A method and system for intelligent normalization and sorting of DICOM sequences

Abstract

This application relates to the technical field of medical image processing, and in particular to a method and system for intelligent normalization and sorting of DICOM sequences. The method includes: acquiring the original DICOM file; parsing the key DICOM tags of the original DICOM file to obtain orientation tags and position information; performing normalization based on the orientation tags and orientation correction according to the patient coordinate system; projecting the coordinates of all slices to the normal vector direction to reconstruct the slice order and statistically analyzing the slice interslice spacing sequence to estimate the true interslice spacing based on the difference between adjacent projections; identifying and removing abnormal slices based on the anomaly comparison target tags and slice interslice spacing sequence; for a set of multiple sequences from the same examination, extracting the sorting feature vector corresponding to each sequence and selecting a sorting rule based on scene complexity to obtain the final sorting key used to characterize the sorting results of the phase dimension and / or time dimension. This application effectively solves the problems of inconsistent orientation, disordered slice order, and difficulty in identifying temporal phase in existing DICOM sequences.

Description

Technical Field

[0001] This application relates to the technical field of medical image processing, and in particular to a method and system for intelligent normalization and sorting of DICOM sequences. Background Technology

[0002] DICOM, or Digital Imaging and Communications in Medicine, defines a medical image format for data exchange that meets clinical needs. It standardizes communication between medical imaging equipment (such as X-ray, CT, MRI, and ultrasound) and other systems, covering aspects such as acquisition, archiving, communication, display, and retrieval of medical digital images. DICOM files store data in "data elements," each containing an identifier, data type, and value. Multiple data elements constitute a "dataset," which can be nested within "sequences" to form complex structures. Each DICOM file typically contains an image and its associated metadata, such as patient information (name, gender, age, etc.), equipment information, and examination parameters.

[0003] In clinical radiology departments, imaging cloud platforms, and AI algorithm development, a large number of DICOM sequences come from different hospitals, different equipment suppliers, and different scanning protocols. Therefore, the following prominent problems exist in actual operation:

[0004] 1. Inconsistent sequence directions:

[0005] Different devices use different patient positions (head-in / feet-in, supine / prone), resulting in inconsistencies between the image row and column orientation and the anatomical position; the combination of fields such as ImageOrientationPatient (patient image orientation label) and PatientPosition (patient position label) in the DICOM head is complex, and traditional processing is prone to parsing errors; some sequences have problems with head-to-feet reversal and left-to-right flipping, affecting 3D reconstruction and subsequent analysis;

[0006] 2. The layer order and spacing are disordered:

[0007] Slices in the same inspection may be sorted by InstanceNumber (instance number) but not by SliceLocation (slice location label), and there may even be duplicates or missing slices; the SpacingBetweenSlices is inconsistent with the true geometric spacing, resulting in measurement errors and reconstruction artifacts; mixed sequences (such as scans from different FOVs or different locations) are misclassified as a single sequence, causing spatial discontinuities;

[0008] 3. Missing or non-standard time dimension and phase information:

[0009] In dynamic contrast-enhanced scans (such as three-phase liver imaging and coronary CTA), the labeling of different phases is inconsistent and difficult to distinguish automatically; the time frame sorting in cardiac or functional imaging relies on vendor-proprietary labels, resulting in poor compatibility; traditional methods often can only sort by file time or simple numbers, and cannot recover the true time sequence.

[0010] 4. Difficulty in integrating data from multiple centers and devices:

[0011] Different vendors use DICOM tags differently, and the same meaning may correspond to different proprietary tags; some checks have protocol configuration errors or technician operation errors, resulting in missing or inconsistent metadata; a robust normalization and sorting algorithm is needed to recover the correct geometric and temporal order under untrusted metadata conditions. Summary of the Invention

[0012] To address the problems of inconsistent orientation, disordered layer order, and difficulty in identifying time phase in existing DICOM sequences, this application provides a method and system for intelligent normalization and sorting of DICOM sequences.

[0013] Firstly, this application provides a method for intelligent normalization and sorting of DICOM sequences, employing the following technical solution:

[0014] A method for intelligent normalization and sorting of DICOM sequences includes the following steps:

[0015] Obtain the original DICOM file;

[0016] The key DICOM tags of the original DICOM file are parsed to obtain orientation tags and body position information. Based on the orientation tags, normalization based on the patient coordinate system is performed, and the orientation of the normalization result is corrected in combination with the body position information.

[0017] Based on geometric vector projection, the coordinates of all normalized slices are projected to the normal vector direction. The layer order is reconstructed based on the result of geometric vector projection, and the slice layer spacing sequence is statistically analyzed to estimate the true layer spacing based on the difference between adjacent projections.

[0018] Abnormal slices are identified and removed based on the anomaly comparison target label and the slice layer spacing sequence;

[0019] For a set of multiple sequences under the same inspection, the sorting feature vector corresponding to each sequence is extracted and sorting rules are selected in combination with scene complexity to obtain the final sorting key used to characterize the sorting results of the phase dimension and / or time dimension. The sorting feature vector includes time features, phase features, and intensity features.

[0020] In some embodiments, after obtaining the original DICOM file, the following steps are also included:

[0021] Obtain the check number and serial number, and group the original DICOM files with the same check number and serial number into the same initial sequence;

[0022] For each of the initial sequences, calculate the field of view size and the standard deviation of the slice center coordinates for all slices;

[0023] If the coefficient of variation of the field of view is greater than a preset value or the standard deviation is greater than a preset value, it is determined to be a mixed sequence;

[0024] K-means clustering is used to split the mixed sequence into slices to obtain several subsequences. Temporary check numbers are reassigned to the subsequences, and the original check numbers and splitting criteria are recorded.

[0025] In some embodiments, key DICOM tags of the original DICOM file are parsed to obtain orientation tags and body position information. Normalization based on the orientation tags is performed using the patient coordinate system, and the normalization result is corrected for orientation according to the body position information. This includes the following steps:

[0026] Obtain the orientation labels of the patient image and extract the row vectors and column vectors, construct the orientation cosine matrix based on the row vectors and column vectors, and map the cosine matrix to the patient coordinate system.

[0027] The normal vector is calculated by cross-multiplying the row vector and the column vector and then normalized to a unit vector.

[0028] Obtain the patient coordinate system type, and obtain the corresponding direction requirements based on the patient coordinate system type;

[0029] Based on the cosine matrix and the normal vector, it is determined whether the direction requirement is met, and if not, the direction cosine matrix is ​​corrected.

[0030] When the orientation label of the patient image is missing, extract the patient location label and the patient position label;

[0031] The normal vector is estimated based on the coordinate distribution represented by the patient location label, and the row vector, column vector, and normal vector are estimated based on the patient position label as position information.

[0032] In some embodiments, the coordinates of all normalized slices are projected onto the normal vector direction based on geometric vector projection. The layer order is reconstructed based on the result of the geometric vector projection, and the slice layer spacing sequence is statistically analyzed to estimate the true layer spacing based on the difference between adjacent projections. This includes the following steps:

[0033] The geometric center coordinates of the slice are calculated based on the patient location labels;

[0034] Project the geometric center coordinates of all slices onto the normal vector direction, and multiply the geometric center coordinates by the normal vector to obtain the projection distance;

[0035] The projection distances are sorted in ascending order to generate a sorting index to rearrange the slices and determine the true layer order.

[0036] The difference in the projected distance between adjacent slices under the true layer sequence is calculated to obtain the slice layer spacing sequence, and the median, mode or mean of the slice layer spacing sequence is taken as the true layer spacing.

[0037] In some embodiments, after estimating the actual interlayer spacing, the following steps are further included:

[0038] Extract the claimed interlayer spacing from the original DICOM file;

[0039] Calculate the relative error between the actual interlayer spacing and the claimed interlayer spacing and determine whether it is greater than a preset value;

[0040] If so, mark it as an abnormal spacing and record it in the quality control log.

[0041] In some embodiments, identifying abnormal slices based on the anomaly comparison target label and the slice layer spacing sequence includes the following steps:

[0042] The algorithm uses density clustering to identify principal spacing clusters and determines whether the slice layer spacing sequence is within the range of the principal spacing clusters. If not, it is marked as a suspected missing layer or multi-layer boundary.

[0043] If so, extract the layer thickness label and / or pixel spacing label of all slices and calculate the difference with the mainstream layer thickness value and / or mainstream pixel spacing value respectively;

[0044] Determine if the difference is greater than the difference threshold; if so, mark it as a metadata anomaly slice.

[0045] If not, compare the Euclidean distance and image overlap of adjacent slices and determine whether the corresponding threshold is met simultaneously. If so, mark it as a duplicate slice.

[0046] Record the abnormal information, which includes the abnormal type, slice number, and abnormal parameters.

[0047] In some embodiments, the abnormal slices are removed, including the following steps:

[0048] The degree of abnormality is determined based on the abnormal parameters in the abnormal information.

[0049] If the anomaly level is severe, then delete the slice and update the sorting index;

[0050] If the degree of anomaly is minor, the slice is retained and its weight is reduced.

[0051] In some embodiments, the sorting feature vector corresponding to each sequence is extracted, and a sorting rule is selected based on scene complexity to obtain the final sorting key used to characterize the sorting results in the phase dimension and / or time dimension, including the following steps:

[0052] The analysis determines whether the extracted sorting feature vector satisfies at least one of the following scenario conditions: no missing / repeated / skipped timestamps, the existence of standard time series labels that are continuous, and the parsable supplier-private phase field.

[0053] If so, it is defined as a simple scenario, and a sorting rule is selected based on the scenario conditions to directly generate the final sorting key through time features and / or phase features;

[0054] If not, it is defined as a complex scene, and the ROI region in the slice is identified and the intensity feature vector is calculated. The predicted phase is generated by combining the intensity change pattern to generate the final sorting key.

[0055] In some embodiments, the time feature includes a collection time tag and a content time tag. If both the collection time tag and the content time tag exist, the collection time tag is used first. If the timestamp corresponding to the collection time tag is missing / repeated / skipped, the content time tag is used.

[0056] Secondly, this application provides a DICOM sequence intelligent normalization and sorting system, which adopts the following technical solution:

[0057] A DICOM sequence intelligent normalization and sorting system is provided to implement the above method.

[0058] The technical solutions provided by the embodiments of this application have the following technical effects:

[0059] 1. Significantly improved geometric consistency: All sequences are unified to the same patient coordinate system, significantly reducing the distortion rate of 3D reconstruction and providing a stable foundation for AI segmentation and measurement;

[0060] 2. Automatic sequence quality screening: Through an abnormal slice detection mechanism, unusable sequences and problematic slices are automatically identified, reducing backend failures and manual troubleshooting workload;

[0061] 3. Strong multi-center compatibility: Designed for heterogeneous devices and non-standard metadata, it can cover mainstream suppliers and multiple batches of equipment, and is a basic component of image cloud platforms and multi-center projects.

[0062] Improve clinical and model iteration efficiency: In a large-scale data environment, it can automatically process a large number of examinations from different hospitals and equipment, significantly shortening data preparation time and accelerating AI model iteration. Attached Figure Description

[0063] Figure 1 This is a schematic diagram illustrating the steps of a DICOM sequence intelligent normalization and sorting method provided in this embodiment.

[0064] Figure 2 This is a logical diagram illustrating the direction normalization process provided in the embodiments of this application.

[0065] Figure 3 This is a logical diagram illustrating the reconstruction layer sequence and estimated layer spacing provided in the embodiments of this application.

[0066] Figure 4 This is a logical diagram illustrating the filtering and removal of abnormal slices provided in the embodiments of this application.

[0067] Figure 5 This is a logical diagram illustrating the sorting of multiple sequences provided in this application embodiment. Detailed Implementation

[0068] To better understand the purpose, technical solutions, and advantages of this application, it has been described and illustrated below with reference to the accompanying drawings and embodiments. However, those skilled in the art should understand that this application can be implemented without these details. In some cases, to avoid obscuring various aspects of this application due to unnecessary description, well-known methods, processes, systems, components, and / or circuits already described at a higher level will not be elaborated upon. It will be apparent to those skilled in the art that various modifications can be made to the embodiments disclosed in this application, and the general principles defined in this application can be applied to other embodiments and application scenarios without departing from the principles and scope of this application. Therefore, this application is not limited to the illustrated embodiments, but conforms to the broadest scope consistent with the scope of protection claimed in this application.

[0069] It should be noted that the descriptions of these embodiments are for the purpose of aiding understanding the present invention, but do not constitute a limitation thereof. Furthermore, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0070] For software implementations, the techniques described herein can be implemented using modules (e.g., programs, functions, etc.) that perform the functions described herein. The software code can be stored in memory units and executed by a processor. The memory units can be implemented within or outside the processor; in the latter case, the memory units can be communicatively coupled to the processor via various methods known in this art.

[0071] In the description of this application, "several" means one or more, "more than" means two or more, "greater than," "less than," and "exceeding" are understood to exclude the stated number, while "above," "below," and "within" are understood to include the stated number. The use of "first" and "second" in the description is merely for distinguishing technical features and should not be construed as indicating or implying relative importance, or implicitly indicating the number of indicated technical features, or implicitly indicating the order of the indicated technical features.

[0072] In the description of this application, the terms "one embodiment," "some embodiments," "illustrative embodiment," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any one or more embodiments or examples.

[0073] like Figure 1 As shown in the figure, this application discloses a method for intelligent normalization and sorting of DICOM sequences, including the following steps:

[0074] S100, Obtain the original DICOM file, which corresponds to a single sequence.

[0075] S200 parses the key DICOM tags of the original DICOM file to obtain orientation tags and position information, performs normalization based on the orientation tags and patient coordinate system, and corrects the orientation of the normalization result by combining the position information.

[0076] To address the issue of inconsistent DICOM sequence orientations under different devices and patient positions, orientation-related labels and patient position information are read from the original DICOM file. The vectors corresponding to the orientation are then used to uniformly transform the row and column orientations of the slices to a standard patient coordinate system. The patient coordinate system is a spatial coordinate system defined based on human anatomy, commonly including LPS (Left-Posterior-Superior) and RAS (Right-Anterior-Superior).

[0077] Meanwhile, the LPS and RAS coordinate systems are inherently inverted. Furthermore, due to the different body positions used by different devices, patients may adopt different postures during examination, such as HFS (head-first supine), FFS (feet-first supine), HFP (head-first prone), and FFP (feet-first prone). Therefore, the normalized result may contain issues such as head-to-feet reversal or left-to-right rotation. Thus, it is necessary to further correct the orientation of the final normalized result by combining specific direction labels and body position labels.

[0078] S300 projects the coordinates of all normalized slices onto the normal vector direction based on geometric vector projection. The layer order is reconstructed based on the result of geometric vector projection, and the slice layer spacing sequence is statistically analyzed to estimate the true layer spacing based on the difference between adjacent projections.

[0079] The traditional method of sorting by instance number (InstanceNumber) is prone to failure in mixed sequences or abnormal export scenarios. Therefore, in this embodiment, the layer order is rearranged based on the projection scalar under geometric vector projection, and the actual interlayer spacing is estimated based on the projection difference between adjacent slices after the layer order is rearranged.

[0080] At the same time, the corrected interlayer spacing will be written into the internal metadata for subsequent 3D reconstruction and quantitative measurement.

[0081] S400 identifies and removes abnormal slices based on the anomaly comparison target label and slice layer spacing sequence.

[0082] Based on the reconstruction layer sequence, an abnormal slice detection module is introduced to automatically identify and remove slices under various abnormal conditions, such as obvious abnormal geometric spacing between slices, abnormal layer thickness, abnormal pixel spacing, and duplicate slices.

[0083] Furthermore, the handling of abnormal slides follows the principle of "preserving geometric continuity as much as possible," employing different processing methods for slides of different abnormality levels. All abnormal information and processing results are written to the quality control log, linking with a multi-layered quality control rule engine to facilitate doctors or engineers in tracing the source of problems.

[0084] S500, for a set of multiple sequences under the same inspection, extracts the sorting feature vector corresponding to each sequence and selects a sorting rule based on the scene complexity to obtain the final sorting key used to characterize the sorting results of the phase dimension and / or time dimension. The sorting feature vector includes time features, phase features, and intensity features.

[0085] In dynamic contrast-enhanced CT, cardiac CT, or functional MRI, multiple sequence sets are often present, making the temporal or phase order particularly important. However, the labels provided by different vendors are often inconsistent (e.g., different vendors use different arterial phase label fields), leading to missing or incorrect labels in some scenarios. Therefore, this application proposes a multi-feature fusion intelligent sorting algorithm that integrates time labels, phase labels, and image intensity features to ensure the accuracy of the temporal / phase order.

[0086] Furthermore, in this embodiment, a specific scene complexity is constructed, and different sorting logics are matched according to the scene complexity. Scene complexity mainly focuses on whether the above-mentioned features exist. If all of them exist, the scene is relatively simple, and sorting is performed directly by combining the existing labels with the sorting rules. However, when there are missing contents such as time and phase, the scene is more complex, and further comprehensive inference and analysis are performed based on the analyzed image intensity features.

[0087] The above method unifies DICOM sequences from different devices and patient positions into a standard patient coordinate system, automatically corrects slice order, accurately estimates the true interslice spacing, identifies and removes abnormal slices, and, in the absence of standardized labels, intelligently infers phase or temporal order through multi-feature fusion, providing a stable and consistent input data foundation for 3D reconstruction, AI analysis, and anatomical coverage assessment. This solves problems such as inconsistent DICOM sequence orientation, disordered slice order, and difficulty in identifying temporal phase.

[0088] In other embodiments, after obtaining the original DICOM file and before normalization, the DICOM file first needs to undergo corresponding preprocessing and grouping, including the following steps:

[0089] S110, obtain the check number and serial number, and group the original DICOM files with the same check number and serial number into the same initial sequence.

[0090] First, the system extracts the check ID (StudyInstanceUID) and serial number (SeriesInstanceUID) from the DICOM file.

[0091] An examination number is an identifier for an examination level, typically generated by the RIS (Research Information System), used to identify a specific examination order or instance for a patient. It uniquely identifies an examination throughout the system, representing all relevant image acquisitions for a single patient examination. For example, different examinations for the same patient (such as CT and MRI) will be assigned different examination numbers.

[0092] A sequence number is a sequence-level identifier used to identify a specific sequence within a single examination. An examination may include multiple sequences, each corresponding to a specific imaging modality or parameter setting. Sequence numbers are unique within a single examination, but can be repeated across different examinations. They help organize and differentiate different image sets within the same examination.

[0093] Therefore, this application first classifies the corresponding original DICOM files into an initial sequence based on the fact that both the serial number and the check number are consistent.

[0094] S120, calculate the field size and standard deviation of the slice center coordinates for each initial sequence.

[0095] For each initial sequence, the field of view (FOV) is calculated for all slices. FOV refers to the area of ​​the body imaged during a scan; it is the extent of the patient's anatomy captured in the image. It is typically expressed in centimeters or millimeters and can be adjusted by the operator before scanning. A larger FOV will capture a larger area of ​​the body, while a smaller FOV will capture a smaller area with higher spatial resolution.

[0096] In this embodiment of the application, the method for calculating the field of view (FOV) is as follows:

[0097] FOV_x=PixelSpacing_x*Columns, FOV_y=PixelSpacing_y*Rows.

[0098] Among them, PixelSpacing is represented by the pixel spacing label, with label number (0028, 0030); Columns is the image column number label, with label number (0028, 0011); and Rows is represented by the image column number label, with label number (0028, 0010).

[0099] The standard deviation is represented as the standard deviation of the Euclidean distance between the center coordinates of adjacent slices.

[0100] S130, if the coefficient of variation of the field of view is greater than the preset value or the standard deviation is greater than the preset value, it is determined to be a mixed sequence.

[0101] This application requires two separate determinations to determine whether a sequence is a mixed sequence:

[0102] Determine whether the coefficient of variation (where the coefficient of variation is represented as the standard deviation of the field of view / mean of the field of view) of FOV_x or FOV_y exceeds 30%;

[0103] The standard deviation of the Euclidean distance is determined to be more than twice the average inter-layer spacing.

[0104] If any of the above conditions are triggered, the sequence is determined to be a mixed sequence.

[0105] S140. K-means clustering is used to split the mixed sequence into slices to obtain several subsequences. Temporary check numbers are reassigned to the subsequences and the original check numbers and splitting criteria are recorded.

[0106] If the sequence is determined to be a mixed sequence, it is further split into segments using K-means clustering. The clustering features are FOV_x, FOV_y, and the center coordinates of the slice (x, y, z). The specific splitting method is implemented using the current conventional clustering algorithm.

[0107] Each subsequence after splitting will be reassigned a temporary check number, and the original check number and the basis for splitting will be recorded for subsequent traceability.

[0108] This solves the spatial discontinuity caused by the misidentification of mixed sequences corresponding to scans at different FOVs or positions as a single sequence.

[0109] like Figure 2 As shown, in some other embodiments, key DICOM tags of the original DICOM file are parsed to obtain orientation tags and body position information. Normalization based on the orientation tags is then performed using the patient coordinate system, and the normalization result is corrected for orientation according to the body position information. This includes the following steps:

[0110] S210: Obtain the orientation labels of the patient image and extract the row and column vectors. Construct an orientation cosine matrix based on the row and column vectors and map the cosine matrix to the patient coordinate system.

[0111] Based on the parsed key DICOM tags, several orientation-related tags were obtained, with a focus on acquiring and utilizing the patient image orientation tag ImageOrientationPatient (0020, 0037).

[0112] The patient image orientation labels are stored in 6 floating-point numbers ( , , , , ),in( , Let u be the unit vector u in the row direction of the image. , , ) is the unit vector v in the column direction of the image.

[0113] The row and column direction vectors of each slice are transformed into a unified patient coordinate system (LPS coordinate system or RAS coordinate system). In this embodiment, the LPS coordinate system is used by default. It is defined as follows: the x-axis is along the left and right direction of the human body, with positive to the left (negative to the left in the RAS coordinate system); the y-axis is along the front and back direction of the human body, with positive to the back (positive to the front in the RAS coordinate system); and the z-axis is along the up and down direction of the human body, with positive to the up.

[0114] First, a cosine matrix is ​​constructed using row and column vectors. The first row of the matrix is ​​row vector u, and the second row is column vector v. The elements of the cosine matrix are the direction cosines of these two vectors with respect to the three axes of the standard coordinate system (LPS / RAS) (i.e., the cosine values ​​of the angles between the vectors and the coordinate system), which quantifies the correspondence between the rows and columns of the image and the anatomical directions.

[0115] For example, if the x-axis cosine of the row vector u is 1 and the others are 0, it means that the row direction of the slice is completely aligned with the left and right directions of the human body; if the cosine value is -1, it means that the row direction is opposite to the left and right directions (left and right reversed).

[0116] S220 calculates the normal vector by performing a cross product of the row vector and the column vector and normalizes it to a unit vector.

[0117] Furthermore, the normal vector n is calculated through the cross product operation of uxv. The normal vector is perpendicular to the slicing plane and reflects the layer thickness direction of the slice. At the same time, the normal vector is normalized (making its magnitude 1) to eliminate scaling differences between devices and ensure the consistency of subsequent projection calculations.

[0118] S230, obtain the patient coordinate system type, and obtain the corresponding direction requirements based on the patient coordinate system type.

[0119] The patient coordinate system type is represented as either LPS or RAS coordinate system. Based on the established transformation table of device coordinates to standard coordinates, and combined with preset mapping rules, the coordinate systems of different devices are uniformly mapped to the coordinate system. Simultaneously, because the x-axis and y-axis directions of the LPS and RAS coordinate systems are defined differently, the corresponding direction requirements need to be obtained through the coordinate system type. For example, if the x-axis cosine value of a row vector u is 1, and all others are 0, it represents normal left and right in the LPS coordinate system, but is represented as left and right reversed in the RAS coordinate system.

[0120] S240 determines whether the direction requirement is met based on the cosine matrix and normal vector, and corrects the direction cosine matrix if it is not met.

[0121] The direction requirements of the corresponding coordinate system are determined by the values ​​of each element in the cosine matrix and the normal vector. If the requirements are not met, the elements in the cosine matrix are multiplied by -1 (to adjust the sign) to flip the direction.

[0122] The overall directional requirements are that the sign of the x-component of the row vector u must be consistent with the x-axis definition of LPS / RAS, and the z-component of the normal vector n must conform to the vertical anatomical direction of the human body. During correction, the x or y vectors need to be inverted to update the vectors u, v, and n.

[0123] S250: When the orientation label of the patient image is missing, extract the patient location label and the patient position label.

[0124] In some special scenarios, patient image orientation labels may be missing and cannot be extracted. In this case, we will further extract the patient position label ImagePositionPatient (0020, 0032) and the patient position label PatientPosition (0018, 5100) from the orientation-related labels.

[0125] Patient location labels are used to obtain the geometric center coordinates (x, y, z) of each slice, which are used for subsequent auxiliary verification of orientation. Patient position labels are used to characterize position information, clarify the original spatial background, and also serve as a basis for orientation when row vectors and column vectors are missing.

[0126] S260 estimates the normal vector based on the coordinate distribution represented by the patient's location label, and estimates the row vector, column vector, and normal vector based on the patient's position label represented by the position information.

[0127] When the orientation labels of patient images are completely missing, the coordinate distribution (the changing trends of x, y, z) of the patient location labels of multiple slices is combined to help estimate the stacking direction of the slices, that is, the normal vector, and then supplement the derivation of the approximate direction of the row / column orientation vectors.

[0128] Furthermore, patient location labels can also be used as cross-validation for orientation correction when patient image orientation labels are not completely missing. When the orientation vector derived from patient image orientation labels is ambiguous, partially missing, or unreliable, the orientation normalization result can be verified by the trend of changes in patient location label coordinates in adjacent slices.

[0129] For example, if the derived slice normal vector points to the "head direction" (positive z-axis direction of the patient coordinate system), but the z-coordinate of the patient position label of the adjacent slice continues to be effective (not in line with the increasing z-axis direction of the head direction), then it can be determined that the normal vector direction is incorrectly derived, and the z-axis flip can be corrected.

[0130] For example, if the row vector is derived to the left (positive x-axis direction of the LPS coordinate system), but the x-coordinate of the adjacent patient location label continues to decrease, the row direction vector sign error can be corrected and the left-right flip can be corrected.

[0131] By using position labels, when the orientation labels of the patient's image are completely missing, the specific content of the position can be represented as image information as auxiliary information. For example, if the current position label is HFS (head first, supine), the default row direction of the slice is left-right, the column direction is front-back, and the normal vector direction is head-to-foot. However, when the position label is FFS (feet first, supine), the default z-axis direction is feet-to-head, and the z-axis direction needs to be flipped. In this way, the direction of the row vector, column direction, and normal vector under ideal conditions can be predicted by using the position information.

[0132] It is important to note that when patient image orientation labels are missing, the corresponding normalization results are configured with low confidence and labeled when estimating row, column, and normal vectors using patient location and position labels, so as to inform technical personnel to conduct quality control review of the results.

[0133] Furthermore, there may be slight numerical errors in the direction cosine (such as ±1 floating-point error). In this case, tolerance processing is required to improve robustness. Specifically,

[0134] The tolerance threshold is set to 1e-3. When the direction cosine value is in the range of [-1-1e-3, -1+1e-3], it is regarded as -1. When it is in the range of [1-1e-3, 1+1e-3], it is regarded as 1. In other cases, the original value is retained and normalized to avoid direction judgment errors caused by floating-point precision errors of the device.

[0135] Finally, the corrected orientation information is stored in the internal metadata structure for subsequent 3D reconstruction and coordinate conversion, without directly modifying the original DICOM file.

[0136] like Figure 3 As shown, in some other embodiments, the coordinates of all normalized slices are projected onto the normal vector direction based on geometric vector projection. The layer order is reconstructed based on the result of geometric vector projection, and the slice layer spacing sequence is statistically analyzed to estimate the true layer spacing based on the difference between adjacent projections. This includes the following steps:

[0137] S310, calculate the geometric center coordinates of the slice based on the patient location label.

[0138] The geometric center coordinates p_i of each slice are calculated using patient location labels. Specifically,

[0139] p_i=ImagePositionPatient+0.5*(Columns* *u+Rows* *v), where,

[0140] Columns represents the number of image columns, and Rows represents the number of image rows. This is the first value of the image spacing label PixelSpacing (pixel spacing in the x-direction). This is the second value of the image spacing label PixelSpacing (pixel spacing in the y-direction).

[0141] The patient location label represents the three floating-point numbers (x, y, z) stored in the image as the spatial coordinates of the top left corner of the image in the patient coordinate system. The above calculation formula obtains the accurate location of the slice center by superimposing the top left corner coordinates with the spatial offset of half the image.

[0142] S320: Project the geometric center coordinates of all slices onto the direction of the normal vector, and multiply the geometric center coordinates by the normal vector to obtain the projection distance.

[0143] The normal vector n represents the core direction of the slice stack and is the reference direction for layer ordering. Therefore, in this embodiment, the geometric center coordinates of all slices are projected onto the normal vector direction to obtain the scalar d_i. It is calculated by d_i = p_i·n (dot product).

[0144] In this embodiment, the scalar is the projection distance, which represents the projection distance from the center of the slice to the origin of the patient's coordinate system in the direction of the normal vector. The distribution of d_i values ​​of slices in the same sequence should show a linear trend. The dot product calculation uses floating-point precision to avoid data overflow or precision loss.

[0145] S330 generates a sorting index by dividing several projected distances into ascending order to rearrange several slices and determine the true layer order.

[0146] Sort d_i in ascending order to adjust the position of each slice. The final sort is the actual layer sort. After sorting, a layer order index is generated to record the position of each slice in the normalized sequence.

[0147] S340, calculate the difference in projected distances between adjacent slices under the true layer sequence to obtain the slice interlayer spacing sequence, and use the median, mode, or mean of the slice interlayer spacing sequence as the true interlayer spacing.

[0148] After rearranging the slice layer order, it is necessary to further evaluate the actual interlayer spacing. In this embodiment, for the sorted layer order sequence d_i, the projection difference Δd_i=d_[i+1]-d_i (i=1,2,...,N-1, N is the total number of slices) is calculated to obtain the slice interlayer spacing sequence.

[0149] Secondly, outliers in Δd_i are removed based on the 3σ principle (i.e., Δd_i exceeding the mean ± 3 times the standard deviation), and the mean, median, or mode of the remaining Δd_i are used as the true interstitial spacing. The specific value to be used can be selected based on the actual situation.

[0150] After estimating the actual interlayer spacing, cross-validation can be used to identify and flag anomalies in the interlayer spacing metadata. Specifically,

[0151] In other embodiments, after estimating the actual interlayer spacing, the following steps are also included:

[0152] S341, Extract the claimed interlayer spacing from the original DICOM file.

[0153] Claimed inter-slice spacing is the "theoretical inter-slice spacing" explicitly marked in DICOM metadata, which is the physical spacing between two adjacent image slices under ideal conditions preset by the device or scanning protocol.

[0154] S342, calculate the relative error between the actual interlayer spacing and the claimed interlayer spacing and determine whether it is greater than the preset value.

[0155] S343, if so, mark it as a spacing anomaly and record it in the quality control log.

[0156] Analyze the relationship between the relative error between the actual interlayer spacing and the claimed interlayer spacing and the preset value. When the error is greater than the preset value, the actual interlayer spacing is marked as an interlayer spacing metadata anomaly and recorded in the quality control log, but the actual interlayer spacing will still be used as the basis for subsequent processing. If the error is not greater than the preset value, it indicates that the cross-validation has passed.

[0157] The preset value in this application embodiment is 10%, and the relative error is: |actual interlayer spacing - claimed interlayer spacing| / actual interlayer spacing.

[0158] Furthermore, for obvious tomographic scans (such as two-segment scans of the chest and abdomen), different boundaries are identified through abrupt changes in slice spacing, generating spatial discontinuity markers. Specifically,

[0159] Criteria for determining abrupt changes in interlayer spacing: If the difference between two consecutive Δd_i exceeds 50% of the average principal spacing, it is considered a spatial discontinuity. For example, if the Δd_i of the first 100 slices is around 5mm (principal spacing 5mm), and the Δd_i of the 101st slice suddenly changes to 20mm, then the 100th and 101st slices are marked as a spatial discontinuity and split into two contiguous blocks (Block1: slices 1-100, Block2: slices 101-N), and the interlayer spacing is estimated for each contiguous block.

[0160] Based on the reconstruction layer sequence, an abnormal slice detection step is introduced to automatically identify and remove the following three common anomalies:

[0161] 1. Slices with significantly abnormal (too large or too small) geometric spacing from other slices in the same sequence;

[0162] 2. Slices whose slice thickness and / or pixel spacing labels differ significantly from mainstream values;

[0163] 3. Slices that are obviously repeated (their geometric positions are almost the same or the overlap is too high).

[0164] like Figure 4 As shown, specifically, in some other embodiments, identifying abnormal slices based on anomaly comparison target labels and slice layer spacing sequences includes the following steps:

[0165] S410 identifies principal spacing clusters based on density clustering algorithm and determines whether the slice layer spacing sequence is within the principal spacing cluster range. If not, it is marked as a suspected missing layer or multi-layer boundary.

[0166] The density clustering algorithm DBSCAN is used to identify the main spacing clusters (cluster_main). The clustering configuration parameters are set as follows: neighborhood radius eps = actual interlayer spacing * 20%; minimum number of samples min_samples = 3 (at least 3 adjacent differences constitute a cluster); after the clustering is completed, the cluster with the largest number of samples is the main spacing cluster, and its mean is used as the reference value of the main spacing.

[0167] Calculate whether the current Δd_i is within the range of the main spacing cluster. In this application, the range of the main spacing cluster is set to [0.5 times the mean, 2 times the mean]. If it is, continue to the next identification step. If it is not, mark it as a suspected missing layer or multi-layer boundary. At this time, it is necessary to record the value of Δd_i and its corresponding position to complete the recording of abnormal information.

[0168] S420, if so, extract the layer thickness label and / or pixel spacing label of all slices and calculate the difference with the mainstream layer thickness value and / or mainstream pixel spacing value respectively.

[0169] The next step is to identify whether the layer thickness and pixel spacing differ too much from the mainstream values. Therefore, the first step is to determine the size of the mainstream values.

[0170] The mainstream value of the layer thickness label is determined by statistically analyzing the median of all layer thickness labels, with the corresponding difference threshold set to 50%. The mainstream value of the pixel pitch label is determined by statistically analyzing the median of all pixel pitch labels, with the corresponding difference threshold set to 30%.

[0171] S430: Determine if the difference is greater than the difference threshold. If so, mark it as a metadata anomaly slice.

[0172] Check if the difference between the layer thickness label and / or pixel spacing label of each slice and the corresponding mainstream value is greater than the corresponding difference threshold. If it is within the tolerance, proceed to the next step of identification; if it is not within the tolerance, mark it as a slice with abnormal metadata. At this point, it is necessary to record the abnormal parameters, actual values, mainstream values, etc., to record the abnormal information.

[0173] Furthermore, when multiple consecutive blocks exist under tomographic scanning, it is also necessary to compare the estimated actual inter-slice spacing of each consecutive block with the aforementioned slice thickness labels to identify any configuration errors. Specifically,

[0174] If the relative error between the actual interlayer spacing and the interlayer thickness spacing exceeds a preset value, it will be marked as "interlayer thickness spacing and interlayer spacing mismatch," because the interlayer spacing and interlayer thickness are usually close. In this case, the error type, block number, actual value, label value, etc., will be recorded.

[0175] S440, if not, compare the Euclidean distance and image overlap of adjacent slices and determine whether the corresponding threshold is met simultaneously. If so, mark it as a duplicate slice.

[0176] Finally, it is necessary to identify whether duplicate slices exist. First, the geometric distance between the two slices (the Euclidean distance between the center coordinates of the slices) is calculated to see if it is less than a preset value. In this embodiment, the preset value is 10% of the actual inter-slice spacing. Next, the intersection-to-union ratio (IoU) is used to calculate the ratio of the intersection area to the union area of ​​the two slices' projection regions in the patient coordinate system to obtain the overlap, and it is determined whether the overlap is greater than a preset value. In this embodiment, the preset value for overlap is 90%.

[0177] If both the Euclidean distance and the image overlap are less than a preset value, the slice is considered a duplicate. In this case, the slice number of the duplicate object needs to be recorded.

[0178] S450 records exception information, including exception type, slice number, and exception parameters.

[0179] After recording the abnormal information corresponding to all the above abnormal situations, the cleaned slice sequence and abnormal report are output.

[0180] Furthermore, isolated slices or slices with obvious abrupt changes are directly removed from the main sequence and recorded in the anomaly information.

[0181] An isolated slice is defined as one whose Δd_i, along with both the preceding and following slices, exceeds twice the mean of the primary interval cluster, and which has no other slice forming a contiguous block with it; a skip slice is defined as one whose Δd_i exceeds three times the mean of the primary interval cluster. After removal, the slice number and index of the sequence need to be updated to ensure the continuity of subsequent processing.

[0182] In other embodiments, the removal of abnormal slices includes the following steps:

[0183] S460 determines the degree of abnormality based on the abnormal parameters in the abnormal information.

[0184] S470, if the anomaly level is severe, delete the slice and update the sort index.

[0185] S480: If the abnormality is minor, the slice is retained and the weighting is reduced.

[0186] The severity of the anomaly is determined by the difference between the value of the abnormal parameter and the preset value. Different severity levels result in different decisions regarding whether to remove the slice.

[0187] For example, when the geometric spacing exceeds twice or less than 0.5 times the average of the main spacing clusters, or when there are duplicate slices or a slice thickness difference exceeding 100%, these are considered severely anomalous slices. These slices will cause significant image deviations during the final image reconstruction and need to be directly removed. On the other hand, slices with geometric spacing exceeding 1.2 times the average of the main spacing clusters or slice thickness differences of 30% are considered slightly anomalous slices and do not need to be forcibly removed. Instead, their weight is reduced during subsequent 3D reconstruction or AI analysis.

[0188] In this way, the geometric continuity of the sequence is preserved as much as possible during the process of removing abnormal slices.

[0189] like Figure 5 As shown, in some other embodiments, the sorting feature vector corresponding to each sequence is extracted and sorting rules are selected in combination with scene complexity to obtain the final sorting key used to characterize the sorting results in the phase dimension and / or time dimension, including the following steps:

[0190] S510, analyze whether the extracted sorting feature vector satisfies at least one scenario condition. Scenario conditions include: no missing / repeated / skipped timestamps, the existence of standard time series labels and their continuity, and the parsing of the supplier's private phase field.

[0191] S520, if so, is defined as a simple scenario, and a sorting rule is selected based on the satisfied scenario conditions to directly generate the final sorting key through time features and / or phase features.

[0192] S530, if not, is defined as a complex scene, and the ROI region in the slice is identified and the intensity feature vector is calculated. The predicted phase is generated by combining the intensity change pattern to generate the final sorting key.

[0193] The time features used in this step correspond to the extracted acquisition time tag AcquisitionTime (0008, 0032) and content time tag ContentTime (0008, 0033).

[0194] The acquisition time stamp records the actual time point when the medical image is scanned by the device. It is generated in real time during the scanning process and is based on the device's built-in clock. Theoretically, it can reflect the true acquisition sequence of the slice / frame. The content time stamp records the time point when the medical image content is finally generated and solidified into the DICOM file. It is filled by the device software or PACS system when the post-acquisition processing is completed, and may be later than the actual acquisition time.

[0195] Both the collection time stamp and the content time stamp are in string format and need to be converted to Unix timestamps for numerical comparison.

[0196] Furthermore, when sorting multiple sequences using time features, if timestamps corresponding to both types of labels exist simultaneously, the timestamp corresponding to the collection time label will be used first, because the collection time is closer to the actual scanning sequence. However, when the collection time label is missing or there are abnormalities such as duplication or skipping, the timestamp corresponding to the content time label will be used as an alternative.

[0197] Phase features include two types: a standard timing label (TemporalPositionIdentifier) ​​and a resolvable vendor-specific label.

[0198] Standard time sequence labels are integers that directly represent the sequence number of a time frame or phase (e.g., 1 represents the first frame, 2 represents the second frame). They are the most reliable standard phase labels. If a label exists and is consecutive without repetition, it can be directly used as the core sorting criterion.

[0199] The supplier's private label uses a pre-defined private phase label mapping table of mainstream suppliers (e.g., "0019, 100a" corresponds to the phase type, and "0019, 100b" corresponds to the phase number). By reading the private label data elements of the DICOM file, the phase information is parsed by matching the mapping table. If the supplier is not in the pre-defined list, or the private label cannot be parsed, then that type of feature is ignored.

[0200] Specifically, the presence or absence of the aforementioned time and phase features determines whether it is a simple scene. If both or any one of them is present, it corresponds to a simple scene; if none of them are present, it corresponds to a complex scene.

[0201] In simple scenarios, timestamps that are free of missing, duplicate, or skipped items, or that have standard time-series labels that are consecutive, or that contain vendor-specific phase fields, can be parsed. In these cases, sorting can be performed directly using rules based on the collection time label and / or content time label and / or standard time-series label and / or vendor-specific label.

[0202] The sorting process requires a weighted voting method for fusion. If reliable timestamps and standard phase labels exist, the timestamp weight is 0.4 and the standard phase label weight is 0.6. If reliable timestamps exist but there are no phase labels, but there are resolvable vendor private labels, the timestamp weight is 0.5 and the private phase label weight is 0.5.

[0203] Meanwhile, if there are slight differences in timestamps during sorting, they are considered to be the same time point, and the original instance number is retained. If the phase numbers are not consecutive (such as the 3rd frame being missing), it is marked as "phase sequence incomplete", and the sorting of existing frames is retained in the output.

[0204] In complex scenarios, the phase sequence is inferred based on pixel intensity statistical features. Specifically, the Region of Interest (ROI) is automatically identified (e.g., in a dynamic contrast-enhanced liver scenario, the liver region is located using an anatomical structure recognition algorithm, or the central abdominal region is preset as the ROI). The average CT value, peak CT value, intensity change rate, and other parameters in the ROI region are calculated as intensity feature vectors. Based on the prior phase intensity pattern (e.g., for the liver: average CT value 40-60 HU in plain scan, 80-120 HU in arterial scan, 100-140 HU in portal venous scan, and 90-130 HU in delayed scan), the phase type is determined by cosine similarity matching.

[0205] Another approach is to use prior patterns such as the darkest area on plain scan, rapid enhancement during the arterial phase, and stable enhancement during the portal venous phase to classify the sequence by phase. Specifically:

[0206] Phase classification employs a template matching method, pre-setting CT value curve templates for each phase (e.g., plain scan phase: stable low intensity; arterial phase: rapid rise; portal venous phase: high and stable; delayed phase: slow decline). The dynamic time-normalized distance (DTW) between the CT value curve of the sequence to be classified and each template is calculated, and the phase corresponding to the template with the smallest distance is the classification result.

[0207] This application also discloses a DICOM sequence intelligent normalization and sorting system for implementing the above method.

[0208] The implementation principle is as follows:

[0209] This method unifies DICOM sequences from different devices and patient positions into a standard patient coordinate system, automatically corrects slice order, accurately estimates the true interslice spacing, identifies and removes abnormal slices, and intelligently infers phase or temporal order through multi-feature fusion in the absence of standardized labels. This provides a stable and consistent input data foundation for 3D reconstruction, AI analysis, and anatomical coverage assessment. In this way, it solves problems such as inconsistent DICOM sequence orientation, disordered slice order, and difficulty in identifying temporal phase.

[0210] It should be understood that although the steps in the flowcharts in the accompanying drawings are shown sequentially as indicated by the arrows, these steps are not necessarily performed in the order indicated by the arrows. Unless otherwise expressly stated herein, there is no strict order in which these steps are performed, and they may be performed in other orders.

[0211] The above are all preferred embodiments of this application, and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the structure, shape and principle of this application should be covered within the scope of protection of this application.

Claims

1. A method for intelligent normalization and sorting of DICOM sequences, characterized in that, Includes the following steps: Obtain the original DICOM file; Obtain the check number and serial number, and group the original DICOM files with the same check number and serial number into the same initial sequence; For each of the initial sequences, calculate the field of view size and the standard deviation of the slice center coordinates for all slices; If the coefficient of variation of the field of view is greater than a preset value or the standard deviation is greater than a preset value, it is determined to be a mixed sequence; K-means clustering is used to split the mixed sequence into slices to obtain several subsequences. Temporary check numbers are reassigned to the subsequences and the original check numbers and splitting criteria are recorded. The key DICOM tags of the original DICOM file are parsed to obtain orientation tags and body position information. Normalization based on the orientation tags and the patient coordinate system is then performed, and the orientation of the normalization result is corrected using the body position information. Specifically, Obtain the orientation labels of the patient image and extract the row vectors and column vectors, construct the orientation cosine matrix based on the row vectors and column vectors, and map the cosine matrix to the patient coordinate system. The normal vector is calculated by cross-multiplying the row vector and the column vector and then normalized to a unit vector. Obtain the patient coordinate system type, and obtain the corresponding direction requirements based on the patient coordinate system type; Based on the cosine matrix and the normal vector, it is determined whether the direction requirement is met, and if not, the direction cosine matrix is ​​corrected. When the orientation label of the patient image is missing, extract the patient location label and the patient position label; The normal vector is estimated based on the coordinate distribution represented by the patient location label, and the row vector, column vector, and normal vector are estimated based on the patient position label represented as position information. Based on geometric vector projection, the coordinates of all normalized slices are projected onto the normal vector direction. The layer order is reconstructed based on the results of the geometric vector projection, and the slice interlayer spacing sequence is statistically analyzed to estimate the true interlayer spacing based on the difference between adjacent projections. Extract the claimed interlayer spacing from the original DICOM file; Calculate the relative error between the actual interlayer spacing and the claimed interlayer spacing and determine whether it is greater than a preset value; If so, mark it as an abnormal spacing and record it in the quality control log; Abnormal slices are identified and removed based on the anomaly comparison target label and the slice layer spacing sequence, wherein, The algorithm uses density clustering to identify principal spacing clusters and determines whether the slice layer spacing sequence is within the range of the principal spacing clusters. If not, it is marked as a suspected missing layer or multi-layer boundary. If so, extract the layer thickness label and / or pixel spacing label of all slices and calculate the difference with the mainstream layer thickness value and / or mainstream pixel spacing value respectively; Determine if the difference is greater than the difference threshold; if so, mark it as a metadata anomaly slice. If not, compare the Euclidean distance and image overlap of adjacent slices and determine whether the corresponding threshold is met simultaneously. If so, mark it as a duplicate slice. Record abnormal information, including abnormality type, slice number, and abnormal parameters; The degree of abnormality is determined based on the abnormal parameters in the abnormal information. If the anomaly level is severe, then delete the slice and update the sorting index; If the degree of anomaly is minor, the slice is retained and its weight is reduced. For a set of multiple sequences examined, the sorting feature vector corresponding to each sequence is extracted, and a sorting rule is selected based on the scene complexity to obtain the final sorting key used to characterize the sorting results in the phase dimension and / or time dimension. Specifically, the sorting feature vector includes time features, phase features, and intensity features. The analysis determines whether the extracted sorting feature vector satisfies at least one of the following scenario conditions: no missing / repeated / skipped timestamps, the existence of standard time series labels that are continuous, and the parsable supplier-private phase field. If so, it is defined as a simple scenario, and a sorting rule is selected based on the scenario conditions to directly generate the final sorting key through time features and / or phase features; If not, it is defined as a complex scene, and the ROI region in the slice is identified and the intensity feature vector is calculated. The predicted phase is generated by combining the intensity change pattern to generate the final sorting key.

2. The intelligent normalization and sorting method for DICOM sequences according to claim 1, characterized in that, Based on geometric vector projection, the coordinates of all normalized slices are projected onto the normal vector direction. The layer order is reconstructed based on the results of geometric vector projection, and the slice layer spacing sequence is statistically analyzed to estimate the true layer spacing based on the difference between adjacent projections. This includes the following steps: The geometric center coordinates of the slice are calculated based on the patient location labels; Project the geometric center coordinates of all slices onto the normal vector direction, and multiply the geometric center coordinates by the normal vector to obtain the projection distance; The projection distances are sorted in ascending order to generate a sorting index, which is used to rearrange the slices and obtain the true layer order. The difference in the projected distance between adjacent slices under the true layer sequence is calculated to obtain the slice layer spacing sequence, and the median, mode or mean of the slice layer spacing sequence is taken as the true layer spacing.

3. The intelligent normalization and sorting method for DICOM sequences according to claim 2, characterized in that, The time features include a collection time tag and a content time tag. If both the collection time tag and the content time tag exist, the collection time tag is used first. If the timestamp corresponding to the collection time tag is missing / repeated / skipped, the content time tag is used.

4. A DICOM sequence intelligent normalization and sorting system, characterized in that, Used to implement the method as described in any one of claims 1-3.

Images