of a DROC interface, and the sample images are labeled with body part names corresponding to the X-ray image data. The predicted label and the actual label are input into a preset loss function, and parameters of the ResNetXt-50 model are optimized according to the loss function; The segmentation model is an MSU-Unet model, the MSU-Unet model comprises an encoder and a decoder, wherein the encoder and the decoder each comprise a plurality of convolutional layers, the convolutional layers in the encoder and the convolutional layers in the decoder are symmetrically arranged, and the output ends of the convolutional layers in the encoder and the input ends of the symmetrically arranged convolutional layers in the decoder are connected in a manner of jump connection; each convolutional layer comprises a first convolutional group composed of a first convolutional block of two 3*3 convolutions arranged in series, a second convolutional group composed of a second convolutional block of two 7*7 convolutions arranged in series, and a third convolutional block of a 1*1 convolution, the image features input into the convolutional layer are obtained as first intermediate features and second intermediate features through the first convolutional group and the second convolutional group respectively, and the first intermediate features and the second intermediate features are combined after channel splicing and then output image features are obtained through the third convolutional block.

2. The digital radiography real-time image quality assessment control system of claim 1, wherein, In the encoder, the image features output by the convolutional layers are down-sampled through a maximum pooling operation.

3. The digital radiography real-time image quality assessment control system of claim 1, wherein, In the decoder, the image features output by the convolutional layers are up-sampled through deconvolution or bilinear.

4. The digital radiography real-time image quality assessment control system of claim 1, wherein, The sample pictures used for training the segmentation model and the classification model are subjected to micro data enhancement, and the operation process of the micro data enhancement specifically comprises one or more of the following operations: random brightness adjustment, random saturation adjustment, random contrast adjustment, random translation adjustment, and random cropping adjustment.

5. The digital radiography real-time image quality assessment control system of claim 1, wherein, The client comprises: a receiving and storing module configured to receive and store the X-ray image data, the segmentation result, the detection result, and the quality control analysis result transmitted by the server; A cropping and displaying module is configured to automatically crop a region requiring quality control in the segmentation result according to the quality control analysis result and to adaptively display the region, and to issue a warning in the client to prompt an operator to timely adjust a shooting parameter if the analysis result is abnormal; A correction module is configured to manually switch information of a detection part when it is detected that the detection part recognized by the server does not match an actual detection part, and to feed back the modified information to the server for correction; An index displaying module is configured to display all quality control indexes related to the detection part, and to issue an abnormality prompt if a certain quality control index is abnormal; An automatic evaluation report generating module is configured to mark a detection result corresponding to a quality control index according to the quality control analysis result, to mark the detection result as normal if the detection result is within a range of an expert consensus quality control standard, and to mark the detection result as abnormal if the detection result is outside the range of the expert consensus quality control standard, and to output the detection result to a report; A historical query management module is configured to convert the report into a document format, to save the report in a storage, and to query and track the report in the storage.

6. The digital radiography real-time image quality assessment control system of claim 1, wherein, The server and the client are both implemented based on ReviewX software.

7. An apparatus, comprising: Comprise: Memory, a processor, and a computer program stored on the memory and executable on the processor, wherein the processor implements the functions of the digital radiography real-time image quality evaluation control system according to any one of claims 1-6 when executing the computer program.

8. A computer-readable storage medium, characterized in that, The computer readable storage medium stores computer instructions, and the computer instructions implement the functions of the digital radiography real-time image quality evaluation control system according to any one of claims 1-6 when executed by a processor.

Images