An abnormal brain area positioning method and system based on plain CT and a storage medium
By combining deep learning and the Voxelmorph registration model with a voxel-structure dual-scale detection framework, the problem of subjectivity and lack of automation in the analysis of plain CT images of the brain is solved, achieving high-precision, low-radiation localization of brain region abnormalities, which is applicable to various types of CT equipment and clinical scenarios.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- XIEHE HOSPITAL ATTACHED TO TONGJI MEDICAL COLLEGE HUAZHONG SCI & TECH UNIV
- Filing Date
- 2026-05-11
- Publication Date
- 2026-08-04
AI Technical Summary
Current brain plain CT image analysis techniques suffer from problems such as high subjectivity, insufficient quantification capabilities, difficulty in cross-time point comparison, and low automation, which cannot meet the clinical needs for rapid, stable, and accurate localization and dynamic assessment of brain abnormalities.
A deep learning segmentation model is used to identify and extract ROI images of the left and right brain regions. These images are then registered using a Voxelmorph registration model. Significance tests and statistical analyses at both voxel and structural scales are performed, and a healthy control database is used to correct the list of abnormal brain regions, thus achieving a multi-level detection framework with cross-validation.
It achieves highly accurate and robust localization of abnormal brain regions under conventional plain CT conditions, and is suitable for emergency, physical examination and various types of CT equipment scenarios, reducing the threshold of radiation examination, preserving the ability to detect small lesions and suppressing noise interference.
Smart Images
Figure CN122175975B_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of medical image analysis, specifically to a method, system, and storage medium for locating abnormal brain regions based on plain CT scans. Background Technology
[0002] Brain CT (computed tomography) scans are a core imaging tool for the early screening and diagnosis of neurological diseases due to their rapid examination, high accessibility in emergency settings, and sensitivity to high-density lesions such as hemorrhage. In clinical practice, the analysis of brain CT scans still mainly relies on the doctor's visual interpretation of the images. The doctor assesses the lesion by observing indicators such as brain parenchyma density, ventricular system morphology, presence of mass effect, and midline shift. This is supplemented by manual CT value (Hounsfield Units, HU) measurement of some suspicious areas to assess the amount of intracranial hemorrhage, early low-density signs, and degree of edema. However, this method, which relies primarily on visual inspection and localized measurements, has significant limitations: First, it is highly subjective and depends on the experience of the doctors reading the images, resulting in poor consistency among different doctors; second, minor lesions (such as small hemorrhages, early ischemic foci, and mild edema) are easily overlooked, leading to missed diagnoses or delays; third, the measurement range of CT values is limited, making it difficult to achieve systematic and quantitative analysis of the entire brain; and fourth, manual comparison between follow-up CT scans at multiple time points is difficult due to differences in scanning angle, slice thickness, and patient position, which can easily lead to errors.
[0003] To overcome the limitations of manual assessment, traditional computer-aided diagnostic methods often employ machine learning approaches such as threshold segmentation, region growing, edge detection, morphological operations, or manual feature extraction. These rule-based or handcrafted feature-based algorithms often exhibit insufficient robustness when dealing with complex and morphologically varied brain lesions, poor adaptability to different devices, imaging conditions, or individual patient differences, and limited generalization performance.
[0004] In recent years, deep learning technology has been increasingly applied to segmentation and registration tasks in medical images, which can improve the ability to identify abnormal brain regions to some extent. However, most existing models are only designed for single tasks (such as segmentation or registration only), and lack the ability to combine automatic segmentation and spatial alignment for the localization and follow-up quantification of abnormal brain regions; moreover, some models rely on enhanced scanning or multimodal imaging, making them difficult to apply to real-world scenarios in clinical settings where only plain CT scans are available.
[0005] In summary, existing brain plain CT image analysis techniques generally suffer from problems such as high subjectivity, insufficient quantification capabilities, difficulty in cross-time point comparison, and low automation, which cannot meet the clinical needs for rapid, stable, and accurate localization and dynamic assessment of brain abnormalities. Summary of the Invention
[0006] This application provides a method, system, and storage medium for locating abnormal brain regions based on plain CT scans, which can solve the technical problems of difficulty in locating abnormal brain regions and high noise in existing technologies that rely on CTP perfusion imaging or traditional voxel comparison methods.
[0007] In a first aspect, embodiments of this application provide a method for locating abnormal brain regions based on plain CT scans, which includes: Deep learning segmentation models were used to process the sequence of plain CT images of the brain of the subjects to identify and extract ROI images of multiple anatomical structures in the left and right brain regions; The Voxelmorph registration model was used to register the ROI images of the left and right brain regions to obtain the correspondence between homologous voxels between the ROI images of the left and right brain regions. Based on the correspondence of homologous voxels between ROI images of the left and right brain regions, the differences in gray values between ROI images of the left and right brain regions are compared at the voxel scale to perform significance tests and generate a preliminary list of abnormal brain regions. At the same time, at the structural scale, the gray distribution of each ROI image is statistically analyzed to generate a list of potential abnormal brain regions. The preliminary list of abnormal brain regions is corrected using a healthy control database; cross-validation is performed using the potential abnormal brain region list and the corrected preliminary abnormal brain region list to output the final abnormal brain region localization results.
[0008] Preferably, the Voxelmorph registration model is used to register the left and right brain region ROI images to obtain the correspondence between homologous voxels between the left and right brain region ROI images. Specifically, this includes the following steps: A ROI image from one side of the brain was selected as the fixed image, and the corresponding ROI image from the other side was selected as the moving image. The fixed and moving images were then input into the Voxelmorph registration model to output a three-dimensional non-rigid deformation field. Three-dimensional non-rigid deformation fields are used to characterize the voxel-level mapping function relationship from the space of a fixed image to the space of a moving image. According to the three-dimensional non-rigid deformation field Establish voxel correspondences according to the following pairing steps: For any voxel position x in a fixed image, its corresponding homologous voxel position in the moving image is y= (x); For any voxel position y in the moving image, its corresponding voxel position in the fixed image is obtained by inverse mapping of the deformation field, x = (y).
[0009] Preferably, based on the correspondence of homologous voxels between left and right brain region ROI images, the grayscale differences between left and right brain region ROI images are compared at the voxel scale, and a significance test is performed to generate a preliminary list of abnormal brain regions. This specifically includes the following steps: The fixed image and the moving image are subjected to multi-scale Gaussian smoothing to obtain the CT value of each voxel at multiple scales. For any voxel position x in the fixed image, according to the three-dimensional non-rigid deformation field Determine the position of its homologous voxel in the moving image, y= (x) is used to obtain voxel pairs; the CT difference of the voxel pairs at each scale is calculated to form a local difference sequence; wherein, the calculation formula is: k is the scale, k = 0, 1, 2, 3. This is the CT difference. To fix the CT value at any voxel position x in the image after smoothing, The CT value of the homologous voxel position y in the smoothed image; The Wilcoxon signed-rank test was used to perform a nonparametric significance test on the locally differential sequences to obtain a significance value P; If the significance value P is less than a preset significance threshold, the voxels in the fixed image are determined to be abnormal voxels, and the original CT difference between the voxel pairs in the unsmoothed fixed image and the moving image is calculated; the absolute value of the original CT value difference is taken as the abnormal amplitude of the abnormal voxel; when the original CT difference > 0, it indicates an increase in grayscale, suggesting hemorrhage or calcification; when the original CT difference < 0, it indicates a decrease in grayscale, suggesting edema or ischemia; where, ; This is the difference between the original CT scans. The CT value at any voxel position x in a fixed image without smoothing. The CT value for the location y of a homologous voxel in an unsmoothed moving image; Traverse all voxel locations in the fixed image to generate a preliminary list of abnormal brain regions, including voxel location, saliency value, direction of grayscale change, and abnormal amplitude.
[0010] Preferably, the fixed image and the moving image are subjected to multi-scale Gaussian smoothing to obtain the CT value of each voxel at multiple scales, specifically including the following steps: A set of preset standard deviation parameters are selected as the spatial scales of the Gaussian kernel; For each spatial scale, the fixed image and the moving image are independently convolved with a three-dimensional isotropic Gaussian kernel; through the convolution operation, a smooth image at the corresponding scale is generated, in which the CT value of each voxel is replaced by the gray value of its local neighborhood weighted average. Repeat the above steps to obtain the CT values of each voxel at multiple scales.
[0011] Preferably, statistical analysis is performed on the grayscale distribution of each ROI image at the structural scale to generate a list of potential abnormal brain regions, specifically including the following steps: Obtain ROI images of the left and right brain regions of the same type of anatomical structure to form an ROI image group; For each ROI image group, perform the following steps: The CT values of all voxels in the ROI images of the left and right brain regions are extracted respectively, and discretized into bins with a preset CT value interval width to generate corresponding histograms; the histograms are used to represent the voxel frequency distribution within each CT value interval. The two histograms are merged into a contingency table with two rows and K columns; where the rows represent the left brain region ROI image and the right brain region ROI image, the columns represent the CT value binning intervals, and the cell value is the number of voxels in the corresponding region that fall into the CT interval. Fisher's exact test was used to perform a two-tailed test on the contingency table to obtain statistically significant differences in the distribution of CT values between the left and right brain region ROI images. If the statistically significant difference value is less than the preset threshold, the ROI image group is determined to have an overall abnormal gray-scale distribution, and the advantage ratio is calculated. When the advantage ratio is >1, it means that the CT value on the left side is higher than that on the right side, indicating a high-density lesion. When the advantage ratio is <1, it means a low-density lesion. Iterate through all ROI image groups, and then summarize all ROI images that are judged to be abnormal to form a list of potentially abnormal brain regions.
[0012] Preferably, the preliminary list of abnormal brain regions is corrected using a healthy control database, specifically including the following steps: Obtain a pre-built database of plain CT images of healthy volunteers; Using the same deep learning segmentation model as the subjects, the same database of plain CT images from healthy volunteers was used to identify cranial anatomical structures. A Voxelmorph registration model was then used to register the left and right brain region (ROI) images to obtain the correspondence of homologous voxels between the left and right ROI images. Finally, the mean CT value of each ROI image in the healthy population was calculated. μ healthy and standard deviation σ healthy; For ROI images of the same type from the subjects, the average CT values μL and μR corresponding to the left and right brain regions were calculated, and combined with the mean CT values of the corresponding ROI images from a database of plain CT images of healthy volunteers. μ healthy and standard deviation σ Healthy calculated the check value of the subject's left brain region. And right brain region verification value ;in, = ; = ; The check values for the left and right brain regions are compared with preset values, and then the preliminary list of abnormal brain regions is corrected based on the comparison results.
[0013] Preferably, the check values for the left and right brain regions are compared with preset values, and then the preliminary list of abnormal brain regions is corrected based on the comparison results, specifically including: like > Preset value, and If the value is less than or equal to the preset value, the left ROI image is considered abnormal. like > Preset value and If the value is less than or equal to the preset value, the right-side ROI image is considered abnormal. like > Preset value and >Preset value, judged as abnormal bilateral ROI image; A voxel is retained only if the ROI image to which it belongs is determined to be abnormal; otherwise, it is removed from the preliminary list of abnormal brain regions.
[0014] Preferably, cross-validation is performed using the list of potential abnormal brain regions and the corrected preliminary list of abnormal brain regions to output the final abnormal brain region localization result, specifically including the following steps: The list of potential abnormal brain regions and the corrected preliminary list of abnormal brain regions were cross-validated according to the following steps: For the corrected preliminary list of abnormal brain regions, if an abnormal voxel is located within an abnormal ROI image in the list of potential abnormal brain regions, the abnormal voxel is retained; otherwise, it is removed. For a list of potential abnormal brain regions, if the abnormal ROI image contains at least one abnormal voxel from the corrected preliminary list of abnormal brain regions, then the abnormal ROI image is retained; otherwise, it is discarded. The list of potential abnormal brain regions after cross-validation and the preliminary list of abnormal brain regions are used as the final results of abnormal brain region localization.
[0015] Thirdly, embodiments of this application provide an abnormal brain region localization system based on plain CT scan, which includes: The first module is used to process the subject's cranial plain CT image sequence using a deep learning segmentation model to identify and extract ROI images of multiple anatomical structures in the left and right brain regions. The second module is used to register the left and right brain region ROI images using the Voxelmorph registration model in order to obtain the correspondence between homologous voxels between the left and right brain region ROI images. The third module is used to compare the gray value differences between the left and right brain region ROI images at the voxel scale based on the correspondence of homologous voxels between the left and right brain region ROI images, perform significance tests, and generate a preliminary list of abnormal brain regions; at the same time, it performs statistical analysis on the gray value distribution of each ROI image at the structural scale to generate a list of potential abnormal brain regions. The fourth module is used to correct the preliminary list of abnormal brain regions using a healthy control database; and to perform cross-validation using the potential abnormal brain region list and the corrected preliminary list of abnormal brain regions to output the final abnormal brain region localization result.
[0016] Fourthly, embodiments of this application provide a computer-readable storage medium storing a program for locating abnormal brain regions based on plain CT scans, wherein when the program for locating abnormal brain regions based on plain CT scans is executed by a processor, it implements the steps of a method for locating abnormal brain regions based on plain CT scans.
[0017] The beneficial effects of the technical solutions provided in this application include: A multi-level detection framework combining prior correction from a health database and voxel-structural dual-scale fusion cross-validation was constructed. First, a deep learning model was used to segment the anatomical regions of interest (ROIs) of the left and right hemispheres of the subject. Then, a high-precision non-rigid homology mapping was established using an unsupervised Voxelmorph model. Next, voxel-level multi-scale verification was performed in parallel to generate a preliminary list of abnormal brain regions. Simultaneously, statistical analysis of the grayscale distribution of each ROI image was performed at the structural scale to generate a list of potential abnormal brain regions. The preliminary list of abnormal brain regions was then corrected for anatomical rationality using a healthy control database. Finally, bidirectional cross-validation of the potential abnormal brain region list and the corrected preliminary list of abnormal brain regions ensured that the output results simultaneously met the requirements of local saliency and global abnormality. This method requires only conventional plain CT scans and does not require CTP perfusion, significantly reducing radiation and examination barriers. It retains the ability to detect small lesions while effectively suppressing noise and artifact interference, achieving highly accurate, robust, and interpretable localization of abnormal brain regions, suitable for emergency departments, physical examinations, and various types of CT equipment scenarios. Attached Figure Description
[0018] Figure 1 A schematic diagram of unilateral brain segmentation provided for this application; Figure 2 A schematic diagram of the grayscale distribution of the left and right brain regions of a patient's anatomical structure (taking the temporal lobe as an example) provided in this application; Figure 3 Gray-scale distribution images of the occipital lobe of multiple healthy volunteers provided for this application; Figure 4 Gray-scale distribution images of the temporal lobes of multiple healthy volunteers provided for this application; Figure 5 This is a flowchart illustrating the abnormal brain region localization method based on plain CT scan involved in the embodiments of this application. Detailed Implementation
[0019] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present application.
[0020] First, some of the technical terms used in this application will be explained to help those skilled in the art understand this application.
[0021] Region of Interest (ROI) is an image analysis method used in machine vision and image processing to delineate regions to be processed. It uses bounding boxes, circles, ellipses, or irregular polygons to outline specific areas, reducing computational load and focusing on key information processing. This region can be obtained through pixel matrix cropping or operator calculation, and is applied in fields such as face recognition, image segmentation, and object detection, supporting dynamic tracking and multi-region processing.
[0022] Voxel-level (Local Scale) Definition: Using a single voxel (3D pixel, e.g., 1×1×1mm³) as the unit of analysis, the differences in CT values (HU values) at corresponding locations in the left and right hemispheres are compared point by point. Operation: For each left hemisphere voxel x, its anatomically homologous voxel y in the right hemisphere is found, and d=CT(x) is calculated. CT(y) was calculated and a statistical significance test was performed.
[0023] Features: Highest spatial resolution (down to sub-millimeter level); capable of detecting tiny lesions (such as bleeding points <5mm); sensitive to noise, prone to isolated false positives.
[0024] Structural scale (Anatomical Structure-level / Regional Scale) Definition: Using the entire anatomical region (e.g., "left temporal lobe") as the unit of analysis, compare the differences in the overall grayscale distribution of the same brain region on the left and right sides.
[0025] Operation method: Extract the CT value histogram of all voxels in the left temporal lobe and perform statistical tests (such as Fisher's exact test) with the histogram of the right temporal lobe to determine whether there is an overall shift.
[0026] Characteristics: Low spatial resolution (in units of ROI); sensitive to large-scale or diffuse lesions; It has strong noise resistance, but may miss small focal lesions.
[0027] The specific details of the training of the Voxelmorph registration model are not the focus of this application; other models with the same function as the Voxelmorph registration model can also be applied.
[0028] If Voxelmorph is not provided in the following text Without this, it becomes impossible to determine which right-brain voxel is homologous to left-brain voxel x, rendering the comparison anatomically meaningless. Therefore, voxel-scale detection relies entirely on point-by-point mappings established by Voxelmorph. Although structural scales perform statistical analysis on the entire region of interest (ROI), the left and right ROIs must be anatomically aligned; otherwise, the left and right temporal lobes may contain different tissues (e.g., the left lobes contain the insula, while the right lobes do not), leading to incomparable histograms and invalidating statistical tests.
[0029] The Voxelmorph registration model ensures that the left temporal lobe ROI and the deformed right temporal lobe ROI are highly anatomically overlapped, making the comparison between the left and right temporal lobes medically reasonable.
[0030] To make the objectives, technical solutions, and advantages of this application clearer, the embodiments of this application will be described in further detail below with reference to the accompanying drawings.
[0031] Firstly, reference Figure 5 As shown, this application provides a method for locating abnormal brain regions based on plain CT scans, which includes: Step 100: Process the subject's cranial CT scan image sequence using a deep learning segmentation model to identify and extract Regions of Interest (ROIs) images of multiple anatomical structures in the left and right brain regions. Deep learning segmentation models such as U-Net and AttentionU-Net are existing models. For ROI images of multiple anatomical structures, refer to [reference needed]. Figure 1 ; Step 200: Use the Voxelmorph registration model to register the left and right ROI images to obtain the correspondence between homologous voxels between the left and right ROI images. Step 300: Based on the correspondence of homologous voxels between the left and right brain region ROI images, compare the differences in gray values between the left and right brain region ROI images at the voxel scale, perform a significance test, and generate a preliminary list of abnormal brain regions; at the same time, perform statistical analysis on the gray distribution of each ROI image at the structural scale to generate a list of potential abnormal brain regions. Step 400: Correct the preliminary list of abnormal brain regions using a healthy control database; perform cross-validation using the potential abnormal brain region list and the corrected preliminary list of abnormal brain regions to output the final abnormal brain region localization result.
[0032] The above constructs a multi-level anomaly detection framework that combines prior correction of a health database with voxel-structure dual-scale fusion cross-validation. First, a deep learning model is used to segment the whole-brain anatomical structures from plain CT scans of subjects, obtaining left and right Regions of Interest (ROI) images. Then, a non-rigid homology mapping is established using Voxelmorph. Based on this, voxel-level Wilcoxon tests and ROI-level Fisher distribution tests are performed in parallel to generate preliminary voxel anomaly lists and potential structural anomaly lists, respectively. Finally, statistical benchmarks from healthy individuals are introduced for correction, and cross-validation is achieved through bidirectional logical constraints.
[0033] This architecture effectively addresses the inherent limitations of single-scale methods. Voxel methods are sensitive but susceptible to noise interference, while structural methods are robust but lack localization accuracy. By employing a dual-scale complementarity, it retains the ability to detect minute lesions while avoiding false positives caused by local artifacts. Health database correction eliminates physiological asymmetries (such as calcification of normal choroid plexus), and cross-validation ensures that all output anomalies simultaneously satisfy both local significance and global abnormality as evidence.
[0034] Compared with existing abnormal brain region identification techniques that rely on CTP perfusion imaging or traditional voxel-by-voxel comparison methods, this invention has the following significant technical advantages: (1) Rapid anomaly localization can be achieved without CTP. This invention is based entirely on conventional plain CT images. It can identify abnormal brain regions by using deep learning brain region segmentation, deep registration, and multi-scale statistical analysis. This avoids the problems of high radiation, complex examination procedures, and high requirements for patient condition in CTP examination, and greatly improves the universality and clinical accessibility of the examination.
[0035] (2) This invention uses the Voxelmorph registration model based on unsupervised deformation field prediction to achieve automatic, efficient and stable voxel-level registration between the left and right brain regions. Compared with traditional optimization-based methods, it has significantly improved in speed, robustness and stability, and has established a reliable spatial consistency foundation for subsequent difference analysis.
[0036] (3) By combining voxel-by-voxel difference detection at the voxel scale with overall grayscale distribution verification at the structural scale, this invention effectively overcomes the instability of single-layer verification caused by CT noise, voxel random error, and differences in reconstruction algorithms, thereby improving the overall accuracy and robustness of anomaly localization. In addition, by introducing a healthy control database and using the grayscale distribution of brain regions of healthy volunteers as a reference, the source of the anomaly (left / right / bilateral) can be determined simultaneously, and the voxel scale can be corrected to avoid misjudgment that may occur if only left-right differences are relied upon.
[0037] (4) This method relies only on plain CT image data and can be run on various types of equipment such as traditional single-energy CT, dual-energy CT and photon counting CT, without being restricted by manufacturers; it is also applicable to various clinical procedures such as physical examination, emergency, and outpatient follow-up.
[0038] Therefore, it effectively solves the technical problems of difficulty in locating abnormal brain regions, high noise, difficulty in establishing symmetry relationships, and unstable results under plain CT conditions, and realizes low radiation, high efficiency, and high accuracy in detecting abnormal brain regions.
[0039] In some embodiments, step 200 involves registering the left and right ROI images using the Voxelmorph registration model to obtain the correspondence between homologous voxels between the left and right ROI images. This specifically includes the following steps: A ROI image from one side of the brain was selected as the fixed image, and the corresponding ROI image from the other side was selected as the moving image. The fixed and moving images were then input into the Voxelmorph registration model to output a three-dimensional non-rigid deformation field. Three-dimensional non-rigid deformation fields are used to characterize the voxel-level mapping function relationship from the space of a fixed image to the space of a moving image. According to the three-dimensional non-rigid deformation field Establish voxel correspondences according to the following pairing steps: For any voxel position x in a fixed image, its corresponding homologous voxel position in the moving image is y= (x); For any voxel position y in the moving image, its corresponding voxel position in the fixed image is obtained by inverse mapping of the deformation field, x = (y).
[0040] In some embodiments, based on the correspondence of homologous voxels between left and right brain region ROI images, the grayscale differences between the left and right brain region ROI images are compared at the voxel scale to perform a significance test and generate a preliminary list of abnormal brain regions. This specifically includes the following steps: The fixed image and the moving image are subjected to multi-scale Gaussian smoothing to obtain the CT value of each voxel at multiple scales. For any voxel position x in the fixed image, according to the three-dimensional non-rigid deformation field Determine the position of its homologous voxel in the moving image, y= (x) is used to obtain voxel pairs; the CT difference of the voxel pairs at each scale is calculated to form a local difference sequence; wherein, the calculation formula is: k is the scale, k = 0, 1, 2, 3. This is the CT difference. To fix the CT value at any voxel position x in the image after smoothing, The CT value of the homologous voxel position y in the smoothed image; The Wilcoxon signed-rank test was used to perform a nonparametric significance test on the locally differential sequences to obtain a significance value P; If the significance value P is less than a preset significance threshold, the voxels in the fixed image are determined to be abnormal voxels, and the original CT difference between the voxel pairs in the unsmoothed fixed image and the moving image is calculated; the absolute value of the original CT value difference is taken as the abnormal amplitude of the abnormal voxel; when the original CT difference > 0, it indicates an increase in grayscale, suggesting hemorrhage or calcification; when the original CT difference < 0, it indicates a decrease in grayscale, suggesting edema or ischemia; where, ; This is the difference between the original CT scans. The CT value at any voxel position x in a fixed image without smoothing. The CT value for the location y of a homologous voxel in an unsmoothed moving image; Traverse all voxel locations in the fixed image to generate a preliminary list of abnormal brain regions, including voxel location, saliency value, direction of grayscale change, and abnormal amplitude.
[0041] The selected ROI image (e.g., the left side) is the fixed image, and the opposite side is the moving image, which is then input into the Voxelmorph registration model. This model learns the dense 3D deformation field from the fixed domain to the moving domain end-to-end by minimizing similarity loss (e.g., NCC) and deformation regularization terms (e.g., diffusion or bending energy). Only forward mapping y= (x) can establish the homologous position y of each voxel x on the opposite side in a fixed image. The Wilcoxon signed-rank test does not depend on the assumption of normal distribution of data and is robust to small samples and CT imaging noise.
[0042] Voxelmorph employs unsupervised training, requiring no paired labels, making it suitable for real-world clinical scenarios. Its non-rigid deformation capability accurately aligns structural shifts caused by individual anatomical differences (such as hippocampal atrophy and ventricular enlargement). Using only unidirectional mapping significantly reduces computational complexity and fully meets the requirement of subsequently traversing only fixed-side voxels. This registration strategy provides geometric consistency guarantees for voxel-level comparisons, a prerequisite for subsequent high-precision anomaly detection. In this embodiment, it should be understood that: Although the left and right hemispheres exist in the same visual space, the normal human brain exhibits physiological asymmetry (e.g., a larger left temporal plane and a more prominent right parietal lobe); lesions (such as tumors or atrophy) can further exacerbate this structural shift; for example: The 100th voxel in the left precentral gyrus of the left brain may be located at (80, 120, 90) on spatial coordinates. Its anatomically homologous region (i.e., the "right precentral gyrus corresponding part" of the right brain) may be located at (85, 118, 92) - not a mirror-symmetric point.
[0043] The Voxelmorph registration model establishes anatomically homologous voxel mappings, rather than spatial coordinate alignment, and outputs a three-dimensional non-rigid deformation field. , so that: For any anatomical location x in the left brain, the most functionally / structurally similar corresponding location in the right brain can be found. (x), thus achieving anatomically aligned left and right, rather than a simple mirror flip or rigid symmetry. Specific operation method (with the left side as the fixed side): Extract the left brain region of interest (ROI) as a fixed image. Extracting right brain regions of interest (ROIs) as moving images; Voxelmorph predicting deformation fields. The right brain is non-rigidly deformed into the space of the left brain; at this time, the left brain voxels and the deformed right brain voxels at the same spatial coordinates (x,y,z) are anatomically comparable.
[0044] In some embodiments, multi-scale Gaussian smoothing is performed on the fixed image and the moving image respectively to obtain the CT value of each voxel at multiple scales, specifically including the following steps: A set of preset standard deviation parameters are selected as the spatial scales of the Gaussian kernel; For each spatial scale, the fixed image and the moving image are independently convolved with a three-dimensional isotropic Gaussian kernel; through the convolution operation, a smooth image at the corresponding scale is generated, in which the CT value of each voxel is replaced by the gray value of its local neighborhood weighted average. Repeat the above steps to obtain the CT values of each voxel at multiple scales.
[0045] Wherein, the standard deviation parameter σ = {0mm, 1mm, 2mm, 3mm}; Small-scale (σ=0–1mm) captures focal abnormalities (such as microbleeds), while large-scale (σ=2–3mm) suppresses quantum noise and enhances diffuse signals, thereby enhancing the sensitivity to lesions at different spatial scales (such as microbleeds and diffuse edema).
[0046] In some embodiments, statistical analysis of the grayscale distribution of each ROI image at the structural scale is performed to generate a list of potential abnormal brain regions, specifically including the following steps: Obtain ROI images of the left and right brain regions of the same type of anatomical structure to form an ROI image group; For each ROI image group, perform the following steps: The CT values of all voxels in the ROI images of the left and right brain regions are extracted respectively, and discretized into bins with a preset CT value interval width to generate corresponding histograms; the histograms are used to represent the voxel frequency distribution (i.e., the distribution of the number of voxels) within each CT value interval. The two histograms are merged into a contingency table with two rows and K columns; where the rows represent the left and right brain region ROI images, the columns represent the CT value binning intervals, and the cell value is the number of voxels within the corresponding CT interval; see reference. Figure 2 A contingency table of anatomical structures of a patient (temporal lobe); Fisher's exact test was used to perform a two-tailed test on the contingency table to obtain statistically significant differences in the distribution of CT values between the left and right brain region ROI images. If the statistically significant difference value is less than the preset threshold, the ROI image group is determined to have an overall abnormal gray-scale distribution, and the odds ratio is calculated. When the odds ratio is greater than 1, it indicates that the CT value on the left side is higher than that on the right side, suggesting a high-density lesion, such as hemorrhage. When the odds ratio is less than 1, it indicates a low-density lesion, such as edema or infarction. Iterate through all ROI image groups, and then summarize all ROI images that are judged to be abnormal to form a list of potentially abnormal brain regions.
[0047] The Fisher's exact test method described above is suitable for small sample and sparse frequency scenarios. It effectively avoids the bias of the traditional chi-square test in low-frequency cells and, under the assumption of no difference between the left and right distributions, observes the probability value of the current or more extreme distribution. Furthermore, the reason for performing structural scale detection above is: When doctors review CT scans, they not only focus on whether a single point is abnormal, but also on whether the entire temporal lobe is darker / brighter than the contralateral side. This is because certain lesions (such as large-area cerebral infarction, tumor edema, and diffuse hemorrhage) can cause a systematic shift in the grayscale distribution of the entire brain region. Even if the changes in individual voxels are not significant, the overall trend is still substantial.
[0048] However, relying solely on the voxel scale may lead to false positives due to noise or minor registration errors; if the lesion is diffuse but small in amplitude, the voxel test may not be significant, resulting in missed diagnoses. Therefore, the method described in this application was developed, which treats the entire ROI image as a statistical unit and compares whether the CT value distribution of the corresponding left and right brain regions is consistent.
[0049] The reason for using histograms in this scheme is that CT values are continuous variables, and directly comparing the means would lose information about the distribution shape (such as bimodality or skewness). Histograms discretize CT values into several bins to form a frequency distribution. This preserves grayscale information and facilitates statistical modeling.
[0050] In some embodiments, the preliminary list of abnormal brain regions is corrected using a healthy control database, specifically including the following steps: Obtain a pre-constructed database of plain CT images of healthy volunteers; this database contains plain CT images of at least 200 healthy volunteers, selected based on age ± 5 years and gender matching; the mean CT value of the ROI images in the healthy population is then calculated. μ healthy and standard deviation σ Healthy can be achieved through Figure 3 and Figure 4 Calculations show that the different colored dots and horizontal lines represent brain regions of healthy individuals on different CT scanners (a total of 18 in the image). Figure 3 It is the occipital lobe. Figure 4 This is the grayscale distribution of the temporal lobe, where the vertical axis is the logarithm of the voxel count, and the x-line represents the median distribution. The current project focuses on CT scans (5 units) of brain regions in healthy individuals whose distribution closely approximates the x-line. The x-line also serves as the standard distribution for this brain region in healthy individuals. Using the same deep learning segmentation model as the subjects, the same database of plain CT images from healthy volunteers was used to identify cranial anatomical structures. A Voxelmorph registration model was then used to register the left and right brain region (ROI) images to obtain the correspondence of homologous voxels between the left and right ROI images. Finally, the mean CT value of each ROI image in the healthy population was calculated. μ healthy and standard deviation σ healthy; For ROI images of the same type from the subjects, the average CT values μL and μR corresponding to the left and right brain regions were calculated, and combined with the mean CT values of the corresponding ROI images from a database of plain CT images of healthy volunteers. μ healthy and standard deviation σ Healthy calculated the check value of the subject's left brain region. And right brain region verification value ;in, = ; = ; The check values for the left and right brain regions are compared with preset values, and then the preliminary list of abnormal brain regions is corrected based on the comparison results.
[0051] In some embodiments, the check values for the left and right brain regions are compared with preset values, and then the preliminary list of abnormal brain regions is corrected based on the comparison results, specifically including: like > Preset value, and If the value is less than or equal to the preset value, the left ROI image is considered abnormal. like > Preset value and If the value is less than or equal to the preset value, the right-side ROI image is considered abnormal. like > Preset value and >Preset value, judged as abnormal bilateral ROI image; A voxel is retained only if the ROI image to which it belongs is determined to be abnormal; otherwise, it is removed from the preliminary list of abnormal brain regions.
[0052] In this embodiment, the Wilcoxon test is very sensitive at the voxel level and can capture minute asymmetries; however, it is easily affected by CT noise, local artifacts, registration residuals, etc., which can produce isolated false positive voxels, but there are no abnormalities around them and the overall ROI image is normal; these false positives need to be removed.
[0053] Structural scale results are highly specific but coarse, unable to pinpoint specific abnormal locations, and can only determine whether the entire ROI is abnormal; therefore, it is not used for self-correction, and correction of the preliminary list of abnormal brain regions is required first.
[0054] In some embodiments, cross-validation is performed using the list of potential abnormal brain regions and the corrected preliminary list of abnormal brain regions to output the final abnormal brain region localization result, specifically including the following steps: The list of potential abnormal brain regions and the corrected preliminary list of abnormal brain regions were cross-validated according to the following steps: For the corrected preliminary list of abnormal brain regions, if an abnormal voxel is located within an abnormal ROI image in the list of potential abnormal brain regions, the abnormal voxel is retained; otherwise, it is removed. For a list of potential abnormal brain regions, if the abnormal ROI image contains at least one abnormal voxel from the corrected preliminary list of abnormal brain regions, then the abnormal ROI image is retained; otherwise, it is discarded. The list of potential abnormal brain regions after cross-validation and the preliminary list of abnormal brain regions are used as the final results of abnormal brain region localization.
[0055] The above cross-validation leverages the advantages of voxel-scale and structural-scale detection, resulting in accurate localization results. Furthermore, the final abnormal brain region localization results are associated with anatomical name, laterality, degree, and pathological direction, forming structured data that can be directly used for report generation.
[0056] The method described in this application is applicable to various types of equipment, including conventional CT, single-energy CT, dual-energy CT, and photon-counting CT. The entire process can be automated using a deep learning framework without manual intervention and can be deployed in PACS workstations, image intelligent analysis platforms, or medical device terminals. This invention is not limited to the specific model described (for example, the brain region segmentation model can be replaced with other segmentation networks, Voxelmorph can be replaced with other deep deformation registration frameworks, and statistical methods can be replaced with other hypothesis testing methods); all such modifications and substitutions fall within the scope of protection of this invention.
[0057] Secondly, embodiments of this application also provide a system for locating abnormal brain regions based on plain CT scans. It includes: The first module is used to process the subject's cranial plain CT image sequence using a deep learning segmentation model to identify and extract ROI images of multiple anatomical structures in the left and right brain regions. The second module is used to register the left and right brain region ROI images using the Voxelmorph registration model in order to obtain the correspondence between homologous voxels between the left and right brain region ROI images. The third module is used to compare the gray value differences between the left and right brain region ROI images at the voxel scale based on the correspondence of homologous voxels between the left and right brain region ROI images, perform significance tests, and generate a preliminary list of abnormal brain regions; at the same time, it performs statistical analysis on the gray value distribution of each ROI image at the structural scale to generate a list of potential abnormal brain regions. The fourth module is used to correct the preliminary list of abnormal brain regions using a healthy control database; and to perform cross-validation using the potential abnormal brain region list and the corrected preliminary list of abnormal brain regions to output the final abnormal brain region localization result.
[0058] Thirdly, embodiments of this application provide an abnormal brain region localization device based on plain CT scan. The abnormal brain region localization device based on plain CT scan can be a personal computer (PC), laptop computer, server, or other device with data processing capabilities.
[0059] In this embodiment of the application, the abnormal brain region localization device based on plain CT scan may include a processor, a memory, a communication interface, and a communication bus.
[0060] The communication bus can be of any type and is used to interconnect the processor, memory, and communication interface.
[0061] The communication interface includes input / output (I / O) interfaces, physical interfaces, and logical interfaces used for interconnecting internal components of the plain CT-based abnormal brain region localization device, as well as interfaces used for interconnecting the device with other devices (such as other computing devices or user equipment). Physical interfaces can be Ethernet interfaces, fiber optic interfaces, ATM interfaces, etc.; user equipment can be displays, keyboards, etc.
[0062] Memory can be various types of storage media, such as random access memory (RAM), read-only memory (ROM), non-volatile RAM (NVRAM), flash memory, optical storage, hard disk, programmable ROM (PROM), erasable PROM (EPROM), electrically erasable PROM (EEPROM), etc.
[0063] The processor can be a general-purpose processor, which can call a program for locating abnormal brain regions based on plain CT scans stored in memory and execute the method for locating abnormal brain regions based on plain CT scans provided in the embodiments of this application. For example, the general-purpose processor can be a central processing unit (CPU). The method executed when the program for locating abnormal brain regions based on plain CT scans is called can be referred to in the various embodiments of the method for locating abnormal brain regions based on plain CT scans in this application, and will not be repeated here.
[0064] Fourthly, embodiments of this application also provide a computer-readable storage medium.
[0065] The present application has a computer-readable storage medium storing a program for locating abnormal brain regions based on plain CT scans, wherein when the program for locating abnormal brain regions based on plain CT scans is executed by a processor, it implements the steps of the method for locating abnormal brain regions based on plain CT scans as described above.
[0066] The method implemented when the abnormal brain region localization procedure based on plain CT is executed can be referred to in the various embodiments of the abnormal brain region localization method based on plain CT of this application, and will not be repeated here.
[0067] It should be noted that the sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.
[0068] The terms "comprising" and "having," and any variations thereof, in the specification, claims, and accompanying drawings of this application are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to such process, method, product, or apparatus. The terms "first," "second," and "third," etc., are used to distinguish different objects, etc., and do not indicate a sequence, nor do they limit "first," "second," and "third" to different types.
[0069] In the description of the embodiments of this application, terms such as "exemplary," "for example," or "for instance" are used to indicate examples, illustrations, or explanations. Any embodiment or design described as "exemplary," "for example," or "for instance" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of terms such as "exemplary," "for example," or "for instance" is intended to present the relevant concepts in a concrete manner.
[0070] In the description of the embodiments of this application, unless otherwise stated, " / " means "or". For example, A / B can mean A or B. The "and / or" in the text is merely a description of the relationship between related objects, indicating that there can be three relationships. For example, A and / or B can mean: A exists alone, A and B exist simultaneously, and B exists alone. In addition, in the description of the embodiments of this application, "multiple" means two or more.
[0071] In some processes described in the embodiments of this application, multiple operations or steps are included in a specific order. However, it should be understood that these operations or steps may not be executed in the order they appear in the embodiments of this application, or they may be executed in parallel. The sequence number of the operation is only used to distinguish different operations, and the sequence number itself does not represent any execution order. In addition, these processes may include more or fewer operations, and these operations or steps may be executed sequentially or in parallel, and these operations or steps may be combined.
[0072] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) as described above, and includes several instructions to cause a terminal device to execute the methods described in the various embodiments of this application.
[0073] The above are merely preferred embodiments of this application and do not limit the patent scope of this application. Any equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.
Claims
1. A method for locating abnormal brain regions based on plain CT, characterized in that, It includes: Deep learning segmentation models were used to process the sequence of plain CT images of the brain of the subjects to identify and extract ROI images of multiple anatomical structures in the left and right brain regions; The Voxelmorph registration model is used to register the Region of Interest (ROI) images of the left and right brain regions to obtain the correspondence of homologous voxels between the ROI images. This step specifically includes: selecting the ROI image of one brain region as the fixed image and the corresponding ROI image of the other brain region as the moving image; inputting the fixed and moving images into the Voxelmorph registration model to output a three-dimensional non-rigid deformation field. A three-dimensional non-rigid deformation field is used to characterize the voxel-level mapping function relationship from the space of a fixed image to the space of a moving image; based on the three-dimensional non-rigid deformation field... Establish voxel correspondences by following the pairing steps: For any voxel position x in a fixed image, its corresponding homologous voxel position y in the moving image is... (x); For any voxel position y in the moving image, its corresponding voxel position in the fixed image is obtained by inverse mapping of the deformation field, x = (y); Based on the correspondence of homologous voxels between left and right brain region ROI images, the grayscale differences between the left and right brain region ROI images are compared at the voxel scale, and a significance test is performed to generate a preliminary list of abnormal brain regions. Simultaneously, at the structural scale, statistical analysis is performed on the grayscale distribution of each ROI image to generate a list of potential abnormal brain regions. Specifically, generating the preliminary list of abnormal brain regions includes: performing multi-scale Gaussian smoothing on both the fixed and moving images to obtain the CT value of each voxel at multiple scales; and for any voxel position x in the fixed image, based on the three-dimensional non-rigid deformation field... Determine the position of its homologous voxel in the moving image, y= (x) is used to obtain voxel pairs; the CT difference of the voxel pairs at each scale is calculated to form a local difference sequence; wherein, the calculation formula is: k is the scale, k = 0, 1, 2, 3; This refers to the CT difference. To fix the CT value at any voxel position x in the image after smoothing; The CT value of the homologous voxel position y in the smoothed moving image is used; the Wilcoxon signed-rank test is used to perform a nonparametric significance test on the local difference sequence to obtain a significance value P; if the significance value P is less than a preset significance threshold, the voxel in the fixed image is determined to be an abnormal voxel, and the original CT difference between the voxel pair in the unsmoothed fixed image and the moving image is calculated; the absolute value of the difference in original CT values is taken as the abnormal amplitude of the abnormal voxel; when the original CT difference > 0, it indicates an increase in gray level, suggesting hemorrhage or calcification; when the original CT difference < 0, it indicates a decrease in gray level, suggesting edema or ischemia; where, ; This is the difference between the original CT scans. The CT value at any voxel position x in a fixed image without smoothing. The CT value of the homologous voxel position y in the unsmoothed moving image is obtained; all voxel positions in the fixed image are traversed to generate a preliminary list of abnormal brain regions containing voxel positions, saliency values, gray-scale change directions, and abnormal amplitudes. The preliminary list of abnormal brain regions is corrected using a healthy control database; cross-validation is performed using the potential abnormal brain region list and the corrected preliminary abnormal brain region list to output the final abnormal brain region localization results.
2. The method of claim 1, wherein the step of locating the abnormal brain region based on the non-contrast CT is performed by using a method comprising: The fixed and moving images are subjected to multi-scale Gaussian smoothing to obtain the CT value of each voxel at multiple scales. Specifically, this includes the following steps: A set of preset standard deviation parameters are selected as the spatial scales of the Gaussian kernel; For each spatial scale, the fixed image and the moving image are independently convolved with a three-dimensional isotropic Gaussian kernel; through the convolution operation, a smooth image at the corresponding scale is generated, in which the CT value of each voxel is replaced by the gray value of its local neighborhood weighted average. Repeat the above steps to obtain the CT values of each voxel at multiple scales.
3. The method of claim 1, wherein the step of locating the abnormal brain region based on the plain CT is characterized by, Statistical analysis of the grayscale distribution of each ROI image at the structural scale is performed to generate a list of potential abnormal brain regions. This includes the following steps: Obtain ROI images of the left and right brain regions of the same type of anatomical structure to form an ROI image group; For each ROI image group, perform the following steps: The CT values of all voxels in the ROI images of the left and right brain regions are extracted respectively, and discretized into bins with a preset CT value interval width to generate corresponding histograms; the histograms are used to represent the voxel frequency distribution within each CT value interval. The two histograms are merged into a contingency table with two rows and K columns; where the rows represent the left brain region ROI image and the right brain region ROI image, the columns represent the CT value binning intervals, and the cell value is the number of voxels in the corresponding region that fall into the CT interval. Fisher's exact test was used to perform a two-tailed test on the contingency table to obtain statistically significant differences in the distribution of CT values between the left and right brain region ROI images. If the statistically significant difference value is less than the preset threshold, the ROI image group is determined to have an overall abnormal gray-scale distribution, and the advantage ratio is calculated. When the advantage ratio is >1, it means that the CT value on the left side is higher than that on the right side, indicating a high-density lesion. When the advantage ratio is <1, it means a low-density lesion. Iterate through all ROI image groups, and then summarize all ROI images that are judged to be abnormal to form a list of potentially abnormal brain regions.
4. The method of claim 1, wherein the method is based on a non-contrast CT. The preliminary list of abnormal brain regions was corrected using a healthy control database, specifically including the following steps: Obtain a pre-built database of plain CT images of healthy volunteers; Using the same deep learning segmentation model as the subjects, the same database of plain CT images from healthy volunteers was used to identify cranial anatomical structures. A Voxelmorph registration model was then used to register the left and right brain region (ROI) images to obtain the correspondence of homologous voxels between the left and right ROI images. Finally, the mean CT value of each ROI image in the healthy population was calculated. μ healthy and standard deviation σ healthy; For ROI images of the same type from the subjects, the average CT values μL and μR corresponding to the left and right brain regions were calculated, and combined with the mean CT values of the corresponding ROI images from a database of plain CT images of healthy volunteers. μ healthy and standard deviation σ Healthy calculated the check value of the subject's left brain region. And right brain region verification value ;in, = ; = ; The check values for the left and right brain regions are compared with preset values, and then the preliminary list of abnormal brain regions is corrected based on the comparison results.
5. The method of claim 4, wherein the step of locating the abnormal brain region based on the plain CT is characterized by, The check values for the left and right brain regions are compared with preset values, and then the preliminary list of abnormal brain regions is corrected based on the comparison results. Specifically, this includes: If > preset value, and ≤ preset value, it is determined that the left ROI image is abnormal. If > preset value and ≤ preset value, it is determined that the right ROI image is abnormal. If > the preset value and > the preset value, it is determined that the bilateral ROI image is abnormal. A voxel is retained only if the ROI image to which it belongs is determined to be abnormal; otherwise, it is removed from the preliminary list of abnormal brain regions.
6. The non-enhanced CT-based abnormal brain region localization method of claim 1, wherein, Cross-validation is performed using the list of potential abnormal brain regions and the corrected preliminary list of abnormal brain regions to output the final abnormal brain region localization results. This process includes the following steps: The list of potential abnormal brain regions and the corrected preliminary list of abnormal brain regions were cross-validated according to the following steps: For the corrected preliminary list of abnormal brain regions, if an abnormal voxel is located within an abnormal ROI image in the list of potential abnormal brain regions, the abnormal voxel is retained; otherwise, it is removed. For a list of potential abnormal brain regions, if the abnormal ROI image contains at least one abnormal voxel from the corrected preliminary list of abnormal brain regions, then the abnormal ROI image is retained; otherwise, it is discarded. The list of potential abnormal brain regions after cross-validation and the preliminary list of abnormal brain regions are used as the final results of abnormal brain region localization.
7. A non-contrast CT based abnormal brain region localization system, characterized in that, It includes: The first module is used to process the subject's cranial plain CT image sequence using a deep learning segmentation model to identify and extract ROI images of multiple anatomical structures in the left and right brain regions. The second module is used to register the left and right brain region ROI images using the Voxelmorph registration model to obtain the correspondence between homologous voxels between the left and right brain region ROI images. This step specifically includes: selecting the ROI image of one brain region as the fixed image and the corresponding ROI image of the other brain region as the moving image, and inputting the fixed image and the moving image into the Voxelmorph registration model to output a three-dimensional non-rigid deformation field. A three-dimensional non-rigid deformation field is used to characterize the voxel-level mapping function relationship from the space of a fixed image to the space of a moving image; according to the three-dimensional non-rigid deformation field... Establish voxel correspondences according to the following pairing steps: For any voxel position x in a fixed image, its homologous voxel position in the moving image is y= (x); For any voxel position y in the moving image, its corresponding voxel position in the fixed image is obtained by inverse mapping of the deformation field, x = (y); The third module is used to compare the grayscale differences between the left and right brain region ROI images at the voxel scale based on the correspondence of homologous voxels between the left and right brain region ROI images, perform significance testing, and generate a preliminary list of abnormal brain regions. Simultaneously, it performs statistical analysis on the grayscale distribution of each ROI image at the structural scale to generate a list of potential abnormal brain regions. Specifically, generating the preliminary list of abnormal brain regions includes: performing multi-scale Gaussian smoothing on the fixed and moving images to obtain the CT value of each voxel at multiple scales; and for any voxel position x in the fixed image, based on the three-dimensional non-rigid deformation field... Determine the position of its homologous voxel in the moving image, y= (x) is used to obtain voxel pairs; the CT difference of the voxel pairs at each scale is calculated to form a local difference sequence; wherein, the calculation formula is: k is the scale, k = 0, 1, 2, 3; This refers to the CT difference. To fix the CT value at any voxel position x in the image after smoothing; The CT value of the homologous voxel position y in the smoothed moving image is used; the Wilcoxon signed-rank test is used to perform a nonparametric significance test on the local difference sequence to obtain a significance value P; if the significance value P is less than a preset significance threshold, the voxel in the fixed image is determined to be an abnormal voxel, and the original CT difference between the voxel pair in the unsmoothed fixed image and the moving image is calculated; the absolute value of the difference in original CT values is taken as the abnormal amplitude of the abnormal voxel; when the original CT difference > 0, it indicates an increase in gray level, suggesting hemorrhage or calcification; when the original CT difference < 0, it indicates a decrease in gray level, suggesting edema or ischemia; where, ; This is the difference between the original CT scans. The CT value at any voxel position x in a fixed image without smoothing. The CT value of the homologous voxel position y in the unsmoothed moving image is obtained; all voxel positions in the fixed image are traversed to generate a preliminary list of abnormal brain regions containing voxel positions, saliency values, gray-scale change directions, and abnormal amplitudes. The fourth module is used to correct the preliminary list of abnormal brain regions using a healthy control database; and to perform cross-validation using the potential abnormal brain region list and the corrected preliminary list of abnormal brain regions to output the final abnormal brain region localization result.
8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a program for locating abnormal brain regions based on plain CT scans, wherein when the program is executed by a processor, it implements the steps of the method for locating abnormal brain regions based on plain CT scans as described in any one of claims 1 to 6.