Configurable quality control system and method for full-spine X-ray film

Abstract

The embodiment of the invention relates to a full-spine X-ray film configurable quality control system and method. The system comprises a configuration and knowledge base which is set to set operation environment parameters through a preset configuration file and to evaluate anatomical standards, a label mapping dictionary and examination rules required by a full-spine X-ray film; the data adaptation and standardization module is set to read full-spine X-ray labeling data according to the operating environment parameters, carry out standardized label mapping according to the label mapping dictionary, and normalize pixel coordinates of an input full-spine X-ray film to a preset standard coordinate space; the anatomical feature extraction module is configured to receive the data of the whole-spine X-ray film after the standardization label and the coordinate normalization, and output a structural feature measurement value required by consistency check; the consistency checking engine is set to carry out quality checking according to the structural feature measurement value and the checking rule; and the report output module is set to generate a quality report containing the quality level according to the inspection result.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to a configurable quality control system and method for whole spine X-ray films. Background Technology

[0002] Spinal deformities are surgically correctable. Accurate assessment of overall spinal alignment and sagittal balance is crucial for successful preoperative planning and surgical correction of spinal deformities such as scoliosis. Core spinopelvic parameters, including the pelvic incidence angle (PI), lumbar lordosis angle (LL), sacral tilt angle (SS), and sagittal vertical axis (SVA), are typically measured via full-spine standing radiographs. Precise identification of anatomical landmarks, such as the highest points of the clavicle and pelvis, the superior endplate of the S1 vertebra, and the midpoint of the femoral head (CFH), is essential for defining the coordinate system and calculating these critical surgical parameters. Errors in locating these key points can directly lead to significant deviations in surgical decisions.

[0003] Given the time-consuming nature and significant operator variability of manual X-ray measurements, deep learning-based artificial intelligence (AI) models have been rapidly developed to automate the detection, segmentation, and measurement of whole-spine X-rays. These AI tools promise to revolutionize preoperative assessment by providing surgeons with rapid, consistent, and highly reliable biometric data. However, the performance and clinical reliability of these AI models are fundamentally limited by the quality and accuracy of the training data. Even subtle topological errors, such as incorrect vertebral labeling or incorrect spinal sequence order, can severely impact the model's learned representations, leading to catastrophic predictive failures in real-world clinical use, especially in surgical planning scenarios where precision is not a compromise.

[0004] Currently available open-source datasets are limited in number, and the quality of labeled data is generally low. In clinical practice, the construction of private datasets still largely requires manual, time-consuming, and costly expert annotation. Due to the lack of unified and standardized quality control standards, this traditional annotation method is prone to systematic biases and random errors, leading to unreliable annotation results. This not only wastes valuable professional resources but also directly impairs the robustness and generalization ability of the model.

[0005] Currently, quality control (QC) for large-scale radiological annotation datasets primarily relies on manual consensus or simple inter-rater consistency checks. For complex tasks such as Cobb angle measurement in scoliosis, existing practices typically involve senior experts performing redundant measurements to reach consensus, or utilizing a panel of multiple radiologists to determine a reference standard through median consensus. While these methods strive for high fidelity, their inherent limitations stem from high costs, reliance on high-level professional resources, and difficulty in establishing automated QC processes. Some methods have introduced statistical consistency metrics such as Krippendorff's Alpha to assess annotation reliability. However, these metrics are generally more suitable for structurally simple target detection tasks and have significant limitations when applied to complex anatomical annotations such as spinal segmentation and landmark localization. Other advanced efforts focus on analyzing annotator behavior or introducing complex automated processes, but these processes typically have extremely high hardware and software requirements, hindering their widespread application in large-scale medical image annotation scenarios. Furthermore, while robust quality management frameworks exist in the field of Natural Language Processing (NLP) (such as annotation process management and dispute resolution), they lack coverage for high-risk, visually complex applications such as spinal surgery planning data, and cannot be directly extended to data quality control scenarios involving measurements of Cobb angles or other spinal parameters. Therefore, there is an urgent need for a more efficient, anatomically aware, and configuration-driven QC solution. Summary of the Invention

[0006] This application provides a configurable quality control system and method for full spinal X-ray images, the purpose of which is to ensure the consistency of annotation through the quality control system, effectively eliminate structural errors, and thus ensure the high robustness and generalization ability of the model.

[0007] In a first aspect, embodiments of this application provide a configurable quality control system for whole-spine X-ray images, including: Configuration and knowledge base, set up to configure runtime environment parameters, as well as anatomical standards, label mapping dictionaries and examination rules required for evaluating full spine X-rays through preset configuration files; The data adaptation and standardization module is configured to read the annotation data of the whole spine X-ray according to the configuration and the runtime environment parameters defined in the knowledge base, and perform standardized label mapping according to the label mapping dictionary to normalize the pixel coordinates of the input whole spine X-ray to a preset standard coordinate space. The anatomical feature extraction module is configured to receive standardized labels and coordinate-normalized whole spine X-ray data output by the data adaptation and standardization module, and output the structural feature measurement values ​​required for consistency inspection. The consistency check engine is configured to perform quality checks based on the structural feature measurement values ​​output by the anatomical feature extraction module and the check rules defined in the configuration and knowledge base. The report output module is configured to generate a quality report containing quality levels based on the inspection results output by the consistency inspection engine.

[0008] Optionally, the configuration and knowledge base are specifically set as follows: The anatomical standards, label mapping dictionary, inspection rules, and runtime environment parameters are defined through a preset configuration file. The runtime environment parameters include the annotation software type, the root directory path of the dataset, the storage structure of the annotation files in the file system, and the annotation orientation of the image. The anatomical criteria include the standard spinal sequence; The label mapping dictionary includes the actual label name, standard label, and the mapping relationship between non-standard labels in the actual label name and standard labels; the types of marker points and key lines and their label mappings, as well as all statistics and position thresholds used for automatic verification checks; The inspection rules include inspection items, absolute position judgment thresholds, anatomical reference relationships of relative positions, definitions of optional annotations, and annotation types.

[0009] Optionally, the data adaptation and standardization module is specifically configured as follows: The pixel coordinates of the full spine X-ray are normalized to a preset standard coordinate space; wherein, the input full spine X-ray has a width of Height is Image, normalized scale Normalized coordinate points The calculation method is as follows , These are the coordinates of the points in the original image corresponding to the normalized coordinate points. This indicates that the option to mirror the X-axis is enabled. This indicates that the option to mirror the X-axis is not enabled. If the input full spine X-ray data lacks data for tools with well-defined image dimensions, dynamic normalization is performed; where the normalization ratio is determined by the coordinate range of the input full spine X-ray, where... and Among all valid pixels in the input full-spine X-ray coordinate system, This represents the maximum X-axis coordinate value. This represents the minimum coordinate value on the X-axis. This represents the maximum Y-axis coordinate value. This represents the minimum coordinate value on the Y-axis. The X-axis is the horizontal direction of the image (from left to right in the positive direction), and the Y-axis is the vertical direction of the image (from top to bottom in the positive direction).

[0010] Optionally, the anatomical feature extraction module is specifically configured as follows: For the segmentation mask of the vertebral bodies in a full spine X-ray, calculate the centroid coordinates and bounding box of each vertebral body; Extract the normalized coordinates of the marker points and key lines, and calculate the center points of the marker points and key lines; Relationship data generated based on the center point of the key marker points.

[0011] Optionally, the anatomical feature extraction module is further configured as follows: If a preset vertebra is missing from the input full spine X-ray, the simulated center point of the preset vertebra is derived based on the missing preset vertebra's number in the standard spine sequence. The expected value of the simulated center point on the normalized Y-axis is used as a reference point for relative position examination in the case of a missed vertebral marker. Among them, the standard spinal sequence index The center point of the pre-set vertebral body is missing. Expected value on the Y-axis in It represents the Y-axis coordinates of the center point of the highest effective vertebral body observed in the current spinal sequence. It represents the Y-axis coordinates of the center point of the lowest effective vertebral body observed in the current spinal sequence. yes and The total number of vertebrae that should theoretically exist between them is used as the length of the reference spinal sequence.

[0012] Optionally, the consistency check engine is specifically configured as follows: Using the centroid coordinates of the vertebral bodies along the normalized Y-axis and the corresponding labels, we check whether the spinal sequence satisfies the medical order of increasing from top to bottom and determine the topological order. Identify omissions and errors, and determine the completeness and existence of vertebral segmentation and labeling. Check the correctness of the names and orientation assignments of paired markers, including applying absolute spatial constraints for lateral orientation checks; in Indicates the normalized median; This indicates the coordinates of the left-hand marker in the normalized image. This indicates the coordinates of the right-hand marker in the normalized image; A relative consistency check is performed based on the relative position and distance between the markers, whereby... in It is a predefined position tolerance, where and These represent the Y-axis coordinates of the center points of the i-th and (x+1)-th vertebrae in the spinal sequence, respectively, in the normalized image coordinate system.

[0013] Optionally, the report output module is specifically configured as follows: A quality report is output based on the errors found in the inspection results and the preset error levels.

[0014] Optionally, the preset error levels include failure and warning; the failure level includes failure of the vertebral topology sequence check, laterality check and integrity check; the warning level includes warnings for relative consistency check failure and non-critical additional labels.

[0015] Secondly, embodiments of this application provide a method for quality control of whole-spine X-ray data, and a configurable quality control system for whole-spine X-rays provided in any embodiment of this application includes: Obtain a full spine X-ray; wherein the full spine X-ray includes images of the vertebral bodies and labeling of the vertebral bodies; The annotation data of the whole spine X-ray is read according to the runtime environment parameters defined by the configuration and knowledge base, and the label mapping is standardized according to the label mapping dictionary to normalize the pixel coordinates of the whole spine X-ray to the preset standard coordinate space. Based on the data from the whole spine X-ray after standardization of labels and coordinate normalization, output the structural feature measurements required for consistency checks; Quality inspection is conducted based on the structural feature measurements and inspection rules. Based on the inspection results, a quality report containing the quality level is generated.

[0016] Optional methods for quality control of whole-spine X-ray data also include: The operating environment parameters, as well as the anatomical standards, label mapping dictionary, and examination rules required for evaluating whole spine X-rays, are set through a preset configuration file.

[0017] This application provides a configurable quality control system and method for whole-spine X-ray images. This system rigorously enforces spinal surgical anatomical standards through automated examination of whole-spine X-ray images. These examinations aim to ensure compliance with specifications and topological integrity, and introduce a hierarchical structure for error grading to clearly distinguish between critical topological errors that severely disrupt the representation of anatomical structures and minor positional deviations. This overcomes key limitations of traditional quality assurance methods, including high dependence on specialized resources, time-consuming processes, and inherent inconsistencies in manual quality control. The primary objective of this system is to ensure consistency of annotation protocols and integrity of topological structures, rather than guaranteeing absolute subpixel accuracy. Attached Figure Description

[0018] Figure 1This is a schematic diagram of a configurable quality control system for whole-spine X-ray images provided in an embodiment of the present invention; Figure 2 A flowchart illustrating a method for quality control of whole-spine X-ray data provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of a whole spine X-ray examination case provided in an embodiment of the present invention. Detailed Implementation

[0019] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Furthermore, it should be noted that, for ease of description, the accompanying drawings show only the parts relevant to the present invention, and not all of the structures.

[0020] Figure 1 This is a schematic diagram of a configurable quality control system for whole-spine X-ray images provided in an embodiment of the present invention. This system can be used for quality control of vertebral body identification and annotation in whole-spine X-ray images. The system includes: The configuration and knowledge base are set up by configuring runtime environment parameters, as well as the anatomical standards, label mapping dictionaries, and examination rules required for evaluating whole-spine X-rays, through a preset configuration file. The configuration and knowledge base, serving as the system's configurable data dictionary and rule set, is defined through a dedicated configuration file. This library module centrally stores all the anatomical standards and runtime environment parameters required for comprehensive evaluation, ensuring the system's flexibility and versatility across different datasets.

[0021] The data adaptation and standardization module is configured to read the annotation data of full-spine X-rays according to the runtime environment parameters defined in the configuration and knowledge base, and perform standardized label mapping according to the label mapping dictionary, normalizing the pixel coordinates of the input full-spine X-ray images to a preset standard coordinate space. This module serves as the system's standardized data input layer, aiming to support a diverse medical annotation ecosystem. This module can natively handle common annotation software formats, including LabelMe JSON files and the multi-file structure of 3D Slicer (e.g., NRRD files for segmentation and .mrk.json files for keypoints). The data adaptation and standardization module is responsible for reading and standardizing the annotation data of anteroposterior (AP) or lateral (LAT) full-spine X-ray images according to the runtime environment parameters defined in the configuration and knowledge base. During this process, it performs label mapping, uniformly converting the actual labels used by the annotator (e.g., numbers 0-17, Segment_0-Segment_17, or medical labels such as C7-L5) into the medical standard terminology used internally by the system. This is a crucial step in ensuring terminology consistency. Secondly, this module performs the essential coordinate scaling and normalization. It normalizes all input raw pixel coordinates, regardless of image size and resolution, to a uniform preset standard space, such as a preset standard space of 0-1000.

[0022] The anatomical feature extraction module is configured to receive standardized labels and coordinate-normalized whole-spine X-ray data output by the data adaptation and standardization module, and output structural feature measurements required for consistency checks. Specifically, the anatomical feature extraction module processes the standardized and normalized input data output by the data adaptation and standardization module to derive the quantitative geometric and structural features required for verification. This module transforms raw point and polygon data into measurements that can be directly used by the automatic consistency check engine.

[0023] The consistency check engine is configured to perform quality checks based on the structural feature measurements output by the anatomical feature extraction module and the check rules defined in the configuration and knowledge base. The consistency check engine is the core logical unit of the entire quality control system; it enforces annotation protocols and anatomical specifications through objective, rule-based logic. This engine receives normalized measurements provided by the anatomical feature extraction module and verifies them against thresholds and rules defined in the configuration and knowledge base.

[0024] The report output module is configured to generate a quality report containing quality levels based on the inspection results output by the consistency inspection engine.

[0025] Optionally, the configuration and knowledge base are specifically set as follows: The anatomical standards, label mapping dictionary, inspection rules, and runtime environment parameters are defined through a preset configuration file. This preset configuration file can be a JSON (JavaScript Object Notation) configuration file. The JSON file fields are configured to define the anatomical standards, label mapping dictionary, inspection rules, and runtime environment parameters. By reading the JSON file, the configuration is parsed into entity classes. The parsed configuration dictionary allows direct access to configuration fields via keys. Then, for different annotation software and different annotation files, the logic corresponding to the configuration is selected to support multiple annotation software, multiple types of annotation files, and multiple file structures.

[0026] The runtime environment parameters include the annotation software type, the root directory path of the dataset, the storage structure of the annotation files in the file system, and the annotation orientation of the images. These parameters can be set by a JSON configuration file. The annotation software type is, for example, LabelMe or 3D Slicer. The storage structure of the annotation files in the file system is, for example, a flat file structure or a nested patient / sequence directory. The annotation orientation of the images is, for example, AP (anteroposterior) or LAT (lateral).

[0027] The anatomical criteria include a standard spinal sequence, for example, from C0 to S1.

[0028] The label mapping dictionary includes actual label names, standard labels, and the mapping relationship between non-standard labels and standard labels in the actual label names (the standard labels use internal medical standard terminology to uniformly convert non-standard labels used by labelers into internal medical standard terminology); the types of marker points and key lines and their label mappings, as well as all statistics and location thresholds used for automatic verification checks; ,in A label mapping function is defined to represent internal medical standard terms (e.g., "C7", "S1"). Used to label various actual names Converted to this internal medical standard terminology: This ensures that non-standard labels are converted into internal medical standard terminology. All possible actual label names can be configured in a JSON file, centered around the medical standard labels. The key is the internal standard name of the quality control system, and the value is a list of actual labels that the labeler might use.

[0029] The inspection rules include inspection items, absolute position judgment thresholds, anatomical reference relationships of relative positions, definitions of optional annotations, and annotation types (e.g., points, lines, or polygons).

[0030] Optionally, the data adaptation and standardization module is specifically configured as follows: The pixel coordinates of the full spine X-ray are normalized to a preset standard coordinate space; wherein, the input full spine X-ray has a width of Height is Image, normalized scale Normalized coordinate points The calculation method is as follows , These are the coordinates of the points in the original image corresponding to the normalized coordinate points. This indicates that the option to mirror the X-axis is enabled. This indicates that the option to mirror the X-axis is not enabled. This standardization process ensures that subsequent anatomical feature extraction and quality control checks are completely independent of the original image size, greatly enhancing the system's versatility and robustness.

[0031] If the input full spine X-ray data lacks a defined image size, dynamic normalization is performed; where the normalization ratio is determined by the coordinate range of the input full spine X-ray. and Of all valid pixels in the input full-spine X-ray coordinate system, This represents the maximum X-axis coordinate value. This represents the minimum coordinate value on the X-axis. This represents the maximum Y-axis coordinate value. This represents the minimum Y-axis coordinate value. For data from tools such as 3D Slicer, the module performs dynamic normalization, determining the scaling ratio based on the minimum / maximum range of all detected keypoints to achieve optimal normalization on the 2D projection. Furthermore, if the option to mirror the X-axis is enabled in the configuration, the module will also flip the X-axis coordinates at this stage to ensure all data is input into the QC engine in a consistent direction.

[0032] Optionally, the anatomical feature extraction module is specifically configured as follows: For a segmentation mask of vertebrae in a full-spine X-ray, calculate the centroid coordinates and bounding box of each vertebra; specifically, for the segmentation mask (polygon), calculate the centroid coordinates of each vertebra from a set of n normalized points. As the geometric centroid representing the anatomical center of the vertebral body: Where n is the number of points defining the feature. These are the normalized coordinates of the points that make up the annotation.

[0033] Extract the normalized coordinates of marker points and key lines and calculate their center points; for marker point and key line data, extract their normalized X-axis and Y-axis coordinates and calculate their center points.

[0034] Relationship data generated based on the center point of the key markers, such as the distance and relative position between key points.

[0035] Optionally, the anatomical feature extraction module is further configured as follows: If a preset vertebra is missing from the input full spine X-ray, the simulated center point of the preset vertebra is derived based on the missing preset vertebra's number in the standard spine sequence. The expected value of the simulated center point on the normalized Y-axis is used as a reference point for relative position examination in the case of a missed vertebral marker. Among them, the standard spinal sequence index The center point of the pre-set vertebral body is missing. Expected value on the Y-axis in It refers to the length of the reference spinal sequence.

[0036] This module features an anatomically knowledge-driven simulation function. During relative consistency checks, if a necessary segmented vertebra is missing from the data, the module does not simply skip the check. Instead, it derives a simulated center point based on the configuration and the standard spinal sequence defined in the knowledge base, using the vertebra's ordinal number within the sequence. This simulated center point is assigned an expected value on the normalized Y-axis, serving as a reference point for relative position checks in the event of a missed vertebra, thereby enhancing the system's robustness to data incompleteness and its check coverage.

[0037] Optionally, the consistency check engine is specifically configured as follows: Using the centroid coordinates of the vertebrae along the normalized Y-axis and the corresponding labels, the spinal sequence is checked to see if it follows a top-to-bottom increasing medical order, thus determining the topological order; sequence reversal (e.g., L1 below L2) is marked as a serious error.

[0038] Identify omissions and errors to determine the completeness and existence of vertebral segmentation and labeling; omissions may be the absence of necessary segmentation segments, while errors may be the presence of redundant or unexpected labels.

[0039] Check the correctness of the names and orientation assignments of paired markers, including applying absolute spatial constraints for lateral orientation checks; in To represent the normalized median, take ; This indicates the coordinates of the left-hand marker in the normalized image. This indicates the coordinates of the right-hand marker in the normalized image. Failure to perform a lateral check (e.g., a left-hand point being placed to the right of the normalized X-axis center 500) is considered a serious structural error.

[0040] A relative consistency check is performed based on the relative position and distance between the markers, whereby... in This is a predefined location tolerance, which can be defined as 5.0 normalized pixels. The relative consistency check aims to flag deviations that exceed the expected anatomical statistics; the results are typically categorized as warnings.

[0041] Optionally, the report output module is specifically configured as follows: A quality report is output based on the errors found in the inspection results and the preset error levels.

[0042] Optionally, the preset error levels include failure and warning; the failure level includes failure of the vertebral topology sequence check, laterality check and integrity check; the warning level includes warnings for relative consistency check failure and non-critical additional labels.

[0043] Failures (primarily structural errors) apply to violations that severely damage the data, fundamentally violating the anatomical topology and rendering the data unusable. This level primarily includes failures of topological order checks, laterality checks (e.g., incorrect left / right assignment), and necessary integrity checks (e.g., missing critical segments).

[0044] Warnings (primarily for consistency deviations) apply to non-critical deviations that require expert review but do not disrupt the basic structure. This level primarily includes alerts for relative consistency check failures (inconsistencies in locations outside the statistical range) and non-critical additional labels (incorrect labeling).

[0045] This module summarizes the results and generates a comprehensive summary report. Crucially, only data without any reported anomalies is considered acceptable for model training and validation; any data with reported anomalies must be manually reviewed and re-labeled until it is fully acceptable before it can be used.

[0046] In some specific implementations, dataset AP-2 is annotated using 3D Slicer (version >= 5.2.0), a tool suitable for high-precision medical image analysis due to its support for DICOM format, 3D visualization, and complex segmentation capabilities. Datasets AP-1 and LAT-1, on the other hand, are annotated using LabelMe (version >= 5.2.1), utilizing its polygonal outline drawing and JSON export functions.

[0047] The target of the vertebral segmentation task is 18 vertebrae from C7 to L5. The contours are drawn by segmenting each vertebra independently, with the annotation boundaries limited to the entire bony area of ​​the vertebral body, excluding structures such as intervertebral discs or pedicles. The annotation process requires free-form polygon drawing, ensuring the contours closely match the actual boundaries and that each vertebra is a closed region. For images of scoliosis or rotational deformities, all identifiable vertebrae must be segmented as accurately as possible. If the boundaries of individual vertebrae are blurred and difficult to identify, annotation can be skipped, but the reason must be noted in the annotation log file to ensure data traceability. Due to historical reasons and differences in tools across different data batches, the vertebral body labels differ across the three datasets: AP-1 uses numbers 0-17, AP-2 uses Segment_0-Segment_17, while LAT-1 directly uses medical labels such as C7-L5. All datasets assign special labels to specific vertebrae (such as L6 or T13).

[0048] Keypoint and keyline labeling tasks are used to assist in training the model to automatically calculate commonly used clinical parameters in spinal surgery, such as the Cobb angle, SVA, and pelvic parameters. In anteroposterior images, keypoints include the highest point of the right clavicle (CR), the highest point of the left clavicle (CL), the highest point of the right ilium (IR), the highest point of the left ilium (IL), and the right and left borders (SL) of the superior sacral endplate. In lateral images, the labeled objects include the straight line marking of the superior sacral endplate (S1) and the midpoint of the line connecting the midpoints of the bilateral femoral heads (CFH).

[0049] These precise segmentations and landmark localizations provide a data foundation for training models to automatically detect and calculate parameters with extremely high clinical value, such as the coronal / sagittal T1 tilt angle, trunk-shift, coronal / sagittal Cobb angle, sagittal SVA, and pelvic incidence angle (PI), highlighting the necessity of quality control of labeled data.

[0050] Table 1

[0051] The database used in the above data study integrates X-ray images from three independent clinical data sources, totaling 2,358 high-quality whole spine images.

[0052] The anteroposterior (AP) dataset consists of two parts: AP-1 and AP-2. Dataset AP-1 comprises 3140 full-length standing anteroposterior spinal X-rays, collected between May 19, 2025 and June 18, 2025. Within AP-1, 1676 cases (AP-1.1) contain only T1-L5 full-spine segmentation data, lacking keypoint / line annotations; the remaining 1464 cases (AP-1.2) contain complete C7-L5 full-spine segmentation and annotations of all keypoints defined in the anteroposterior view manual. All patients enrolled in this dataset underwent preoperative full-length standing anteroposterior spinal X-ray examinations, and patients were strictly excluded due to diseases that might affect standing posture, poor image quality, the presence of metallic implants causing significant artifacts or occlusion, or incomplete clinical or imaging data. Dataset AP-2 contains 691 cases, collected between July 28, 2014 and July 13, 2023. AP-2 includes complete segmentation (C7-L5) and annotation of all key points defined in the annotation manual. Inclusion criteria are patients diagnosed with scoliosis by a specialist and with access to original DICOM images. Exclusion criteria are more stringent, excluding images with incorrect standing posture, severe rotation, incomplete coverage of the spine or pelvis, and images with internal metal fixation devices or braces, casts, or other devices that significantly affect vertebral morphology assessment.

[0053] The LAT-1 lateral radiograph dataset contains 391 full-spine lateral radiographs, acquired between June 1, 2014, and June 30, 2025. LAT-1 includes all keypoints / line annotations defined in the complete segmentation and annotation manual. This dataset only includes patients with idiopathic scoliosis undergoing initial surgery and excludes images with poor quality, significant artifacts, occluded metal implants, or incomplete clinical or imaging data.

[0054] Table 1 details the annotation tools, file formats, and specific annotation content for each subset of the dataset.

[0055] This analysis phase aims to evaluate the compliance of the dataset annotations and the consistency of the anatomical topology. The analysis covers multiple dimensions, including the existence of annotations, the correctness of annotation labels, the compliance of vertebral annotation order, the correctness of the absolute left and right positions of keypoints, and the correctness of the relative positions of the annotation results. These judgments are all verified based on the configuration and predefined annotation content and label specifications in the knowledge base.

[0056] The process began with an existence check and label correctness verification. Following this, a topological order compliance check of the vertebral body annotations was performed, completed by an automated compliance check engine. For the 18 vertebrae of the entire spine (C7-L5, with T13 and L6 being optional), they must be arranged strictly according to the medical sequence from top to bottom. This vertical positional relationship was calculated from the simulated vertebral body center points extracted by the anatomical feature extraction module.

[0057] Regarding keypoint annotation, the system first performs a laterality check to determine whether paired keypoints meet the absolute left-right position criteria. This determination is made by checking the relationship between the X-axis coordinates in the normalized coordinate system (0-1000) of the data adaptation and standardization module and the center threshold of 500. Simultaneously, the system also performs a relative consistency check, judging the accuracy of the relative positions of the annotations based on preset anatomical topological relationships. For example, the highest points of the left and right clavicles in the anteroposterior view should be above the relative label T4; the highest points of the left and right pelvis in the anteroposterior view should be below T12; the highest points of the upper margins of the left and right S1 vertebrae in the anteroposterior view, the straight line of the upper endplate of the sacrum in the lateral view, and the midpoint of the line connecting the midpoints of the femoral heads in the lateral view should all be below the relative label L5. These vertical judgments of relative positions also rely on the simulated vertebral body center points extracted by the anatomical feature extraction module. Example of the check: Figure 3 As shown.

[0058] Finally, the error classification and reporting module output an analysis report. Among these, errors in annotation existence, annotation label correctness, annotation order compliance, and the absolute left-right position of key points were categorized as failures (structural errors), indicating fundamental topological errors in the data requiring re-annotation. Errors in relative position checks were categorized as warnings (consistency deviations), applicable to non-critical positional deviations.

[0059] To scientifically verify the critical impact of data reliability on AI model performance, this study compared the effectiveness of training deep learning models on two different datasets: an unverified original dataset and a manually selected control dataset. The aim was to rigorously quantify the performance improvement achieved solely by eliminating labeled anomalies, thereby establishing a benchmark for evaluating the necessity of subsequently introducing a CQCS automated quality control solution.

[0060] like Figure 2 The method for quality control of whole-spine X-ray data, as shown, utilizes a configurable quality control system for whole-spine X-rays provided in any of the above embodiments of this application, including: Step 210: Obtain a full spine X-ray; wherein the full spine X-ray includes images of the vertebral bodies and labeling of the vertebral bodies; Step 220: Read the annotation data of the whole spine X-ray according to the runtime environment parameters defined by the configuration and knowledge base, and perform standardized label mapping according to the label mapping dictionary to normalize the pixel coordinates of the whole spine X-ray to the preset standard coordinate space. Step 230: Based on the data from the whole spine X-ray after standardization of labels and coordinate normalization, output the structural feature measurement values ​​required for consistency inspection; Step 240: Perform quality inspection according to the measured values ​​of the structural features and the inspection rules; Step 250: Based on the inspection results, generate a quality report containing the quality level.

[0061] Optional methods for quality control of whole-spine X-ray data also include: The operating environment parameters, as well as the anatomical standards, label mapping dictionary, and examination rules required for evaluating whole spine X-rays, are set through a preset configuration file.

[0062] Although the present invention has been described in detail above with general descriptions, specific embodiments, and experiments, modifications or improvements can be made to it, which will be obvious to those skilled in the art. Therefore, all such modifications or improvements made without departing from the spirit of the present invention fall within the scope of protection claimed by the present invention.

Claims

1. A configurable quality control system for whole-spine X-ray films, characterized in that, include: The configuration and knowledge base are set up to configure the runtime environment parameters and the anatomical standards, label mapping dictionary and examination rules required to evaluate whole spine X-rays through a preset configuration file. The data adaptation and standardization module is configured to read the annotation data of the whole spine X-ray according to the configuration and the runtime environment parameters defined in the knowledge base, and perform standardized label mapping according to the label mapping dictionary to normalize the pixel coordinates of the input whole spine X-ray to a preset standard coordinate space. The anatomical feature extraction module is configured to receive the standardized labels and coordinate-normalized whole spine X-ray data output by the data adaptation and standardization module, and output the structural feature measurement values ​​required for consistency inspection. The consistency check engine is configured to perform quality checks based on the structural feature measurement values ​​output by the anatomical feature extraction module and the check rules defined in the configuration and knowledge base. The report output module is configured to generate a quality report containing quality levels based on the inspection results output by the consistency inspection engine.

2. The configurable quality control system for the whole spinal X-ray film according to claim 1, characterized in that, The specific configuration and knowledge base are set as follows: The anatomical standards, label mapping dictionary, inspection rules, and runtime environment parameters are defined through the preset configuration file. The runtime environment parameters include the annotation software type, the root directory path of the dataset, the storage structure of the annotation files in the file system, and the annotation orientation of the image. The anatomical criteria include the standard spinal sequence; The label mapping dictionary includes the actual label name, standard label, and the mapping relationship between non-standard labels in the actual label name and standard labels; the types of marker points and key lines and their label mappings, as well as all statistics and position thresholds used for automatic verification checks; The inspection rules include inspection items, absolute position judgment thresholds, anatomical reference relationships of relative positions, definitions of optional annotations, and annotation types.

3. The configurable quality control system for the whole spine X-ray film according to claim 1, characterized in that, The data adaptation and standardization module is specifically configured as follows: The pixel coordinates of the full spine X-ray are normalized to a preset standard coordinate space; wherein, the input full spine X-ray has a width of Height is Image, normalized scale Normalized coordinate points The calculation method is as follows If the input full spine X-ray data lacks a defined image size, dynamic normalization is performed; where the normalization ratio is determined by the coordinate range of the input full spine X-ray. and Among all valid pixels in the input full-spine X-ray coordinate system, This represents the maximum X-axis coordinate value. This represents the minimum coordinate value on the X-axis. This represents the maximum Y-axis coordinate value. This represents the minimum Y-axis coordinate value.

4. The configurable quality control system for the whole spinal X-ray film according to claim 1, characterized in that, The anatomical feature extraction module is specifically configured as follows: For the segmentation mask of the vertebral bodies in a full spine X-ray, calculate the centroid coordinates and bounding box of each vertebral body; Extract the normalized coordinates of the marker points and key lines, and calculate the center points of the marker points and key lines; Relationship data generated based on the center point of the key marker points.

5. The configurable quality control system for the whole spinal X-ray film according to claim 4, characterized in that, The anatomical feature extraction module is further configured to: If a preset vertebra is missing from the input full spine X-ray, the simulated center point of the preset vertebra is derived based on the missing preset vertebra's number in the standard spine sequence. The expected value of the simulated center point on the normalized Y-axis is used as a reference point for relative position examination in the case of a missed vertebral marker. Among them, the standard spinal sequence index The center point of the pre-set vertebral body is missing. Expected value on the Y-axis in It represents the Y-axis coordinates of the center point of the highest effective vertebral body observed in the current spinal sequence. It represents the Y-axis coordinates of the center point of the lowest effective vertebral body observed in the current spinal sequence. yes and The total number of vertebrae that should theoretically exist between them is used as the length of the reference spinal sequence.

6. The configurable quality control system for the whole spinal X-ray film according to claim 2, characterized in that, The consistency check engine is specifically configured as follows: Using the centroid coordinates of the vertebral bodies along the normalized Y-axis and the corresponding labels, we check whether the spinal sequence satisfies the medical order of increasing from top to bottom and determine the topological order. Identify omissions and errors, and determine the completeness and existence of vertebral segmentation and labeling. Check the correctness of the names and orientation assignments of paired markers, including applying absolute spatial constraints for lateral orientation checks; in Indicates the normalized median; A relative consistency check is performed based on the relative position and distance between the markers, whereby... in It is a predefined position tolerance, where and These represent the Y-axis coordinates of the center points of the i-th and (x+1)-th vertebrae in the spinal sequence, respectively, in the normalized image coordinate system.

7. The whole spinal X-ray film configurable quality control system according to claim 6, characterized in that, The report output module is specifically configured as follows: A quality report is output based on the errors found in the inspection results and the preset error levels.

8. The whole spinal X-ray film configurable quality control system according to claim 7, characterized in that, The preset error levels include failure and warning; the failure level includes failure of the vertebral topology sequence check, laterality check and integrity check; the warning level includes warnings for failure of relative consistency check and non-critical additional labels.

9. A method for quality control of whole-spine X-ray data, characterized in that, A configurable quality control system for the whole spine X-ray as described in any one of claims 1-8 includes: Obtain a full spine X-ray; wherein the full spine X-ray includes images of the vertebral bodies and labeling of the vertebral bodies; The annotation data of the whole spine X-ray is read according to the runtime environment parameters defined by the configuration and knowledge base, and the label mapping is standardized according to the label mapping dictionary to normalize the pixel coordinates of the whole spine X-ray to the preset standard coordinate space. Based on the data from the whole spine X-ray after standardization of labels and coordinate normalization, output the structural feature measurements required for consistency checks; Quality inspection is conducted based on the structural feature measurements and inspection rules. Based on the inspection results, a quality report containing the quality level is generated.

10. The method for quality control of whole-spine X-ray data according to claim 9, characterized in that, Also includes: The operating environment parameters, as well as the anatomical standards, label mapping dictionary, and examination rules required for evaluating whole spine X-rays, are set through a preset configuration file.

Images