An automatic method and system for distinguishing plain scan and enhanced CT based on aortic semantic segmentation and blood vessel region intensity statistics

Abstract

This invention discloses an automatic discrimination method and system for plain and enhanced CT scans based on aortic semantic segmentation and vascular region intensity statistics. The method inputs the three-dimensional CT volume data to be discriminated into a pre-trained aortic semantic segmentation model to obtain a three-dimensional segmentation mask. A binary vascular region of interest mask and an effective intensity mask are then constructed sequentially. A joint mask is obtained through voxel-level logical AND operation. CT intensity values ​​within the joint mask are extracted to form an intensity set. Statistics of the intensity set are calculated and compared with a preset threshold to complete CT type discrimination. The system executes the corresponding method and includes an aortic semantic segmentation module, a vascular mask construction module, an effective intensity mask construction module, a joint mask processing module, a statistical calculation module, and a discrimination output module. This invention does not rely on DICOM metadata or manual image reading, has strong anti-interference capabilities, high discrimination accuracy and robustness, configurable parameters, and good versatility.

Description

Technical Field

[0001] This invention relates to the fields of image processing and artificial intelligence technology. Specifically, this invention relates to an automatic discrimination method and system for plain and enhanced CT scans based on aortic semantic segmentation and vascular region intensity statistics. Background Technology

[0002] Computed tomography (CT) is one of the most widely used imaging techniques in clinical diagnosis and medical imaging research. Depending on whether an iodine-based contrast agent is injected intravenously during the scan, CT scans can be divided into plain CT and contrast-enhanced CT. During contrast-enhanced CT, the iodine contrast agent is distributed through the bloodstream to the blood vessels and blood-rich tissues, causing a significant increase in the CT value of the corresponding area compared to the plain scan. In clinical diagnosis and medical imaging research, both plain and contrast-enhanced CT sequences are often present for the same patient or research cohort. Downstream tasks such as lesion detection and segmentation, radiomics feature extraction, radiotherapy planning, and screening of multicenter retrospective study data all have strict requirements regarding the type of CT scan. Accurately distinguishing between plain and contrast-enhanced CT is an indispensable and crucial step in the medical imaging data preprocessing pipeline.

[0003] Current methods for distinguishing between plain CT and enhanced CT scans mainly fall into three categories. The first category relies on DICOM tags or descriptive fields in the metadata of the radiology information system. However, the standardization of these fields varies significantly across different hospitals and equipment manufacturers, often resulting in missing fields, errors, or inconsistent descriptions. This leads to low reliability when applied to multi-center retrospective imaging data. The second category uses statistical analysis of the CT value distribution across the entire CT image. Since CT volume contains a large amount of mixed components such as air, bone, and soft tissue, the overall statistical values ​​are easily affected by scanning parameters, metal artifacts, partial volume effects, and differences in patient body composition, significantly limiting the accuracy of the distinction. The third category uses end-to-end deep learning classification models to directly perform binary classification on CT slices or volumes. This type of method requires a large amount of labeled training data, suffers from insufficient model interpretability, and struggles to guarantee generalization ability to out-of-distribution samples. Current methods fail to fully utilize the most fundamental and stable difference between enhanced CT and plain CT: the high-density characteristic of iodine contrast agent within the blood vessel lumen in enhanced CT. This is primarily due to technical bottlenecks in multi-phase CT vessel segmentation, preventing existing methods from fully leveraging this essential characteristic for differentiation. Existing methods also struggle to simultaneously meet the application requirements of automated discrimination, high robustness, and high accuracy.

[0004] Therefore, there is an urgent need for an automatic discrimination method for plain and enhanced CT scans that combines automatic aortic segmentation with vascular region intensity statistics. This method would address the technical problems of existing plain and enhanced CT discrimination methods, such as low reliability due to reliance on metadata, insufficient accuracy of full-image statistical discrimination, and poor interpretability and generalization ability of deep learning classification models. The goal is to achieve automatic and robust discrimination between plain and enhanced CT scans without manual intervention or reliance on metadata. Summary of the Invention

[0005] Based on existing technologies, the objective of this invention is to provide an automatic discrimination method and system for plain and enhanced CT scans based on aortic semantic segmentation and vascular region intensity statistics. This method can achieve fully automatic and robust discrimination between plain and enhanced CT scans, solving the technical problems of existing discrimination methods such as reliance on metadata, weak anti-interference ability of full-image statistics, and insufficient interpretability of deep learning classification models.

[0006] In a first aspect, the present invention provides an automatic discrimination method for plain and enhanced CT scans based on aortic semantic segmentation and vascular region intensity statistics, used to determine whether three-dimensional CT volume data is plain CT or enhanced CT, the method comprising: The three-dimensional CT volume data to be judged is input into the pre-trained aortic semantic segmentation model to obtain a three-dimensional segmentation mask that is spatially aligned with the three-dimensional CT volume data to be judged. Each voxel of the three-dimensional segmentation mask corresponds to a discrete category label of different anatomical segments or backgrounds of the aorta. Based on a preset set of blood vessel-related category labels, corresponding voxels are extracted from the three-dimensional segmentation mask to generate a binary blood vessel region of interest mask. Based on the preset effective intensity range of CT values, corresponding voxels are selected from the three-dimensional CT volume data to generate an effective intensity mask; The binary vascular region of interest mask and the effective intensity mask are subjected to a voxel-level logical AND operation to obtain a joint mask. The CT intensity values ​​corresponding to the voxels covered by the joint mask are then extracted to form an intensity set. When the intensity set is not empty, a statistic of the intensity set is calculated, the statistic is compared with a preset threshold, and the CT type of the CT volume data is determined based on the comparison result, wherein the CT type includes enhanced CT and plain CT.

[0007] Furthermore, the loss function used by the aortic semantic segmentation model during the training phase is a weighted combination of Dice loss and cross-entropy loss.

[0008] Furthermore, the method also includes performing morphological post-processing on the binary vascular region of interest mask after generating the mask to eliminate discrete false positive voxels caused by segmentation errors.

[0009] Furthermore, the effective intensity range is used to exclude at least one of air, fat, and partial volume effect voxels with CT values ​​below the effective intensity range, and / or at least one of calcification, metal artifacts, and bone voxels with CT values ​​above the effective intensity range.

[0010] Furthermore, the statistic includes at least one of the following: median, specified percentile, truncated mean, and weighted mean.

[0011] Furthermore, the preset threshold is determined in advance through the following steps: Provide a calibration dataset, which includes enhanced CT data and plain CT data of known types; and The average intensity value of the vascular region is calculated for both the enhanced CT data and the plain CT data, and an intensity distribution histogram is plotted. The preset threshold is determined by receiver operating characteristic (ROC) curve analysis or based on the optimal segmentation point between the distributions.

[0012] A second aspect of the present invention provides an automatic discrimination system for plain and enhanced CT scans based on aortic semantic segmentation and vascular region intensity statistics, used to execute the method described in the first aspect of the present invention, the system comprising: The aortic semantic segmentation module is configured to input three-dimensional CT volume data into a pre-trained aortic semantic segmentation model to obtain a three-dimensional segmentation mask that is spatially aligned with the three-dimensional CT volume data to be judged. Each voxel of the three-dimensional segmentation mask corresponds to a discrete category label of different anatomical segments of the aorta or background. The vascular mask construction module is configured to extract corresponding voxels from the three-dimensional segmentation mask based on a preset set of vascular-related category labels, and generate a binary vascular region of interest mask. An effective intensity mask construction module is configured to select corresponding voxels from the three-dimensional CT volume data according to a preset effective intensity range of CT values ​​to generate an effective intensity mask; The joint mask processing module is configured to perform a voxel-level logical AND operation on the binary blood vessel region of interest mask and the effective intensity mask to obtain a joint mask, and extract the CT intensity values ​​corresponding to the voxels covered by the joint mask to form an intensity set. A statistical calculation module, configured to calculate statistics for the intensity set when the intensity set is not empty; and The discrimination output module is configured to compare the statistics with a preset threshold, determine whether the three-dimensional CT volume data is enhanced CT or plain CT, and output the result.

[0013] Furthermore, the aortic semantic segmentation module includes an aortic semantic segmentation model with a three-dimensional convolutional neural network structure, the model comprising: The encoder unit includes multiple cascaded downsampling stages for extracting multi-scale features from the input 3D CT volumetric data; and The decoder unit comprises multiple cascaded upsampling stages for restoring the spatial resolution of features.

[0014] Furthermore, the decoder unit fuses the features of the corresponding level of the encoder unit through skip connections, and the output layer of the decoder unit generates a three-dimensional segmentation mask that is spatially aligned with the three-dimensional CT volume data input to the aortic semantic segmentation module.

[0015] Furthermore, the discrimination output module is also configured to output the difference between the statistic and the preset threshold, and the confidence index calculated based on the number of voxels in the intensity set.

[0016] The present invention has at least the following beneficial effects: This invention narrows the criteria for CT type discrimination from overall image grayscale statistics to the semantically located aortic region. It directly utilizes the physical effect of increased CT values ​​caused by iodine contrast agents within the blood vessel lumen as the discrimination signal source, resulting in a high signal-to-noise ratio and strong anti-interference capability. Compared with existing methods based on overall image mean or histogram, this invention is unaffected by changes in CT value distribution in non-vascular tissues such as bone, air, and fat, exhibiting better consistency across different body parts and patient body types.

[0017] This invention introduces an aortic semantic segmentation model to obtain spatial localization information of blood vessels, eliminating the need for manual annotation of blood vessel regions or manual placement of regions of interest, thus achieving a fully automated processing flow. The aortic semantic segmentation model is based on a three-dimensional convolutional neural network with an encoder-decoder architecture. Through a blood vessel annotation scheme specifically designed for this invention, its clinical application value is guaranteed, and it exhibits good cross-device generalization ability after pre-training on large-scale multi-center datasets.

[0018] This invention, by setting an effective intensity range, eliminates the influence of extreme CT value voxels such as calcified plaques and metal artifacts on the statistical results, thus enhancing the robustness of the discrimination method in actual clinical data. The effective intensity range can screen out a reasonable range of CT values ​​for normal or enhanced blood within the blood vessel lumen, avoiding statistical bias caused by extreme values.

[0019] The blood vessel category label set, effective intensity range, and discrimination threshold of this invention are all configurable parameters, making them easy to adapt to different segmentation model versions, different label definition protocols, and different scanning devices and imaging parameters, thus exhibiting good versatility and portability. In practical applications, users can adjust the relevant parameters according to the actual application scenario without having to redesign the entire discrimination process.

[0020] This invention overcomes the technical bottleneck of relying on the description field in the DICOM header information in existing technologies, and avoids the risk of discrimination failure caused by missing, incorrect, or non-standard metadata. It is particularly suitable for retrospective image data integration scenarios involving multiple centers and multiple sources. It can independently determine the type of CT scan (plain CT and enhanced CT) without relying on the metadata support of the radiology information system, and its discrimination accuracy and precision are significantly improved compared to existing methods.

[0021] In summary, this invention first defines, based on clinical medical mechanisms and computer vision principles, arterial vessel types that are technically easy to identify and clinically stable, including the aorta (including the ascending aorta, aortic arch, and descending aorta) and bilateral common iliac arteries, as model learning targets, and successfully trains a high-precision aortic semantic segmentation model. Based on this innovative aortic semantic segmentation model, it achieves stable resolution of existing plain and enhanced CT discriminant methods. Furthermore, this invention employs a technique combining fully automated aortic semantic segmentation with vascular region intensity statistics, solving the technical problems of existing plain and enhanced CT discriminant methods, such as reliance on metadata, insufficient full-image statistical discrimination accuracy, and poor interpretability and generalization ability of deep learning classification models. This approach is suitable for large-scale image quality control, cohort screening, and stratification of clinical research data. Attached Figure Description

[0022] To further illustrate the advantages and other features of the various embodiments of the present invention, a more specific description of the embodiments of the present invention will be presented with reference to the accompanying drawings. It is understood that these drawings depict only typical embodiments of the invention and are therefore not intended to limit its scope. In the drawings, identical or corresponding parts will be indicated by the same or similar reference numerals for clarity.

[0023] Figure 1 A flowchart of an automatic discrimination method for plain CT and enhanced CT scans in one embodiment of the present invention is shown.

[0024] Figure 2 An architecture diagram of an aortic semantic segmentation model according to one embodiment of the present invention is shown.

[0025] Figure 3 A structural diagram of an automatic discrimination system according to one embodiment of the present invention is shown.

[0026] List of reference numerals in the attached figures: 100-Discrimination system, 101-Aortic semantic segmentation module, 101-1-Encoder unit, 101-2-Decoder unit, 102-Vascular mask construction module, 103-Effective intensity mask construction module, 104-Joint mask processing module, 105-Statistical calculation module, 106-Discrimination output module. Detailed Implementation

[0027] It should be noted that the components in the various figures may be shown exaggeratedly for illustrative purposes and are not necessarily to scale. In each figure, the same reference numerals are used for components that are identical or have the same function.

[0028] In this invention, the various embodiments are merely intended to illustrate the solutions of the invention and should not be construed as limiting.

[0029] In this invention, unless otherwise specified, the quantifiers “a” and “one” do not exclude scenarios involving multiple elements.

[0030] It should also be noted that, in the embodiments of the present invention, only a portion of the components or parts may be shown for clarity and simplicity. However, those skilled in the art will understand that, under the teachings of the present invention, necessary components or parts can be added as needed for specific scenarios. Furthermore, unless otherwise stated, features in different embodiments of the present invention can be combined with each other. For example, a feature in the second embodiment can replace a corresponding or functionally identical or similar feature in the first embodiment, and the resulting embodiment will also fall within the scope of disclosure or description of this application.

[0031] It should also be noted that within the scope of this invention, the terms "same", "equal", and "equal to" do not mean that the two values ​​are absolutely equal, but allow for a certain reasonable error. In other words, the terms also cover "substantially the same", "substantially equal", and "substantially equal to".

[0032] In this invention, the modules of the system according to the invention can be implemented using software, hardware, firmware, or a combination thereof. When a module is implemented using software, its function can be implemented through computer program flow. For example, the module can be implemented using code segments (such as code segments in languages ​​like C and C++) stored in a storage device (such as a hard disk, memory, etc.), wherein the corresponding function of the module can be implemented when the code segment is executed by a processor. When a module is implemented using hardware, its function can be implemented by setting a corresponding hardware structure. For example, the module's function can be implemented by hardware programming a programmable device such as a field-programmable gate array (FPGA), or by designing an application-specific integrated circuit (ASIC) that includes multiple transistors, resistors, capacitors, and other electronic devices. When a module is implemented using firmware, the module's function can be written into a read-only memory such as an EPROM or EEPROM in the form of program code, and the corresponding function of the module can be implemented when the program code is executed by a processor. In addition, some functions of the module may need to be implemented by separate hardware or by working in cooperation with the hardware. For example, the detection function is implemented by a corresponding sensor (such as a proximity sensor, accelerometer, gyroscope, etc.), the signal transmission function is implemented by a corresponding communication device (such as a Bluetooth device, infrared communication device, baseband communication device, Wi-Fi communication device, etc.), the output function is implemented by a corresponding output device (such as a display, speaker, etc.), and so on.

[0033] Furthermore, the numbering of the steps in the methods of the present invention does not limit the execution order of the method steps. Unless otherwise specified, the method steps may be executed in different orders.

[0034] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0035] Figure 1 A flowchart of an automatic discrimination method for plain CT and enhanced CT scans in one embodiment of the present invention is shown.

[0036] In one embodiment of the present invention, an automatic discrimination method for plain CT and enhanced CT based on aortic semantic segmentation and vascular region intensity statistics is provided. The method is used to automatically discriminate three-dimensional CT volume data and output the result of the CT volume data as plain CT or enhanced CT. Specifically, it includes the following steps S100 to S500.

[0037] Step S100, aortic semantic segmentation and 3D segmentation mask generation: In one embodiment of the present invention, the 3D CT volume data to be judged is input into a pre-trained aortic semantic segmentation model to obtain a 3D segmentation mask that is spatially aligned with the CT volume data and has the same resolution. The value of each voxel in the 3D segmentation mask is a discrete category label, which corresponds to different anatomical segments of the aorta or background categories.

[0038] Figure 2 An architecture diagram of an aortic semantic segmentation model according to one embodiment of the present invention is shown.

[0039] like Figure 2 As shown, in this embodiment, the aortic semantic segmentation model employs a three-dimensional fully convolutional neural network with an encoder-decoder structure. The encoder path comprises six cascaded downsampling stages, each containing at least one three-dimensional convolutional layer, a normalization layer, and a non-linear activation function. The convolutional kernel size is preferably 3×3×3, and the normalization includes instance normalization or batch normalization. The non-linear activation function can be, for example, LeakyReLU. At the end of each downsampling stage, a convolutional operation with a stride greater than 1 is used to reduce spatial resolution and increase the feature channel dimension, progressively extracting multi-scale features from low-level texture to high-level semantics. The decoder path is symmetrically set to the encoder path, containing six cascaded upsampling stages. Each upsampling stage restores spatial resolution through a transposed convolutional operation, and simultaneously uses skip connections to concatenate the feature maps output from the same level stage in the encoder path with the feature map of the current stage in the decoder path, preserving semantic information while restoring spatial details. The last layer of the decoder path is a 1×1×1 convolutional layer. The number of output channels is equal to the sum of the predefined number of aortic anatomical segment categories and the number of background categories. After voxel-level Softmax or Argmax operations, a 3D segmentation mask is generated.

[0040] In one embodiment of the present invention, the aortic semantic segmentation model uses a weighted combination of Dice loss and cross-entropy loss as the loss function during the training phase. Dice loss is used to optimize the overlap measurement of segmented regions and is robust to class imbalance (such as small blood vessels relative to large background); cross-entropy loss is used to provide voxel-level probabilistic calibration supervision.

[0041] In one embodiment of the present invention, the total loss function used by the aortic semantic segmentation model during the training phase is defined as: in, For the training set True-value one-hot encoding of individual elements; The category predicted by the aortic semantic segmentation model probability; This is a smoothing constant used to prevent division by zero. is the weighting coefficient, used to balance the weights of the cross-entropy term.

[0042] In one embodiment of the present invention, a data augmentation strategy including random rotation, random scaling, random elastic deformation, random contrast adjustment, random gamma transformation, random mirror flipping, and random pruning is adopted during the training process of the aortic semantic segmentation model to improve the model's generalization ability under different scanning devices, scanning parameters, and patient body sizes.

[0043] In one embodiment of the present invention, during the model inference stage, the pre-trained aortic semantic segmentation model first performs intensity normalization and isotropic resampling preprocessing on the input three-dimensional CT volume data. When the size of the input CT volume exceeds the receptive field range of the model in a single processing, a sliding window strategy combined with Gaussian weighted average of overlapping regions is used to perform block-by-block inference on the entire volume. Finally, the prediction results of each block are stitched together to output a three-dimensional segmentation mask.

[0044] Step S200: Construction of binary blood vessel region of interest mask: In one embodiment of the present invention, a voxel-level attribution determination is performed on the three-dimensional segmentation mask obtained in step S100 based on a pre-configured set of blood vessel-related category labels to generate a binary blood vessel region of interest mask.

[0045] In one embodiment of the present invention, based on a pre-configured set of blood vessel-related category labels... In the 3D segmentation mask obtained in step S100, voxel-level attribution determination is performed to classify the category labels as belonging to the set. The voxels are set to 1, and the rest are set to 0, thus generating a binary vascular region of interest mask. , represented as: in, For the 3D segmentation mask in spatial coordinates The category label value at that location.

[0046] In one embodiment of the present invention, a set of blood vessel-related category labels The selection principle is to prioritize large systemic arterial structures (such as the aorta and common iliac artery) with large luminal cross-sectional areas and stable contrast agent filling levels in enhanced CT, to ensure the stability of the statistical estimation. (Vascular-related category label set) The specific values ​​correspond to the label definition protocol of the segmentation model used, and the configuration can be replaced according to the mapping relationship between different segmentation models or label versions.

[0047] Step S300, Effective Intensity Mask Generation: In one embodiment of the present invention, an effective intensity range for CT values ​​is preset in the CT intensity domain, using HU as the unit of measurement. The effective intensity range is configured to represent extreme CT value voxels that are physiologically unlikely to belong to normal or enhanced intravascular blood, including: below Low-density voxels produced by air, fat, or partial volume effect, and higher than Calcification, metal artifacts, or bone voxels. In this embodiment, voxel-level intensity filtering is performed on the input 3D CT volume data, and an effective intensity mask is generated from the CT volume data. , represented as: in, CT volume in spatial coordinates The voxel intensity value at that location.

[0048] Step S400, Joint Mask Generation and Intensity Set Extraction: The binary blood vessel region of interest mask obtained in step S200 is generated. The effective intensity mask obtained in step S300 Perform voxel-level logical AND operations to obtain the joint mask. , represented as: exist Extract the corresponding CT intensity values ​​from all voxel locations to form an intensity set. .

[0049] Step S500, Statistical Calculation and CT Type Determination: In one embodiment of the present invention, when the intensity set obtained in step S400... When the set of intensities is not empty, calculate the intensity set. The statistical measure is calculated and compared with a preset discrimination threshold. Based on the comparison result, the CT type of the 3D CT volume data is determined, where the CT type includes enhanced CT and plain CT. In this embodiment, the intensity set... The statistic is the arithmetic mean. In other alternative implementations, the statistic may be replaced by at least one of the following: median, specified percentile, truncated mean, and weighted mean.

[0050] In one embodiment of the present invention, the preset discrimination threshold is determined through pre-calibration. Specifically, the calibration method is as follows: First, a calibration dataset is acquired, which includes multiple cases of enhanced CT data and multiple cases of plain CT data of known types, covering clinical data from different scanning devices, different scanning protocols, and different patient body types. For both the enhanced CT data and the plain CT data in the calibration dataset, the intensity statistics of the corresponding vascular regions are calculated, and intensity distribution histograms of the two sets of data are plotted. Through receiver operating characteristic (ROC) curve analysis, the preset threshold is determined by the optimal segmentation point that maximizes the Youden index.

[0051] In a specific embodiment of the present invention, in the application scenario of determining whether a given three-dimensional CT volume is enhanced CT, the entire process from CT volume input to determination result output is as follows: In this embodiment, in step S100, a 3D CT volume data is input, stored in NIfTI format (.nii.gz) with a voxel value of HU. This CT volume is then input into a pre-trained aortic semantic segmentation model. The aortic semantic segmentation model is a fully convolutional encoder-decoder network based on a 3D U-Net architecture. The encoder of the aortic semantic segmentation model contains 6 downsampling stages with feature channels of 32, 64, 128, 256, 320, and 320 respectively. The decoder is symmetrical to the encoder, and each stage fuses features at the same level as the encoder through skip connections. The model has been trained on over 1000 CT data samples covering multiple centers and multiple devices, enabling voxel-level semantic annotation of the aorta and its main branches. Label 1 corresponds to the aorta (including the ascending aorta, aortic arch, and descending aorta), label 2 corresponds to the left common iliac artery, label 3 corresponds to the right common iliac artery, and label 0 corresponds to the background. The aortic semantic segmentation model uses a sliding window strategy during inference with a window overlap rate of 50%, and Gaussian weighted average fusion is applied to the overlapping areas. The inference output is a 3D segmentation mask of the same size as the input CT scan. It is also stored in NIfTI format in a subdirectory of the examination directory, with the file name matching the corresponding CT volume to ensure a pairing relationship.

[0052] In this embodiment, in step S200, voxel-level attribution determination is performed on the three-dimensional segmentation mask obtained in step S100 based on the pre-configured set of blood vessel-related category labels to generate a binary blood vessel region of interest mask. The rule for voxel-level attribution determination is: class labels in the 3D segmentation mask belong to the set. One voxel is assigned a value of 1, and the rest are assigned a value of 0, thereby generating a binary vascular region of interest mask. .

[0053] In this embodiment, in step S300, voxel-level intensity filtering is performed on the input three-dimensional CT volume data to generate an effective intensity mask from the CT volume data. The preset effective intensity range for CT values ​​is [0 HU, 250 HU]. This range covers the main intensity range of non-calcified blood within the blood vessel lumen (approximately 30 to 50 HU on plain scan) and iodine-enhanced blood (typically 100 to 250 HU). The voxel-level intensity screening rule is as follows: voxels with CT intensity values ​​within this effective intensity range in the 3D CT volume data are assigned a value of 1, and all other voxels are assigned a value of 0, thus generating an effective intensity mask. This effective intensity range can avoid the influence of extreme CT value voxels on subsequent statistical calculations.

[0054] In this embodiment, in step S400, the binary blood vessel region of interest is masked. With effective intensity mask A voxel-level logical AND operation is performed to obtain a joint mask. The rule for the voxel-level logical AND operation is: only when the binary blood vessel region of interest mask is obtained... and effective intensity mask When all voxel values ​​are 1 at the same spatial coordinates, the joint mask... The voxel value at this coordinate is assigned a value of 1, and the value in all other cases is assigned a value of 0. At all spatial coordinates in the joint mask where the voxel value is 1, the CT intensity values ​​corresponding to the 3D CT volume data are extracted. All extracted CT intensity values ​​constitute an intensity set. .

[0055] In this embodiment, in step S500, if the intensity set If the set is empty (i.e., there are no voxels in the aortic region that meet the effective intensity constraints, possibly due to segmentation failure or the CT coverage area not including the aorta), return a null value (None) and classify the input 3D CT volume data as plain CT according to a preset strategy. If the intensity set... If the expression is not empty, calculate its arithmetic mean. .

[0056] In this embodiment, in step S600, the preset discrimination threshold determined by calibration is... =80HU. The discrimination rule is: when the intensity set arithmetic mean When the value is greater than 80 HU, the 3D CT volumetric data is determined to be enhanced CT, and True is returned; when the intensity set... arithmetic mean When less than or equal to 80HU, or If the value is empty, the 3D CT volume data is determined to be a plain CT scan, and False is returned.

[0057] Possible variations of the method of the present invention will be described below with reference to other embodiments of the present invention.

[0058] In one embodiment of the invention, the aortic region may contain a small number of high-density or low-density voxels introduced by segmentation errors (such as voxels at the edge of adjacent calcified plaques being mistakenly assigned to the aortic lumen), causing the arithmetic mean to be skewed. This embodiment replaces the statistic in step S500 with a truncated mean: removing the intensity set. After selecting the lowest 5% and highest 5% of voxels, the average value is calculated for the remaining voxels. Correspondingly, the discrimination threshold in step S6... It needs to be redefined on the calibration dataset, for example, after determining the optimal operating point through ROC analysis. HU. In another embodiment, the statistic uses the median as the discrimination threshold. Recalibrate as before.

[0059] In one embodiment of the present invention, for scenarios with low segmentation quality or significant fragmentation of the vascular region, after generating a binary vascular region of interest mask in step S200, morphological post-processing is performed on the binary vascular region of interest mask to eliminate discrete false positive voxels caused by segmentation errors and improve the purity of vascular region localization. Morphological post-processing specifically includes two steps: performing three-dimensional connected component analysis on the binary vascular region of interest mask, calculating the number of voxels in each connected component, retaining only the connected component with the largest number of voxels, and assigning the voxels of the remaining smaller connected components to 0; subsequently, performing a three-dimensional morphological opening operation on the processed three-dimensional segmentation mask, with the structuring element being a sphere with a radius of 2 voxels, performing an erosion operation followed by a dilation operation to smooth the mask boundaries and eliminate burrs caused by segmentation noise.

[0060] In one embodiment of the present invention, when outputting the discrimination result in step S600, the result is also based on statistics and a preset discrimination threshold. The difference, and the intensity set The number of voxels is calculated and a confidence index is output. Specifically, the confidence index is calculated based on statistics and a preset discrimination threshold. The absolute or relative difference between them, the larger the difference, the higher the confidence level; and whether the number of effective voxels covered by the joint mask exceeds the preset minimum voxel count threshold, the more effective voxels, the higher the confidence level.

[0061] In one embodiment of the present invention, in a large-scale queue processing scenario, a single patient examination catalog may contain multiple 3D CT volume data (corresponding to different scanning phases or different reconstruction parameters), and each volume has a corresponding segmentation mask in a sub-catalog. The above-described discrimination process is executed one by one for all 3D CT volumes in the examination catalog, and the discrimination results are written to a structured record (such as JSON or CSV), marking the enhancement / plain scan attributes of each volume. Downstream workflows (such as lesion mask screening, maximum tumor ROI extraction, and radiomics feature calculation) only perform subsequent processing on volumes marked as enhanced CT, automatically skipping plain scan volumes, thereby avoiding the impact of non-enhanced data on the accuracy of downstream analysis.

[0062] Figure 3 A structural diagram of an automatic discrimination system according to one embodiment of the present invention is shown.

[0063] In one embodiment of the present invention, an automatic discrimination system for plain and enhanced CT scans (hereinafter referred to as the "discrimination system") 100 based on aortic semantic segmentation and vascular region intensity statistics is provided. The system includes an aortic semantic segmentation module 101, a vascular mask construction module 102, an effective intensity mask construction module 103, a joint mask processing module 104, a statistical calculation module 105, and a discrimination output module 106. Each module works together to execute the discrimination method in the foregoing embodiment of the present invention to achieve fully automatic type discrimination of three-dimensional CT volume data.

[0064] In one embodiment of the present invention, the aortic semantic segmentation module 101 is configured to receive three-dimensional CT volume data to be judged, automatically identify the spatial location of the aorta and its main branches through a pre-trained deep learning model, and output a three-dimensional segmentation mask spatially aligned with the input data. The value of each voxel in the three-dimensional segmentation mask is a discrete category label, corresponding to different anatomical segments of the aorta or background categories.

[0065] In one embodiment of the present invention, the aortic semantic segmentation module 101 is a three-dimensional convolutional neural network, including an encoder unit 101-1 and a decoder unit 101-2. The encoder unit 101-1 includes multiple cascaded downsampling stages for extracting multi-scale features from the input three-dimensional CT volumetric data; the decoder unit 101-2 includes multiple cascaded upsampling stages for restoring the spatial resolution of the features. The decoder unit 101-2 fuses features from corresponding layers of the encoder unit 101-1 through skip connections, and the output layer of the decoder unit 101-2 generates a three-dimensional segmentation mask.

[0066] In one embodiment of the present invention, the vascular mask construction module 102 is configured to receive a three-dimensional segmentation mask output by the aortic semantic segmentation module 101, and extract corresponding voxels to generate a binary vascular region of interest mask based on a pre-configured set of vascular-related category labels. In this embodiment, the preset set of vascular-related category labels is L={1,2,3}, corresponding to the three types of systemic circulation large artery structures: the aorta, the left common iliac artery, and the right common iliac artery. The vascular mask construction module 102 performs voxel-level classification determination on the three-dimensional segmentation mask, assigning a value of 1 to voxels whose category labels belong to set L, and assigning a value of 0 to the remaining voxels, thereby generating a binary vascular region of interest mask.

[0067] In one embodiment of the present invention, the vascular mask construction module 102 is further configured to perform morphological post-processing on the generated binary vascular region of interest mask to eliminate discrete false positive voxels caused by segmentation errors and improve the purity of vascular region localization.

[0068] In one embodiment of the present invention, the effective intensity mask construction module 103 is configured to receive raw three-dimensional CT volume data and generate an effective intensity mask by selecting corresponding voxels according to a preset effective intensity range of CT values.

[0069] In one embodiment of the present invention, the joint mask processing module 104 is configured to simultaneously receive the binary vascular region of interest mask output by the vascular mask construction module 102 and the effective intensity mask output by the effective intensity mask construction module 103. The two masks are then subjected to a voxel-level logical AND operation to obtain a joint mask, and the CT intensity values ​​corresponding to the voxels covered by the joint mask are extracted to form an intensity set. The joint mask processing module 104 extracts the CT intensity values ​​corresponding to the three-dimensional CT volume data at the spatial coordinates where all voxels in the joint mask have a value of 1. All extracted CT intensity values ​​constitute an intensity set. When the intensity set S is empty, that is, when the number of effective voxels covered by the joint mask is 0, the joint mask processing module 104 sends an empty set signal to the statistical calculation module 105 to trigger the preset discrimination strategy.

[0070] In one embodiment of the present invention, the statistical calculation module 105 is configured to receive the intensity set output by the joint mask processing module 104. In intensity set When the set is not empty, the statistics of the intensity set are calculated. When the statistical calculation module 105 receives an empty set signal, it does not perform statistical calculations, but directly sends an empty set flag to the discrimination output module 106.

[0071] In one embodiment of the present invention, the discrimination output module 106 is configured to receive statistics or empty set markers output by the statistical calculation module 105, compare the statistics with a preset threshold, determine whether the three-dimensional CT volume data is enhanced CT or plain CT, and output the result. When an empty set marker is received, the discrimination output module 106 outputs the discrimination result according to a preset strategy.

[0072] In one embodiment of the present invention, the discrimination output module 106 is further configured to output a confidence index, which is calculated based on the difference between a statistic and a preset threshold, and the number of voxels in the intensity set. Specifically, the confidence index is calculated as follows: based on the absolute or relative difference between the statistic and the preset threshold, the larger the difference, the higher the confidence; and based on whether the number of effective voxels covered by the joint mask exceeds a preset minimum voxel count threshold, the more effective voxels, the higher the confidence. When the confidence index is lower than a preset confidence lower limit, the discrimination output module 106 adds an uncertainty warning mark to the output result, indicating that the output result needs to be manually reviewed.

[0073] In one specific embodiment of the present invention, the discrimination system 100 is deployed on one or more servers equipped with graphics processing units (GPUs). The inference computation of the aortic semantic segmentation module 101 is performed in parallel on the GPUs using 3D convolution operations; the computations of the vascular mask construction module 102, the effective intensity mask construction module 103, the joint mask processing module 104, the statistical calculation module 105, and the discrimination output module 106 are performed on the central processing unit (CPU). The discrimination system 100 can be deployed as a service, receiving the storage path of input CT volume data, automatically performing the entire discrimination process, and returning structured discrimination results, making it suitable for batch processing scenarios of large-scale clinical imaging data.

[0074] Effect data of the invention In one application embodiment of the present invention, the discrimination results of the method and system provided in the foregoing embodiments of the present invention on the category of three-dimensional CT volume data were verified. At the same time, the classification methods or end-to-end deep learning classification models that rely on DICOM metadata in the prior art were used as a comparison. The discrimination results are shown in Table 1.

[0075] Table 1 Comparison of discrimination results As shown in Table 1, in this embodiment, the method and system provided by the foregoing embodiments of the present invention can more accurately and precisely determine the category of three-dimensional CT volume data compared with the two methods in the prior art, thus verifying the technical effect of the method and system provided by the foregoing embodiments of the present invention in the discrimination of three-dimensional CT volume data.

[0076] This method was tested on 103 chest and abdominal CT images, including 49 plain scans and 54 enhanced scans. The model achieved an accuracy of 99.0%, a sensitivity of 98.1%, and a specificity of 100.0%, with an average processing time of 1.3 seconds per image. Compared to traditional end-to-end deep learning algorithms, the accuracy is improved by 6%; compared to manual image interpretation, this method is more than 10 times faster, effectively distinguishing between plain and enhanced CT images. It demonstrates good stability in batch image screening scenarios, meeting the needs of rapid clinical interpretation and research.

[0077] Although various embodiments of the invention have been described above, it should be understood that they are presented by way of example only and not as limitations. It will be apparent to those skilled in the art that various combinations, modifications, and alterations can be made without departing from the spirit and scope of the invention. Therefore, the breadth and scope of the invention disclosed herein should not be limited by the exemplary embodiments disclosed above, but should be defined solely by the appended claims and their equivalents.

Claims

1. An automatic discrimination method of plain and enhanced CT based on aortic semantic segmentation and vessel region intensity statistics, for discriminating whether a three-dimensional CT volume data is a plain CT or an enhanced CT, characterized in that, The method includes: The three-dimensional CT volume data to be judged is input into the pre-trained aortic semantic segmentation model to obtain a three-dimensional segmentation mask that is spatially aligned with the three-dimensional CT volume data to be judged. Each voxel of the three-dimensional segmentation mask corresponds to a discrete category label of different anatomical segments or backgrounds of the aorta. Based on a preset set of blood vessel-related category labels, corresponding voxels are extracted from the three-dimensional segmentation mask to generate a binary blood vessel region of interest mask. Based on the preset effective intensity range of CT values, corresponding voxels are selected from the three-dimensional CT volume data to generate an effective intensity mask; The binary vascular region of interest mask and the effective intensity mask are subjected to a voxel-level logical AND operation to obtain a joint mask. The CT intensity values ​​corresponding to the voxels covered by the joint mask are then extracted to form an intensity set. When the intensity set is not empty, a statistic of the intensity set is calculated, the statistic is compared with a preset threshold, and the CT type of the CT volume data is determined based on the comparison result, wherein the CT type includes enhanced CT and plain CT.

2. The method of claim 1, wherein, The loss function used in the training phase of the aortic semantic segmentation model is a weighted combination of Dice loss and cross-entropy loss.

3. The method of claim 1, wherein, The method further includes performing morphological post-processing on the binary vascular region of interest mask after generating the mask to eliminate discrete false positive voxels caused by segmentation errors.

4. The method of claim 1, wherein, The effective intensity range is used to exclude at least one of air, fat, and partial volume effect voxels with CT values ​​below the effective intensity range, and / or at least one of calcification, metal artifacts, and bone voxels with CT values ​​above the effective intensity range.

5. The method of claim 1, wherein, The statistics include at least one of the following: median, specified percentile, truncated mean, and weighted mean.

6. The method of claim 1, wherein, The preset threshold is determined in advance through the following steps: Provide a calibration dataset, which includes enhanced CT data and plain CT data of known types; and The average intensity value of the vascular region is calculated for both the enhanced CT data and the plain CT data, and an intensity distribution histogram is plotted. The preset threshold is determined by receiver operating characteristic (ROC) curve analysis or based on the optimal segmentation point between the distributions.

7. An automatic discrimination system of plain and enhanced CT based on aortic semantic segmentation and vessel region intensity statistics, characterized in that, The system for performing the method according to any one of claims 1 to 6, the system comprising: The aortic semantic segmentation module is configured to input three-dimensional CT volume data into a pre-trained aortic semantic segmentation model to obtain a three-dimensional segmentation mask that is spatially aligned with the three-dimensional CT volume data to be judged. Each voxel of the three-dimensional segmentation mask corresponds to a discrete category label of different anatomical segments of the aorta or background. The vascular mask construction module is configured to extract corresponding voxels from the three-dimensional segmentation mask based on a preset set of vascular-related category labels, and generate a binary vascular region of interest mask. An effective intensity mask construction module is configured to select corresponding voxels from the three-dimensional CT volume data according to a preset effective intensity range of CT values ​​to generate an effective intensity mask; The joint mask processing module is configured to perform a voxel-level logical AND operation on the binary blood vessel region of interest mask and the effective intensity mask to obtain a joint mask, and extract the CT intensity values ​​corresponding to the voxels covered by the joint mask to form an intensity set. A statistical calculation module, configured to calculate statistics for the intensity set when the intensity set is not empty; and The discrimination output module is configured to compare the statistics with a preset threshold, determine whether the three-dimensional CT volume data is enhanced CT or plain CT, and output the result.

8. The system of claim 7, wherein, The aortic semantic segmentation module includes an aortic semantic segmentation model with a three-dimensional convolutional neural network structure, the model comprising: The encoder unit includes multiple cascaded downsampling stages for extracting multi-scale features from the input 3D CT volumetric data; and The decoder unit comprises multiple cascaded upsampling stages for restoring the spatial resolution of features.

9. The system of claim 8, wherein, The decoder unit fuses the features of the corresponding level of the encoder unit through skip connections, and the output layer of the decoder unit generates a three-dimensional segmentation mask that is spatially aligned with the three-dimensional CT volume data input to the aortic semantic segmentation module.

10. The system of claim 7, wherein, The discrimination output module is further configured to output the difference between the statistic and a preset threshold, and a confidence index calculated based on the number of voxels in the intensity set.

Images