Automated ASPECTS scoring method based on plain scan CT, computer device, readable storage medium and program product
Through the automated ASPECTS scoring method, deep learning model and image registration technology, the problems of poor stability and time-consuming artificial reading maps are solved, and fast and accurate ASPECTS scoring is achieved, which is suitable for the diagnosis of acute stroke.
Patent Information
- Application Number
- CN202210406433.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-04-18
- Publication Date
- 2024-11-22
- Estimated Expiration
- 2042-04-18
AI Technical Summary
The existing ASPECTS scoring methods rely on manual reading of pictures, which have poor stability and time-consuming problems, making it difficult to quickly and accurately score in acute stroke.
An automated ASPECTS scoring method based on plain scanning CT was adopted to filter keyframes through deep learning models, locate brain flies, perform image registration, and use multiple scoring models to score partitions to improve the accuracy and efficiency of scoring.
A fast, accurate and stable ASPECTS score is achieved, reducing artificial interference, improving diagnosis speed, and suitable for cases with asymmetric brain shape.
Smart Images

Figure CN114708240B_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the field of medical engineering, and in particular to an automated ASPECTS scoring method based on plain scan CT, a computer device, a readable storage medium and a program product. Background Art
[0002] Stroke, also known as infarction or cerebral infarction, is a major disease that seriously threatens population health and hinders social and economic development. Acute ischemic stroke is the most common type of stroke, accounting for about 60% to 80% of all strokes. It is an acute cerebrovascular disease caused by insufficient blood supply to local brain tissue and has a very high risk of death.
[0003] Non-Contrast CT (NCCT) is the most common imaging technology and method used to diagnose brain lesions. It has the advantages of fast imaging speed, convenient scanning and relatively low price. The ASPECTS (Alberta Stroke Program Early CT Score) based on NCCT is an important basis for the diagnosis and treatment of ischemic stroke.
[0004] This scoring method is for patients with acute stroke. Based on their brain NCCT imaging data, the important levels of the middle cerebral artery blood supply are divided into 10 areas, including the caudate nucleus head (C), lenticular nucleus (L), internal capsule (IC), insular zone (I), M1 (anterior cortical area of the middle cerebral artery), M2 (lateral cortical area of the middle cerebral artery insular lobe), M3 (posterior cortical area of the middle cerebral artery), and M4 (middle cerebral artery cortex above M1), M5 (middle cerebral artery cortex above M2) and M6 (middle cerebral artery cortex above M3) at the level of the superior ganglia. The above 10 areas have the same weight, each accounting for 1 point, and the total score is 10 points. The number of areas with early ischemic changes is subtracted from the total score, and the resulting value is used as the scoring result to provide a basis for disease diagnosis and treatment.
[0005] In current clinical applications, the ASPECTS scoring method still relies mainly on clinicians to manually read the images to judge each area for evaluation. On the one hand, due to differences in imaging equipment, differences in patient conditions, and the subjectivity of readers, the manual reading method has always lacked a certain degree of stability. On the other hand, manual reading is also time-consuming. As a disease with relatively urgent time requirements, ischemic stroke requires patients to race against time for diagnosis and treatment to avoid rapid deterioration of the disease. Therefore, it is of great clinical significance to quickly, accurately, and stably perform ASPECTS scoring on patients. Summary of the invention
[0006] Based on this, it is necessary to provide an automated ASPECTS scoring method based on plain scan CT to address the above technical issues.
[0007] This application is based on the automated ASPECTS scoring method for plain scan CT, including:
[0008] Obtain an image sequence based on plain scan CT, screen the image sequence, and obtain a number of key frames, wherein the number of key frames include a boundary frame where the lateral ventricles are connected, and the method for obtaining the number of key frames includes: selecting sample images based on the image sequence of plain scan CT, performing deep learning of binary classification on the sample images using a deep learning model, obtaining a trained screening model, and screening a number of key frames of the output image sequence using the trained screening model;
[0009] Positioning the falx cerebri on the plurality of key frames, wherein the falx cerebri appears as a dividing line on the plurality of key frames;
[0010] Performing image registration on the key frames for locating the dividing line to obtain pairs of partitions for ASPECTS scoring, and performing circumscribed rectangle processing and image interpolation processing on any pair of partitions;
[0011] Using the training data to train the deep learning model to obtain a scoring model, wherein there are multiple scoring models, and for any one of the scoring models, the training data correspondingly includes a pair of partitions of the plurality of key frames, and for any one of the scoring models, a pair of partitions includes a first side partition and a second side partition located at two opposite sides relative to the dividing line, and the training data includes a mirror image of the first side partition relative to the dividing line, and / or a mirror image of the second side partition relative to the dividing line;
[0012] The scoring model is used to perform ASPECTS scoring on a number of key frames to be scored to obtain an ASPECTS scoring result.
[0013] Optionally, a method of obtaining a plurality of key frames by obtaining an image sequence based on plain scan CT includes:
[0014] Obtaining an image sequence based on plain scan CT, and obtaining an upright image sequence based on the image sequence based on plain scan CT;
[0015] Calculating the bounding rectangle and intercepting the bounding rectangle in sequence according to the upright image sequence to obtain the intercepted image sequence;
[0016] Select a sample image from the captured image sequence, use a deep learning model to perform binary classification on the sample image, and obtain a trained screening model;
[0017] The trained screening model is used to judge the intercepted image sequence and output a number of key frames of the image sequence, wherein the key frames include a boundary frame connected to the lateral ventricles.
[0018] Optionally, locating the falx cerebri on the plurality of key frames specifically includes:
[0019] Ellipses are intercepted based on the several key frames to obtain several elliptical key frames, and the falx cerebri is located on the several elliptical key frames.
[0020] Optionally, locating the falx cerebri on the several elliptical key frames specifically includes: binarizing the several elliptical key frames to obtain corresponding binary images, adjusting the HU value of the binary image to a first threshold value so that a dividing line of the binary image is clearly displayed relative to the entire binary image, and the dividing line is the located falx cerebri.
[0021] Optionally, when performing image registration on the plurality of key frames for locating the dividing line, the method further includes: rotating the plurality of elliptical key frames according to the dividing line to make the dividing line vertical, thereby completing fine tilt correction.
[0022] Optionally, the two categories included in the binary classification are a first category of images belonging to the upper ganglion layer and a second category of images not belonging to the upper ganglion layer.
[0023] Optionally, for any scoring model, the training data includes a first set of data and a second set of data; the first set of data includes a first comparison image and a second side partition; the second set of data includes a second comparison image and a first side partition;
[0024] The first contrast image is obtained by performing image registration and difference processing on the mirror image of the first side partition relative to the dividing line and the second side partition;
[0025] The second contrast image is obtained by performing image registration and difference processing on the mirror image of the second side partition relative to the boundary line and the first side partition.
[0026] The present application also provides a computer device, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the automated ASPECTS scoring method based on plain scan CT described in the present application.
[0027] The present application also provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the steps of the automated ASPECTS scoring method based on plain scan CT described in the present application are implemented.
[0028] The present application also provides a computer program product, comprising computer instructions, which, when executed by a processor, implement the steps of the automated ASPECTS scoring method based on plain scan CT described in the present application.
[0029] The automated ASPECTS scoring method based on plain scan CT in this application has at least the following effects:
[0030] The present application locates the dividing line on several key frames, performs image registration after the location, and after the registration, each pair of partitions located on the two opposite sides of the dividing line is used as training data for the scoring model. Since the image information on the two opposite sides of the dividing line has an auxiliary judgment effect on the judgment of cerebral infarction, the training data including each pair of partitions can improve the accuracy of the scoring model, which is more applicable to cases with asymmetric brain shapes. BRIEF DESCRIPTION OF THE DRAWINGS
[0031] Figure 1 This is a flow chart of an automated ASPECTS scoring method based on plain scan CT in one embodiment of the present application;
[0032] Figure 2 A schematic diagram of a circumscribed rectangle of an erect image sequence in an embodiment of the present application;
[0033] Figure 3 Figure 2 Schematic diagram of the image after cutting off 70% of its length and width;
[0034] Figure 4 Schematic diagram of an image sequence based on plain scan CT in an embodiment of the present application;
[0035] Figure 5 This is a schematic diagram of the residual structure of the ResNet50 model in one embodiment of the present application;
[0036] Figure 6 A schematic diagram of converting an image sequence captured by a screening model into an array in one embodiment of the present application;
[0037] Figure 7 A schematic diagram of a key frame before intercepting an ellipse in an embodiment of the present application;
[0038] Figure 8 A schematic diagram of an ellipse key frame after intercepting an ellipse in one embodiment of the present application
[0039] Fig. 9 for Figure 7 Binary image of ;
[0040] Fig.10 for Figure 8 Binary image of ;
[0041] Fig.11A schematic diagram of a fine tilt correction process in an embodiment of the present application;
[0042] Fig.12 This is a schematic diagram of the ratio of a pair of partitions to the entire image in one embodiment of the present application;
[0043] Fig.13 The partition located on one side of the dividing line after the ROI is intercepted and the size is uniformly changed in one embodiment of the present application;
[0044] Fig.14 The partition on the other side of the dividing line after the ROI is intercepted and the size is uniformly changed in one embodiment of the present application;
[0045] Fig.15 A schematic diagram of one of the comparison images included in the training data in one embodiment of the present application;
[0046] Fig.16 This is a schematic diagram of the structure of a scoring model in one embodiment of the present application;
[0047] Fig.17 This is a schematic diagram of the process of training a scoring model in one embodiment of the present application;
[0048] Fig.18 A schematic diagram of a process of scoring model prediction in one embodiment of the present application;
[0049] Fig.19 1 is a diagram of the internal structure of a computer device in one embodiment of the present application. DETAILED DESCRIPTION
[0050] In current clinical applications, the ASPECTS scoring method mainly relies on clinicians to manually read the images and make judgments and assessments on each area. On the one hand, due to differences in imaging equipment, differences in patient conditions, and subjectivity of readers, manual reading lacks stability. On the other hand, manual reading is also time-consuming, so it is of great clinical significance to quickly, accurately, and stably perform ASPECTS scoring on patients.
[0051] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.
[0052] To solve the above technical problems, see Figure 1 The present application provides an automated ASPECTS scoring method based on plain scan CT, comprising the following steps S100 to S500, wherein:
[0053] Step S100, obtaining an image sequence based on plain scan CT, screening the image sequence, obtaining a number of key frames, the number of key frames including a boundary frame connected to the lateral ventricle. In one embodiment, step S100 includes steps S110 to S140. Among them:
[0054] Step S110, obtaining an image sequence based on plain scan CT, and obtaining an upright image sequence based on the image sequence based on plain scan CT.
[0055] This step includes skull removal, translation and preliminary tilt correction of the image sequence based on plain scan CT. Among them: (1) Skull removal includes: using the binary method and the maximum connected domain method to remove the skull part, for example, using open source image processing libraries such as OpenCV and skimage to remove it. (2) Translation includes: obtaining the image centroid and translating the intracranial image to the center of the image. (3) Preliminary tilt correction includes: rotating the image with the centroid as the origin between -80° and +80° with a step size of 1 degree, calculating the number of pixels overlapping on both sides of each angle value after horizontal mirroring and the longest axis size of the image in the vertical direction, and assigning weights α and β to the two respectively. When the sum of α and β is the largest, confirm the rotation angle required for alignment, and rotate each image to this angle to obtain an upright image sequence. The weight α can be, for example, 1, and the weight β can be, for example, 75.
[0056] Step S120 , calculating the bounding rectangle and intercepting the bounding rectangle in sequence according to the upright image sequence to obtain the intercepted image sequence.
[0057] See also Figure 2 and Figure 3 In this step, the method of intercepting the circumscribed rectangle includes intercepting the length and width of the circumscribed rectangle respectively, excluding the area not containing the head, so that the centroid of the intercepted image sequence is in the center of each image. Since the shape of each person's head may be very different, in order to exclude the influence of the head shape on the computer's automatic layer selection algorithm, the length and width of the above rectangle are intercepted by 70% respectively. That is, the interception is completed by the following formula:
[0058]
[0059] in,
[0060] x1' is the minimum x-axis coordinate in the intercepted image;
[0061] x2' is the maximum value of the x-axis coordinate in the intercepted image;
[0062] x1 is the minimum x-axis coordinate in the image;
[0063] x2 is the maximum value of the x-axis coordinate in the image;
[0064] y1' is the minimum value of the y-axis coordinate in the intercepted image;
[0065] y2' is the maximum value of the y-axis coordinate in the intercepted image;
[0066] y1 is the minimum value of the y-axis coordinate in the image;
[0067] y2 is the maximum value of the y-axis coordinate in the image;
[0068] It can be understood that the captured image does not include areas outside the skull, and the HU values (hounsfield unit) of each coordinate (x1, x2, y1, and y2) in the image sequence are not zero.
[0069] In one embodiment, the automatic screening of key frames is completed through step S130 and step S140 .
[0070] Step S130, selecting a sample image from the captured image sequence, and performing binary classification deep learning on the sample image using a deep learning model to obtain a trained screening model;
[0071] It is known from common knowledge of intracranial imaging that the basal ganglia layer and the superior ganglia layer are close in position. In this embodiment, whether the upper and lower lateral ventricles are connected is used as the boundary between the basal ganglia layer and the superior ganglia layer (expressed as a boundary frame in the CT image sequence), and the sample image at least includes the boundary frame where the lateral ventricles are connected.
[0072] See also Figure 4 If the key frames selected by this method are used for ASPECTS scoring, three images can be taken as images of the basal ganglia layer and the superior ganglia layer, and a total of six images can be taken as the above-mentioned key frames. Among them, the first image connected to the lateral ventricle in the sequence (the boundary frame with sequence number 14) is used as the first image of the superior ganglia layer; the next two connected images are used as the second and third images of the superior ganglia layer (images with sequence numbers 15 and 16); the upper three connected images are used as three images of the basal ganglia layer (images with sequence numbers 11, 12, and 13). In this embodiment, the sample images include images of the basal ganglia layer and images of the superior ganglia layer, specifically: Figure 4 The six images are numbered 11 to 16.
[0073] In step S130, during the training phase, for example, a ResNet50 model with pre-trained weights can be used. The so-called ResNet50 model with pre-trained weights is to apply the weight parameters of the ResNet50 model pre-trained on the ImageNet dataset to the training task of this deep learning model through transfer learning. The cross entropy loss function, Adam optimizer, batch size of 64, learning rate of 2e-4, and a learning rate decay strategy of reducing the learning rate by 10% if the network performance does not improve for every 100 batches are used for model training. After 50 epochs, the model accuracy can reach up to 99.95% for the training set and 92.7% for the validation set. Regarding the meaning of the relevant parameters involved in the training process, please refer to the prior art, and the applicant will not elaborate on it.
[0074] In step S130, the two categories included in the binary classification deep learning are the first category of images belonging to the upper ganglion layer and the second category of images not belonging to the upper ganglion layer.
[0075] Specifically, since the three images of the basal ganglia layer are connected to the three images of the superior ganglia layer in the plain scan CT sequence, and the characteristics of the lateral ventricle of the superior ganglia layer are more obvious, there is no need to set the deep learning task as a three-classification task of the basal ganglia layer key frames, the superior ganglia layer key frames, and the remaining image frames. It can be directly set as a two-classification task of the superior ganglia layer key frames and the remaining image frames.
[0076] Step S140, using the trained screening model to judge the captured image sequence and output a number of key frames of the image sequence, wherein the key frames include a boundary frame connected to the lateral ventricles.
[0077] Step S140 specifically includes steps S141 to S142, wherein:
[0078] Step S141, see Figure 6 , using the trained screening model to convert the intercepted image sequence into an array, the array including the first type of image marked with the first symbol and the second type of image marked with the second symbol;
[0079] Specifically, the captured image sequence obtained in step S130 is input into the trained screening model, and the trained screening model outputs a 1000-dimensional vector. This 1000-dimensional vector is connected to the ReLu activation function and then to the fully connected layer to output a two-dimensional vector. By comparing the size relationship between the two values of the two-dimensional vector, it is determined whether the frame belongs to the upper ganglion layer.
[0080] It can be understood that the binary classification deep learning is used in step S130, so the output result in step S140 corresponds to the above binary classification. That is, the first type of image belongs to the upper ganglion layer, and the second type of image does not belong to the upper ganglion layer. In this embodiment, the first type of image and the second type of image are marked as "1" and "0" respectively, that is, the trained screening model predicts each frame of the intercepted image sequence and outputs two marking results of "1" and "0", and each marking result forms an array as a whole, and the array includes n elements, and n corresponds to the number of outputs of the trained screening model.
[0081] Step S142: the area where the first symbol appears most times continuously is the target area. The trained screening model determines the first frame of the image sequence of the target area as the boundary frame, and obtains a number of key frames according to the boundary frame and the image frames connected to the boundary frame.
[0082] In step S142, due to the presence of a small number of false detections in the output of the trained screening model, such as mislabeling the second symbol "0" as the first symbol "1" (false positive). This embodiment uses the method of finding the target area with the most consecutive "1" occurrences in the array (i.e., the section with the most consecutive "1" occurrences), and using the target area as the image area of the superior ganglion layer, which can eliminate the impact of a small number of false detections. Furthermore, the first frame of the target area is determined as the boundary frame between the basal ganglia layer and the superior ganglia layer. The key frames specifically include basal ganglia layer key frames and superior ganglia layer key frames, specifically including boundary frames and image frames connected to them before and after. The basal ganglia layer key frames and the superior ganglia layer key frames can be, for example, three frames each.
[0083] The screening process of the trained screening model can be found in Figure 4 , the trained screening model determines the boundary frame as Figure 4 The image with sequence number 14 in the figure. According to the boundary frame, the key frames are determined as three frames of images of the upper ganglion layer and three frames of images of the basal ganglion layer. Among them, the three frames of images of the upper ganglion layer include images with sequence numbers 14, 15, and 16, and the three frames of images of the basal ganglion layer include Figure 4 Images with sequence numbers 11, 12, and 13.
[0084] In steps S110 to 140 of the present application, a 30%-reduced brain image circumscribed rectangle is used as the ROI area as the model input. This innovation can not only reduce the interference of skull shape features and improve model accuracy, but also reduce the size of the original image input and speed up model training and prediction. A two-classification model is used instead of a three-classification model for the selection of the basal ganglia layer and the superior ganglia layer in the ASPECTS scoring task. According to the layer selection target of this task, it is only necessary to find the boundary frame between the two to improve the accuracy of the layer selection.
[0085] Steps S110 to 140 of the present application: realize automatic layer selection, high layer selection accuracy, and fast layer selection speed, and can eliminate subjective interference of the scorer, so as to facilitate rapid and accurate automatic scoring of patients. It avoids layer selection errors, can greatly improve the diagnosis speed of patients, and has high clinical application value.
[0086] Step S200, locating the falx cerebri on a number of key frames, where the falx cerebri appears as a dividing line on a number of key frames.
[0087] Since the shape of the brain contour is different in each case, a fine tilt correction method is needed to avoid the influence of image tilt on subsequent image registration and infarction prediction, which is referred to as "fine tilt correction" in each embodiment. In intracranial plain scan CT, the HU value of the falx cerebri is significantly higher than that of brain tissue, and the image shows a line from the front to the back of the brain (such as Figure 7 The boundary line between the left and right hemispheres shown in FIG. 1 is also clearer in the upper ganglion layer, and the positioning of the falx cerebri is shown as the positioning boundary line in the image sequence. The sub-steps S210 to S230 of step S200 are described in detail.
[0088] Step S210, intercepting ellipses based on a number of key frames to obtain a number of elliptical key frames, locating the falx cerebri on the several elliptical key frames, and the falx cerebri represents the boundary line of the several key frames.
[0089] See also Figure 7 and Figure 8 The method of obtaining several elliptical key frames includes: calculating the brain image centroids of several key frames and the horizontal and vertical brain image lengths (pixel values) passing through the centroids. Based on these values, draw an elliptical template with a long axis and a short axis of 2 / 3 of the vertical and horizontal length values respectively.
[0090] It can be understood that all pixels outside the elliptical template of several elliptical key frames are discarded. For example, the discarding method can be to set all HU values of pixels in the area to 0 to discard pixels with higher HU values in the outer circle of the image to avoid interfering with the search for the boundary line as much as possible.
[0091] Step S220, binarization is performed on a plurality of elliptical key frames to obtain a corresponding binary image, and the HU value of the binary image is adjusted to a first threshold value so that the boundary line of the binary image is clearly displayed relative to the entire binary image.
[0092] See also Figure 7 to Figure 10In this step, several ellipse key frames are binarized and the binarized HU (Hounsfield Unit) values are determined. Specifically, the HU value is adjusted from 20 to 50 and traversed in a step of 1 to obtain a first threshold. Clear display can be further understood as the number of pixels in the ellipse template exceeding the first threshold accounting for no more than 1.5% of the total number of pixels.
[0093] Specifically, the traversal is stopped when the number of pixels above the HU threshold drops below 1.5% of the total number of pixels, and this is used as the first threshold for binarization and then the dilation operation. Fig. 9 and Fig.10 , Fig. 9 for Figure 7 The binary image of Fig.10 for Fig. 9 Binary graph of . Fig.10 Relative to Fig. 9 , its boundary line can be shown more clearly compared to the whole.
[0094] Step S230, rotating a plurality of elliptical key frames according to the dividing line to make the dividing line vertical, thereby completing fine tilt correction.
[0095] When the image is displayed in the default state, it has a horizontal direction, and the vertical direction is perpendicular to the horizontal direction. Fig.10 and Fig.11 , on the binary black and white image, draw a vertical line (i.e. Fig.11 A line passing through the centroid in the vertical direction (in the vertical direction) and having a length equal to the length of the major axis of the ellipse template is rotated with the centroid as the origin between the angle values of -70° and +70° in steps of 1 degree, such as Fig.11 As shown. The number of white pixel values for each angle value is recorded. The angle value with the largest number of white pixel values is the tilt angle value of the brain image. Reverse rotation of the angle completes the fine tilt correction. In this step, for example, the upper ganglion layer key frame can be used for fine tilt correction. It can be understood that the lines of the upper ganglion layer key frame are clearer. It can be understood that the fine tilt correction in this step can improve the convenience of image registration in step S300.
[0096] In step S210 to step S230, the characteristics of the difference between the HU values of the falx cerebri and the brain tissue in the intracranial plain scan CT are extracted to perform tilt correction, which is more robust than the existing algorithm and more suitable for cases with asymmetric brain shapes.
[0097] Step S300, performing image registration on a plurality of key frames of the positioning boundary line to obtain pairs of partitions for ASPECTS scoring.
[0098] A pair of partitions is a scoring area for ASPECTS scoring. Since the shapes of each of the ten scoring areas are quite different, and the infarction manifestations and noise characteristics are also inconsistent, it is necessary to train a model for each partition separately. In this step, the ASPECTS partition templates of ten partitions of the left and right brain are prepared, and the nonlinear image registration algorithm is used to register the partition templates to six key frames in a plain scan CT sequence of a case through affine transformation and B-spline transformation.
[0099] Step S400, using the training data to train the deep learning model to obtain a scoring model, there are multiple scoring models, and for any scoring model, the training data correspondingly includes a pair of partitions of a plurality of key frames.
[0100] See also Figure 12 to Figure 16 In step S400, "the corresponding training data includes" means that a pair of partitions are used as a set of training data to train a scoring model. Since the image information of the two opposite sides has an auxiliary judgment effect on the judgment of cerebral infarction, the model input adds the mirror image of the two opposite side partitions relative to the dividing line.
[0101] like Fig.12 As shown in the figure, since the size of each pair of partitions is very small compared to the entire image, it cannot be directly used as the input of the deep learning model. For any pair of partitions, it is necessary to perform bounding rectangle processing and image interpolation processing. Before the training starts, it also includes cutting out the image region of interest (ROI) and changing the image size of the region of interest.
[0102] Further, for any scoring model, a pair of partitions includes a first side partition and a second side partition located on opposite sides of the dividing line, and the training data includes a mirror image of the first side partition relative to the dividing line, and / or a mirror image of the second side partition relative to the dividing line.
[0103] In one embodiment, for any scoring model, the training data includes a first set of data and a second set of data; the first set of data includes a mirror image of the first side partition relative to the dividing line, and the second side partition; the second set of data includes a mirror image of the second side partition relative to the dividing line, and the first side partition.
[0104] In one embodiment, see Fig.15 and Fig.16For any scoring model, the training data includes a first set of data and a second set of data; the first set of data includes a first contrast image and a second side partition; the second set of data includes a second contrast image and the first side partition; the first contrast image is obtained by aligning and performing difference processing on the mirror image of the first side partition relative to the dividing line and the second side partition; the second contrast image is obtained by aligning and performing difference processing on the mirror image of the second side partition relative to the dividing line and the first side partition.
[0105] The image registration method is used to make the mirror image of one side partition relative to the dividing line closer to the image shape of the other side partition. While increasing the amount of information input to the model, it reduces the noise input and improves the reliability of the scoring model.
[0106] Step S500: Use the scoring model to perform ASPECTS scoring on a number of key frames to be scored to obtain an ASPECTS scoring result.
[0107] See also Fig.17 and Fig.18 In the scoring model training stage, the deep learning model is trained using the training data, and the scoring model is obtained after the training is completed. In the scoring model prediction stage, several key frames after registration are predicted, and the zonal infarction judgment is performed to obtain the scores of ten scoring areas as the scoring results. It can be understood that the data format processed during the scoring model prediction is the same as the data format used during its training. The scoring model prediction targets several key frames after the zonal image registration is completed, which will not be repeated here.
[0108] In step S200 to step S500: First, six images for image registration are obtained, and the six images include three images of the basal ganglia layer and three images of the superior ganglia layer. Based on the six images, ten pairs of partitions are registered and output, and each pair of partitions includes three frames on each side of the left and right brain, a total of 60 partition images, and each image only contains a specific partition area part of the original image, and the grayscale value of the rest is 0. It can be understood that among the pairs of partitions used for scoring, seven pairs of partitions are on the three images of the basal ganglia layer, and three pairs of partitions are on the three images of the superior ganglia layer.
[0109] Two 3D images are generated from the three images of the left and right brains, respectively. Each 3D image also uses the above-mentioned circumscribed rectangle method to intercept the image ROI area in the width and height dimensions. Since the sizes of the intercepted images are different, it is necessary to change the image size to (3,128,128) through image interpolation processing. The image interpolation processing can be, for example, 4×4 pixel neighborhood bicubic interpolation. The results after intercepting the ROI and uniformly changing the size are shown in the figure below. Fig.13 and Fig.14The problem of inconsistent and irregular shapes of ASPECTS partition images input to the convolutional neural network can be solved by cutting off the minimum bounding rectangle of the three-dimensional image area (i.e. the above-mentioned 3D image) and re-changing it to a uniform size.
[0110] Secondly, obtain the contrast images on the two opposite sides of the dividing line. The model input of a certain scoring area on one side not only includes the processed image of that side of the area, but also includes the contrast image of the difference between that side and the opposite side of the area. Fig.15 As shown in FIG. 1 , since the shapes of the two sides of the same partition of the same case may not be similar, the mirror image of the contralateral image based on the dividing line is nonlinearly transformed by image registration, so that the shape of the mirror image of the contralateral image is as close as possible to the image of the side. The above-mentioned contrast image can be obtained by performing a difference process between the image of the side and the mirror image of the contralateral image.
[0111] Finally, the scoring model is designed. The image on one side and its contrast image are used as two sets of inputs, respectively connected to two sets of 3DCNN modules (the modules must contain BN structures), and two sets of feature maps are output. The two sets of feature maps are connected in the channel dimension, and after being connected to another set of CNN modules, the judgment of whether there is an ischemic stroke lesion in the area is output through global pooling, fully connected layers, and Sigmoid functions. The model diagram is as follows Fig.16 shown.
[0112] During the scoring model training phase, contrast images can be obtained from both the left and right sides of a certain area of a case. Both sides can be used as data sets, that is, the amount of training set data for each area is twice the number of training set cases.
[0113] In the scoring model prediction stage, ten trained models are used to predict the infarct areas of the left and right brain for each case, and the ASPECTS scores of the left and right brain are output.
[0114] It should be understood that although Figure 1 The steps in the flowchart are shown in sequence as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified in this document, there is no strict order restriction for the execution of these steps, and these steps can be executed in other orders. Moreover, Figure 1 At least part of the steps may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily executed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least part of the sub-steps or stages of other steps.
[0115] In one embodiment, a computer device is provided. The computer device may be a terminal, and its internal structure diagram may be as follows: Fig.19 As shown. The computer device includes a processor, a memory, a network interface, a display screen and an input device connected through a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The network interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, an automated ASPECTS scoring method based on plain scan CT is implemented. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer device can be a touch layer covered on the display screen, or a key, trackball or touchpad set on the computer device housing, or an external keyboard, touchpad or mouse, etc.
[0116] In one embodiment, a computer device is provided, including a memory and a processor, wherein a computer program is stored in the memory, and when the processor executes the computer program, the following steps are implemented:
[0117] Step S100, obtaining an image sequence based on plain scan CT, screening the image sequence, and obtaining a number of key frames, wherein the key frames include a boundary frame connected to the lateral ventricles;
[0118] Step S200, locating the falx cerebri on a number of key frames, where the falx cerebri appears as a dividing line on the number of key frames;
[0119] Step S300, performing image registration on a plurality of key frames of the positioning dividing line to obtain each pair of partitions for ASPECTS scoring;
[0120] Step S400, training the deep learning model using the training data to obtain a scoring model, where there are multiple scoring models, and for any scoring model, the training data correspondingly includes a pair of partitions of a plurality of key frames;
[0121] Step S500: Use the scoring model to perform ASPECTS scoring on a number of key frames to be scored to obtain an ASPECTS scoring result.
[0122] In one embodiment, a computer readable storage medium is provided, on which a computer program is stored, and when the computer program is executed by a processor, the following steps are implemented:
[0123] Step S100, obtaining an image sequence based on plain scan CT, screening the image sequence, and obtaining a number of key frames, wherein the key frames include a boundary frame connected to the lateral ventricles;
[0124] Step S200, locating the falx cerebri on a number of key frames, where the falx cerebri appears as a dividing line on the number of key frames;
[0125] Step S300, performing image registration on a plurality of key frames of the positioning dividing line to obtain each pair of partitions for ASPECTS scoring;
[0126] Step S400, training the deep learning model using the training data to obtain a scoring model, where there are multiple scoring models, and for any scoring model, the training data correspondingly includes a pair of partitions of a plurality of key frames;
[0127] Step S500: Use the scoring model to perform ASPECTS scoring on a number of key frames to be scored to obtain an ASPECTS scoring result.
[0128] In one embodiment, a computer program product is provided, comprising computer instructions, which when executed by a processor implement the following steps:
[0129] Step S100, obtaining an image sequence based on plain scan CT, screening the image sequence, and obtaining a number of key frames, wherein the key frames include a boundary frame connected to the lateral ventricles;
[0130] Step S200, locating the falx cerebri on a number of key frames, where the falx cerebri appears as a dividing line on the number of key frames;
[0131] Step S300, performing image registration on a plurality of key frames of the positioning dividing line to obtain each pair of partitions for ASPECTS scoring;
[0132] Step S400, training the deep learning model using the training data to obtain a scoring model, where there are multiple scoring models, and for any scoring model, the training data correspondingly includes a pair of partitions of a plurality of key frames;
[0133] Step S500: Use the scoring model to perform ASPECTS scoring on a number of key frames to be scored to obtain an ASPECTS scoring result.
[0134] In this embodiment, the computer program product includes a program code portion for executing the steps of the automated ASPECTS scoring method based on plain scan CT in each embodiment of the present application when the computer program product is executed by one or more computing devices. The computer program product can be stored on a computer-readable recording medium. The computer program product can also be provided via a data network (e.g., via a RAN, via the Internet and / or via an RBS) for downloading. Alternatively or additionally, the method can be encoded in a field programmable gate array (FPGA) and / or an application specific integrated circuit (ASIC), or the functionality can be provided for downloading with the aid of a hardware description language.
[0135] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).
[0136] The technical features of the above embodiments can be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification. When the technical features in different embodiments are embodied in the same figure, it can be regarded that the figure also discloses the combination examples of the various embodiments involved.
[0137] The above-mentioned embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent of the present application shall be subject to the attached claims.
Claims
1. Automated ASPECTS scoring method based on plain scan CT, It is characterized in that include: Obtain an image sequence based on plain scan CT, screen the image sequence, and obtain a number of key frames, wherein the number of key frames include a boundary frame where the lateral ventricles are connected, and the method for obtaining the number of key frames includes: selecting sample images based on the image sequence of plain scan CT, performing deep learning of binary classification on the sample images using a deep learning model, obtaining a trained screening model, and screening a number of key frames of the output image sequence using the trained screening model; Positioning the falx cerebri on the plurality of key frames, wherein the falx cerebri appears as a dividing line on the plurality of key frames; Performing image registration on the key frames for locating the dividing line to obtain pairs of partitions for ASPECTS scoring, and performing circumscribed rectangle processing and image interpolation processing on any pair of partitions; Using the training data to train the deep learning model to obtain a scoring model, wherein there are multiple scoring models, and for any one of the scoring models, the training data correspondingly includes a pair of partitions of the plurality of key frames, and for any one of the scoring models, a pair of partitions includes a first side partition and a second side partition located at two opposite sides relative to the dividing line, and the training data includes a mirror image of the first side partition relative to the dividing line, and / or a mirror image of the second side partition relative to the dividing line; The scoring model is used to perform ASPECTS scoring on a number of key frames to be scored to obtain an ASPECTS scoring result.
2. The automated ASPECTS scoring method according to claim 1, It is characterized in that The methods of obtaining several key frames specifically include: Obtaining an image sequence based on plain scan CT, and obtaining an upright image sequence based on the image sequence based on plain scan CT; Calculating the bounding rectangle and intercepting the bounding rectangle in sequence according to the upright image sequence to obtain the intercepted image sequence; Select a sample image from the captured image sequence, use a deep learning model to perform binary classification on the sample image, and obtain a trained screening model; The trained screening model is used to judge the intercepted image sequence and output a number of key frames of the image sequence, wherein the key frames include a boundary frame connected to the lateral ventricles.
3. The automated ASPECTS scoring method according to claim 1, It is characterized in that Positioning the falx cerebri on the plurality of key frames specifically includes: Ellipses are intercepted based on the several key frames to obtain several elliptical key frames, and the falx cerebri is located on the several elliptical key frames.
4. The automated ASPECTS scoring method according to claim 3, It is characterized in that Positioning the falx cerebri on the plurality of elliptical key frames specifically includes: The plurality of elliptical key frames are binarized to obtain a corresponding binary image, and the HU value of the binary image is adjusted to a first threshold value so that the boundary line of the binary image is clearly displayed relative to the entire binary image, and the boundary line is located, and the boundary line is the located falx cerebri.
5. The automated ASPECTS scoring method according to claim 3, It is characterized in that When performing image registration on the plurality of key frames for positioning the dividing line, the method further comprises: The plurality of elliptical key frames are rotated according to the dividing line to make the dividing line vertical, thereby completing fine tilt correction.
6. The automated ASPECTS scoring method according to claim 2, It is characterized in that The two categories included in the binary classification are a first category of images belonging to the upper ganglion layer and a second category of images not belonging to the upper ganglion layer.
7. The automated ASPECTS scoring method according to claim 1, It is characterized in that For any scoring model, the training data includes a first set of data and a second set of data; the first set of data includes a first comparison image and a second side partition; the second set of data includes a second comparison image and a first side partition; The first contrast image is obtained by performing image registration and difference processing on the mirror image of the first side partition relative to the dividing line and the second side partition; The second contrast image is obtained by performing image registration and difference processing on the mirror image of the second side partition relative to the boundary line and the first side partition.
8. Computer equipment, including memory, processor and computer programs stored in the memory, It is characterized in that The processor executes the computer program to implement the steps of the automated ASPECTS scoring method based on plain scan CT according to any one of claims 1 to 7.
9. A computer-readable storage medium having a computer program stored thereon, It is characterized in that When the computer program is executed by a processor, the steps of the automated ASPECTS scoring method based on plain scan CT according to any one of claims 1 to 7 are implemented.
10. Computer program products, including computer instructions, It is characterized in that When the computer instructions are executed by a processor, the steps of the automated ASPECTS scoring method based on plain scan CT according to any one of claims 1 to 7 are implemented.
