Aortic dissection detection method and system

Abstract

The invention discloses an aortic dissection detection method and system, and belongs to the technical field of medical image processing, and the method comprises the steps: segmenting an aortic region in each layer of slice of a CTA image through employing a first segmentation network, and obtaining a slice image of the aortic region in each layer of slice; further segmenting an inner valve region in the aorta region slice image by adopting a second segmentation network; an inner valve area is accurately detected step by step in a step-by-step segmentation mode; in the process, after the slice image of the aorta area in each layer of slice is obtained, slice images with wrong division in the formed slice sequence are filtered through a sliding window and local continuity, and guarantee is provided for further accurate segmentation of the inner valve area in the second stage; according to the method, the aortic dissection can be accurately detected by accurately segmenting the inner valve area in the CTA image.

Description

Technical Field

[0001] The present invention belongs to the technical field of medical image processing, and more specifically, relates to an aortic dissection detection method and system. Background Art

[0002] Aortic dissection (AD) is a life-threatening aortic disease. The aorta consists of three layers: the inner, middle and outer membranes, and is divided into three parts: the ascending aorta, the aortic arch and the descending aorta. Rupture of the inner membrane layer causes blood to seep into the middle layer, resulting in aortic dissection. Soon after the onset of symptoms, the mortality rate of untreated patients with acute aortic dissection increases by 1%–2% per hour. Manual detection is prone to miss some subtle lesions and is affected by factors such as experience and fatigue. Therefore, rapid and accurate intelligent detection methods are crucial to shorten the preoperative time.

[0003] Computed tomography angiography (CTA) is an effective means of detecting aortic dissection. The intimal valve is the most significant imaging feature of aortic dissection. The area and number of the intimal valve can be used to determine the type of aortic dissection under the DeBakey classification. However, due to imaging features such as calcification, artifacts, and low contrast, the intimal valve in the aorta cannot be accurately segmented, and thus aortic dissection cannot be accurately detected. Summary of the invention

[0004] In view of the above defects or improvement needs of the prior art, the present invention provides an aortic dissection detection method and system to solve the technical problem that the prior art cannot accurately detect aortic dissection.

[0005] In order to achieve the above object, in a first aspect, the present invention provides a method for detecting aortic dissection, comprising:

[0006] The first segmentation network is used to segment the CTA image to be detected, and the slice images of the aorta region in each layer of slices stacked in the CTA image and the corresponding aorta category are obtained; the slice images of each aorta region are sorted according to the order of the corresponding slice layers in the CTA image to be detected along the stacking direction to form a slice sequence;

[0007] A sliding window is used to slide on a slice sequence to divide the slice sequence into slice subsequences of different aortic categories, and slice images that do not belong to the aortic category in each slice subsequence are removed and reassembled to obtain a new slice sequence; wherein, when the sliding window is at different positions, the number of slice images of different aortic categories contained in the sliding window is counted respectively, and when the aortic category begins to change when the number exceeds a preset number threshold, the changed aortic category is used as the current category, and the first slice image of the current category in the current sliding window is separated from the previous slice image to segment the slice sequence;

[0008] The second segmentation network is used to segment the inner valve region in each slice image in the new slice sequence, and the number of slice images containing the inner valve region obtained by segmentation under different aortic categories is counted respectively, so as to obtain the DeBakey classification of aortic dissection in the CTA image to be detected.

[0009] Further preferably, the aorta categories include: aortic arch, ascending aorta and descending aorta;

[0010] Count the number of slice images N containing the inner valve area obtained by segmentation under the categories of aortic arch, ascending aorta and descending aorta respectively a 、N u and N d , and then calculate the ascending and descending aortic dissection scores;

[0011] Among them, the ascending and descending aortic dissection scores S u , S d They are:

[0012]

[0013]

[0014] Th is the preset threshold.

[0015] Further preferably, when the ascending and descending aortic dissection scores are both less than or equal to a preset score, the CTA image to be detected is determined to be normal;

[0016] When the ascending and descending aortic dissection scores are both greater than the preset scores, the aortic dissection of the CTA image to be detected is determined to be type I according to the DeBakey classification;

[0017] When the ascending aortic dissection score is greater than or equal to the preset score, and the descending aortic dissection score is less than or equal to the preset score, the aortic dissection of the CTA image to be detected is determined to be DeBakey type II dissection;

[0018] When the ascending aortic dissection score is less than a preset score and the descending aortic dissection score is greater than a preset score, the aortic dissection of the CTA image to be detected is determined to be a DeBakey type III dissection.

[0019] Further preferably, the first segmentation network comprises: a cascaded encoding module, a dilation module and a decoding module;

[0020] The encoding module includes M cascaded residual modules; M≥2; the input end of each residual module is also connected to its input end through a residual;

[0021] The dilation module consists of one or more cascaded atrous convolutional layers.

[0022] Further preferably, the first segmentation network is a pre-trained segmentation network, and its training process includes:

[0023] Inputting the first training set into the first segmentation network, and training the first segmentation network by minimizing the first loss function;

[0024] The first training set includes: CTA images of normal people and CTA images of patients with different types of aortic dissection according to DeBakey classification; wherein the labels carried by each slice of the CTA image include: the true segmentation result of the aorta region in the slice image and the corresponding aorta category;

[0025] The first loss function includes: the difference loss between the aorta region segmentation result of each slice in the CTA image in the first training set and the corresponding true segmentation result.

[0026] Further preferably, the second segmentation network includes: a feature extraction module, a semantic extractor, a spatial feature extraction module, an attention module, a feature fusion module and a segmentation head;

[0027] The feature extraction module is used to extract the features of the slice image step by step, and obtain N feature images f1, f2, ..., f with decreasing scales from large to small. N ; N ≥ 3;

[0028] The semantic extractor is used to transform f1, f2, ..., f N spliced ​​together, and semantic extraction is performed based on the channel attention mechanism and the spatial attention mechanism to obtain the global semantic encoding information; the global semantic encoding information includes: f1, f2, ..., f N The corresponding semantic encoding information b1, b2, ..., b N ;

[0029] The spatial feature extraction module is used to extract the spatial features f of the slice image;

[0030] The feature fusion module includes N semantic injection units;

[0031] The jth semantic injection unit is used to inject N-j+1 Upsampling is performed to make it consistent with f N-j+1 After the size is the same as f N-j+1 Based on the attention mechanism, the jth decoding feature d is obtained. j ; j = 1, 2, ..., N-2;

[0032] The N-1th semantic injection unit is used to upsample b2 to make it the same size as f2, and then fuse it with the features extracted by the attention module of f2 based on the attention mechanism to obtain the N-1th decoding feature d N-1 ;

[0033] The Nth semantic injection unit is used to upsample b1 to make it the same size as f, and then fuse it with f based on the attention mechanism to obtain the Nth decoding feature d N ;

[0034] The splitting head is used based on d1, d2, ..., d N The spliced ​​features are used to segment the inner valve region in the slice image.

[0035] Further preferably, the semantic extractor comprises: a cascaded channel attention module and a spatial attention module.

[0036] Further preferably, the attention module comprises: one or more cascaded channel attention modules.

[0037] Further preferably, the spatial feature extraction module comprises: a plurality of cascaded CNN modules and a spatial attention module connected to the output end of the last stage CNN module.

[0038] Further preferably, the second segmentation network is a pre-trained segmentation network, and its training process includes:

[0039] Inputting the second training set into the second segmentation network, and training the second segmentation network by minimizing the second loss function;

[0040] The second training set includes: slice images of aorta regions of different aorta categories; the labels carried by each slice image include: the true segmentation result and the true boundary line image of the inner valve region in the slice image;

[0041] During the training process, the second segmentation network further includes a boundary line detection module, which is used to extract the boundary line of the inner valve region in the slice image based on f2 to obtain a boundary line image of the inner valve region in the slice image;

[0042] The second loss function includes: the difference loss between the inner valve area segmentation result of the slice image in the second training set obtained by the second segmentation network and the corresponding true segmentation result, and the difference loss between the boundary line image obtained by the boundary line detection module and the corresponding true boundary line image.

[0043] Further preferably, the second training set further includes: slice images with shadows at the edge of the inner valve region.

[0044] In a second aspect, the present invention provides an aortic dissection detection system, comprising: a memory and a processor, wherein the memory stores a computer program, and the processor executes the method provided in the first aspect of the present invention when executing the computer program.

[0045] In a third aspect, the present invention further provides a computer-readable storage medium, which includes a stored computer program, wherein when the computer program is executed by a processor, the device where the storage medium is located is controlled to execute the method provided in the first aspect of the present invention.

[0046] In general, the above technical solutions conceived by the present invention can achieve the following beneficial effects:

[0047] 1. The present invention provides a method for detecting aortic dissection, which uses a first segmentation network to segment the aortic region in each layer of slices of a CTA image to obtain slice images of the aortic region in each layer of slices, and then uses a second segmentation network to further segment the inner valve region in the slice images of the aortic region; through a step-by-step segmentation method, the inner valve region is accurately detected step by step; and in this process, after the slice images of the aortic region in each layer of slices are obtained, the slice sequences constituted are also filtered through sliding windows and local continuity to remove incorrectly divided slice images in the slice sequence, thereby providing a guarantee for further accurate segmentation of the inner valve region in the second stage; the present invention can accurately detect aortic dissection by accurately segmenting the inner valve region in the CTA image.

[0048] 2. Furthermore, in the aortic dissection detection method provided by the present invention, the first segmentation network includes a cascaded encoding module, a dilation module and a decoding module; the encoding module includes a plurality of cascaded residual modules; the input end of each residual module is also connected to its input end through a residual; the dilation module includes one or more cascaded hole convolution layers, thereby effectively expanding the network's receptive field, improving the ability to capture global information, and being able to more accurately locate the aortic area.

[0049] 3. Furthermore, in the aortic dissection detection method provided by the present invention, the second segmentation network includes: a feature extraction module, a semantic extractor, a spatial feature extraction module, an attention module, a feature fusion module and a segmentation head; while ensuring the second segmentation network's ability to extract and summarize the semantic features of the inner valve, the spatial information expression capability of the second segmentation network is enriched, and the inner valve area can be located more accurately.

[0050] 4. Furthermore, in the aortic dissection detection method provided by the present invention, during the training process of the second segmentation network, the second segmentation network also includes a boundary line detection module, which extracts the boundary line of the inner valve area in the slice image based on f2, thereby realizing early guidance of shallow features, and constructing a difference loss between the boundary line image extracted by the boundary line detection module and the corresponding real boundary line image to supervise the spatial features, which can better take into account the spatial and semantic information, so that the second segmentation network can more accurately locate the inner valve area. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 A flow chart of a method for detecting aortic dissection provided by an embodiment of the present invention;

[0052] Figure 2 A schematic diagram of an aortic dissection detection process provided by an embodiment of the present invention;

[0053] Figure 3 A schematic diagram of the structure of a first segmentation network provided in an embodiment of the present invention;

[0054] Figure 4 A schematic diagram of the structure of a second segmentation network provided by an embodiment of the present invention;

[0055] Figure 5 A schematic diagram of an aorta region slice image and an inner valve region segmentation result of the aorta region slice image provided by an embodiment of the present invention;

[0056] Figure 6 A schematic diagram of the structure of a semantic extractor LSSE provided in an embodiment of the present invention;

[0057] Figure 7 A schematic diagram of the structure of a spatial feature extraction module LSPath provided in an embodiment of the present invention;

[0058] Figure 8 A schematic diagram of a process of acquiring a boundary line image and a real boundary line image provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0059] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention 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 invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0060] In order to achieve the above objectives, firstly, Figure 1 As shown, the present invention provides a method for detecting aortic dissection, comprising:

[0061] 1) The first stage of segmentation:

[0062] The first segmentation network is used to segment the CTA image to be detected, and the slice images of the aorta region in each layer of slices stacked in the CTA image and the corresponding aorta category are obtained; the slice images of each aorta region are sorted according to the order of the corresponding slice layers in the CTA image to be detected along the stacking direction to form a slice sequence; wherein the aorta categories include: aortic arch, ascending aorta and descending aorta. One slice of the CTA image corresponds to one scanning segment.

[0063] In an optional implementation manner, the specific detection process is as follows Figure 2 As shown, in order to provide more detailed advice for surgical operations, the diameter of the aorta in the slice image of the aorta region is also measured. Specifically, the aortic diameter is measured based on the aortic segmentation results. For the ascending aorta, the average diameter from the aortic root to the aortic arch is calculated; for the descending aorta, the portion from the aortic arch to the abdominal aorta is detected. In order to verify the accuracy of the diameter measurement, the doctor can manually measure the aortic diameter of the selected vascular segment every five scan slices.

[0064] Preferably, in an optional embodiment, the CTA image to be detected is preprocessed before being input into the first segmentation network. The preprocessing operation may include one or more operations such as denoising and image enhancement. Specifically, in an optional embodiment, the CTA image to be detected is denoised using the SRAD algorithm that reduces speckle anisotropic diffusion, and the resolution of the CTA image to be detected is adjusted to 256×256. In an optional embodiment, considering that different doses of contrast enhancement agents will be used in patient examinations, the brightness of the CTA image to be detected is adjusted to half of the original brightness for data enhancement, and the image is rotated 90° counterclockwise.

[0065] It should be noted that the first segmentation network can be a segmentation network such as SegNet, Mask-RCNN, DeepLab, etc. Preferably, in an optional implementation, the first segmentation network includes: a cascaded encoding module, an expansion module, a decoding module and a mapping layer;

[0066] The encoding module includes M cascaded residual modules; M≥2; the input end of each residual module is also connected to its input end through a residual; preferably, each residual module includes multiple cascaded convolutional layers;

[0067] The dilation module includes one or more cascaded dilated convolutional layers;

[0068] The mapping layer maps the decoded feature map of the slice output by the decoding module to the aorta region pixel by pixel to obtain the aorta region segmentation result of the slice. Generate segmentation results. This function calculates the probability of each pixel in the decoded feature map being the aortic region or the non-aortic region. When the probability of the aortic region is large, the pixel is divided into the aortic region. Otherwise, the pixel is divided into the non-aortic region.

[0069] In one embodiment, a first segmentation network having 21 hidden layers is provided, and its structural diagram is shown as follows: Figure 3 As shown; in this embodiment, the introduction of the residual block can effectively suppress the gradient vanishing phenomenon, and the addition of the hole convolution at the bottleneck of the first CNN can improve the global information capture capability of the convolutional neural network so as to accurately locate the aorta; the use of dropout and rectified linear unit (ReLU) can improve the convergence speed.

[0070] In an optional implementation manner, the first segmentation network is a pre-trained segmentation network, and its training process includes:

[0071] Inputting the first training set into the first segmentation network, and training the first segmentation network by minimizing the first loss function;

[0072] The first training set includes: CTA images of normal people and CTA images of patients with different types of aortic dissection according to DeBakey classification; wherein the labels carried by each slice of the CTA image include: the true segmentation result of the aorta region in the slice image and the corresponding aorta category;

[0073] The first loss function includes: the difference loss between the aorta region segmentation result of each slice in the CTA image in the first training set and the corresponding true segmentation result.

[0074] It should be noted that the loss function for measuring the difference loss between the aorta region segmentation result of each slice in the CTA image of the first training set and the corresponding true segmentation result can be multiple, such as Dice loss function, cross entropy loss, Focal loss function, Jaccard loss function, etc. Among them, the Dice loss function is:

[0075]

[0076] Among them, X and Y represent the prediction result and the true label respectively, and the calculation process does not consider the background pixels.

[0077] Furthermore, in an optional embodiment, an experienced doctor uses the medical image processing software 3DSlicer to perform pixel-level annotation on the aorta region in each slice of the CTA image to obtain the corresponding label. Before the CTA images in the first training set are input into the first segmentation network, they are also preprocessed in the same way as the preprocessing method performed on the CTA images to be detected.

[0078] 2) Filter the wrongly divided slice images in the slice sequence by sliding window and local continuity:

[0079] A sliding window is used to slide on a slice sequence to segment the slice sequence into slice subsequences of different aortic categories, and slice images that do not belong to the aortic category in each slice subsequence are removed and reassembled to obtain a new slice sequence; wherein, when the sliding window is at different positions, the number of slice images of different aortic categories contained in the sliding window is counted respectively, and when the aortic category begins to change when the number exceeds a preset number threshold (in an optional implementation manner, the value is 10), the changed aortic category is used as the current category, and the first slice image of the current category in the current sliding window is segmented from the previous slice image to segment the slice sequence.

[0080] In an optional implementation, the length of the sliding window is 5 image slices, and the step length is 2 image slices. Specifically, a sliding window of 5 slices is used to detect the junction of different scanning segments of the aorta, and the entire CTA volume is divided into three segments: the aortic arch, the descending aorta, and the ascending and descending aorta; each slice is classified into one of the three segments according to the number and shape of the aortic regions detected in the slice; once the category of the slice does not match the scanning segment to which it belongs, the slice will be discarded.

[0081] In an optional implementation, the slice images of each aorta region are sorted in order according to the position of the corresponding slice layer in the CTA image to be detected along the stacking direction to form a slice sequence. Taking the stacking direction as the direction from top to bottom (i.e., the scanning direction of the CTA image) as an example, at this time, the aorta categories of the first and last image slices in the slice sequence are the aortic arch and the descending aorta, respectively.

[0082] When the sliding window is at different positions, the number of image slices of different aorta categories contained in the sliding window is obtained respectively; in the initial sliding stage, the number of image slices of the aortic arch category exceeds a preset threshold; after multiple sliding, when the number of image slices of the ascending aorta category exceeds the preset threshold, the first image slice of the ascending aorta category in the current sliding window is separated from the previous image slice; when the number of image slices of the descending aorta category exceeds the preset threshold, the first image slice of the descending aorta category in the current sliding window is separated from the previous image slice.

[0083] 3) Second stage segmentation:

[0084] The second segmentation network is used to segment the inner valve region in each slice image in the new slice sequence, and the number of slice images containing the inner valve region obtained by segmentation under different aortic categories is counted respectively, so as to obtain the DeBakey classification of aortic dissection in the CTA image to be detected.

[0085] In an optional implementation, the number of slice images N containing the inner valve region obtained by segmentation in the aortic arch, ascending aorta and descending aorta categories is counted respectively. a 、N u and N d , and then calculate the ascending and descending aortic dissection scores;

[0086] Among them, the ascending and descending aortic dissection scores S u , S d They are:

[0087]

[0088]

[0089] Th is a preset threshold value, and in an optional implementation manner, the value is 10.

[0090] In an optional implementation, when the ascending and descending aortic dissection scores are both less than or equal to a preset score, the CTA image to be detected is determined to be normal;

[0091] When the ascending and descending aortic dissection scores are both greater than the preset scores, the aortic dissection of the CTA image to be detected is determined to be type I according to the DeBakey classification;

[0092] When the ascending aortic dissection score is greater than or equal to the preset score, and the descending aortic dissection score is less than or equal to the preset score, the aortic dissection of the CTA image to be detected is determined to be DeBakey type II dissection;

[0093] When the ascending aortic dissection score is less than a preset score and the descending aortic dissection score is greater than a preset score, the aortic dissection of the CTA image to be detected is determined to be a DeBakey type III dissection.

[0094] Furthermore, surgical recommendations can be given based on the classification results. For patients with type I dissection, ascending aorta / aortic arch replacement and antegrade TEVAR are recommended. For patients with type II dissection, ascending aorta / aortic arch replacement is recommended, while for patients with type III dissection, only TEVAR is recommended, which can effectively improve the diagnostic efficiency of aortic dissection.

[0095] It should be noted that the second segmentation network can be a segmentation network such as SegNet, Mask-RCNN, DeepLab, etc.

[0096] Preferably, in an optional embodiment, as Figure 4 As shown, the second segmentation network includes: a feature extraction module, a semantic extractor LSSE, a spatial feature extraction module LSPath, an attention module, a feature fusion module and a segmentation head;

[0097] The feature extraction module includes N cascaded feature extraction units; the feature extraction module is used to extract the features of the slice image step by step to obtain N feature images f1, f2, ..., f with decreasing scales from large to small. N ; N ≥ 3; Preferably, N is 4;

[0098] The semantic extractor LSSE is used to transform f1, f2, …, f N spliced ​​together, and semantic extraction is performed based on the channel attention mechanism and the spatial attention mechanism to obtain the global semantic encoding information; the global semantic encoding information includes: f1, f2, ..., f N The corresponding semantic encoding information b1, b2, ..., b N ;

[0099] The spatial feature extraction module LSPath is used to extract the spatial features f of the slice image;

[0100] The feature fusion module includes N semantic injection units;

[0101] The jth semantic injection unit is used to inject N-j+1 Upsampling is performed to make it consistent with f N-j+1 After the size is the same as fN-j+1 Based on the attention mechanism, the jth decoding feature d is obtained. j ; j = 1, 2, ..., N-2;

[0102] The N-1th semantic injection unit is used to upsample b2 to make it the same size as f2, and then fuse it with the features extracted by the attention module of f2 based on the attention mechanism to obtain the N-1th decoding feature d N-1 ;

[0103] The Nth semantic injection unit is used to upsample b1 to make it the same size as f, and then fuse it with f based on the attention mechanism to obtain the Nth decoding feature d N ;

[0104] The splitting head is used based on d1, d2, ..., d N The spliced ​​features are used to segment the inner valve region in the slice image.

[0105] like Figure 5 Shown are the aorta region slice image and the inner valve region segmentation results of the aorta region slice image.

[0106] In an optional implementation, the semantic injection unit includes an upsampling unit and a SIM module; the upsampling unit is used to perform an upsampling operation; and the SIM module is used to implement a fusion operation based on an attention mechanism.

[0107] In an optional implementation, the semantic extractor LSSE includes: a cascaded channel attention module and a spatial attention module. Preferably, Figure 6 As shown, in an optional implementation, the semantic extractor LSSE includes a multi-layer cascaded channel attention module and a multi-layer cascaded spatial attention module; wherein the output end of the last-level channel attention module is connected to the input end of the first-level spatial attention module.

[0108] In one optional implementation, the attention module includes: one or more cascaded channel attention modules. In another optional implementation, the attention module includes: one or more cascaded spatial attention modules.

[0109] In an optional implementation, the spatial feature extraction module LSPath includes: multiple cascaded CNN modules and a spatial attention module connected to the output end of the last-level CNN module; wherein the multiple cascaded CNN modules are used to extract low-dimensional detailed spatial features of the slice image.

[0110] Introducing additional network layers to obtain more feature information will inevitably increase the overall computational complexity of the network and reduce the speed of network reasoning. Taking the above issues into consideration, this implementation method designs a lightweight spatial feature extraction module LSPath to obtain additional spatial information. Its structure is as follows: Figure 7 shown.

[0111] Generally speaking, small convolution kernels focus on details, while large convolution kernels focus on contextual information. Considering that the purpose of the spatial feature extraction module LSPath is to extract spatial information, 3×3 convolution layers are used in the first three layers of the spatial feature extraction module LSPath to process features. In addition, a large number of channels often represents multi-dimensional abstract features, and a low number of channels corresponds to richer detail information. Therefore, the number of channels in the first three layers of the spatial feature extraction module LSPath are designed to be 16, 32, and 32, respectively. Small convolution kernels and low number of channels also reduce the burden on network reasoning caused by the introduction of the spatial feature extraction module LSPath. And convolution with a step size of 2 is used in the first two layers of convolution to achieve fast downsampling of feature maps, further reducing the amount of network calculations. In this implementation, the last layer of convolution of LSPath is designed as a 1×1 convolution kernel to fuse the channel information of the first three layers of features, and at the same time upgrade the 32-channel features to 64 channels to match the input requirements of the subsequent feature fusion module. And two cascaded spatial refinement modules are connected after the convolution to enhance the spatial feature representation of the path, and the final size is the original Figure 1 / 4 feature map is used to replace the features output by the first-level feature extraction unit (stage1) in the original feature extraction backbone network as the input of the feature fusion module.

[0112] In an optional implementation manner, the second segmentation network is a pre-trained segmentation network, and its training process includes:

[0113] Inputting the second training set into the second segmentation network, and training the second segmentation network by minimizing the second loss function;

[0114] The second training set includes: slice images of aorta regions of different aorta categories; the labels carried by each slice image include: the true segmentation result and the true boundary line image of the inner valve region in the slice image;

[0115] During the training process, the second segmentation network further includes a boundary line detection module FlapHead, which is used to extract the boundary line of the inner valve region in the slice image based on f2 to obtain a boundary line image of the inner valve region in the slice image;

[0116] The second loss function includes: the difference loss between the inner valve area segmentation result of the slice image in the second training set obtained by the second segmentation network and the corresponding true segmentation result, and the difference loss between the boundary line image obtained by the boundary line detection module and the corresponding true boundary line image.

[0117] It should be noted that the aortic region slice images in the second training set can be obtained by direct artificial clipping, or by extracting the aortic region in each layer of the CTA image slices through the first segmentation network of the present invention, which is not limited here.

[0118] It should be noted that the boundary line detection module FlapHead is a neural network, which can be CNN, YOLO, SSD, etc. Preferably, in an optional implementation, the boundary line detection module FlapHead includes a cascaded first convolution layer, a BN layer, an activation layer (preferably using ReLu) and a second convolution layer; in this implementation, the second convolution layer is a 1×1 convolution layer.

[0119] In an optional implementation, the second training set further includes: slice images with shadows at the edge of the inner valve region.

[0120] In an optional implementation, when obtaining the labels carried by the slice image, the true segmentation result of the inner valve area in the slice image is represented by a mask, which can be obtained by manual annotation or by performing preliminary inner valve detection on each slice image based on the Hessian matrix (HM) method. The supervision information of the inner valve demarcation (the true boundary line image) is obtained by the mask of the inner valve area; Figure 8 As shown, when extracting the inner valve boundary, two types of labels are included, namely, inner valve and non-inner valve elements. The contour information is obtained by moving the Laplace operator on the mask annotation with a variety of different moving steps to obtain edge contour maps of different scales. After the edge contour maps of different scales are transformed into images of the same size as the mask, they are averaged to obtain an average edge contour map. The pixels in the average edge contour map with values ​​greater than the preset contour point threshold are divided into the boundary area to obtain the inner valve contour map. The inner valve contour map is expanded to obtain the real boundary line image. Specifically, the moving step sizes are set to 1, 2, and 4, respectively, so that the size is consistent with the original image, and the size is the original Figure 1 / 2 and 1 / 4 edge contour images to obtain contour information of different scales. Figure 1 The edge contour images of size 1 / 2 and 1 / 4 are upsampled to the same size as the original image, and the average of the three contour images is calculated. The pixels with the average value greater than the preset contour point threshold (0.1) are recorded as contour points (points in the boundary area), and the inner valve contour image is obtained. Considering that the feature map itself only has the original Figure 1 / 8 resolution, and it is too demanding to directly obtain accurate contours from shallow features. Therefore, this paper uses the expansion operation to thicken the inner valve contour map to obtain its soft contour map, so as to avoid over-fitting of the network caused by too strong constraints. The formula for the expansion operation is:

[0121]

[0122] In the above formula, R d is the size of the rectangular dilation operator, Represents the inner valve contour, g e It is the contour image after dilation, which is used as the real boundary line image.

[0123] Considering that the number of boundary pixels is far less than that of non-boundary pixels, boundary prediction is a class-imbalanced binary classification problem. Therefore, in this embodiment, binary cross entropy and Dice loss are used to jointly optimize boundary learning. The expression of this joint loss is:

[0124] L Flap (p e , g e )=L bce (p e , g e )+L dice (p e , g e )

[0125] Among them, p e , g e They represent the boundary line image obtained by the boundary line detection module FlapHead and the real boundary line image respectively. bce and L dice Respectively represent binary cross entropy loss and Dice loss, and their expressions are:

[0126]

[0127]

[0128] Where H and W represent the height and width of the boundary line image obtained by the boundary line detection module, respectively.

[0129] Binary cross entropy loss is used for binary classification problems. For inner valve prediction, the inner valve region image can be divided into two types of elements: foreground and background, corresponding to two types of label values ​​0 or 1. The elements in the inner valve region image are probability values ​​between 0 and 1. Binary cross entropy loss measures the performance of the model by comparing the probability distribution of the network output with the distribution of the actual label. Since the predicted label is evaluated using the log function, it is more sensitive when the predicted value deviates greatly from the true value, which can speed up the iteration of the network.

[0130] However, the binary cross entropy loss usually calculates the loss of each pixel independently without considering their spatial connection, and finally takes the average of the loss of each pixel, which will make it difficult to take into account errors in details. In addition, this feature also reduces its performance in class imbalance problems such as inner valve boundary prediction, because the network tends to predict pixels as categories with a larger number, resulting in vague results for other categories. From this perspective, additional losses need to be added to alleviate this problem. Compared with the binary cross entropy loss, the Dice loss measures the network accuracy by measuring the overlap between the boundary line image obtained by the boundary line detection module FlapHead and the true boundary line image. In addition, the Dice loss only considers foreground pixels during the calculation process, which has a better effect on class imbalance problems such as boundary line extraction and evaluation. Therefore, the embodiment combines it with the binary cross entropy loss to make up for its shortcomings and further improve the effect of the network.

[0131] In addition, since the output results of the boundary line detection module FlapHead will not be re-input into the backbone network, they are only used in the training phase to guide the direction of feature extraction of the backbone network, and will not increase the overhead of network inference.

[0132] Two branches are added to the original network structure. The starting point of the LSPath branch of the feature extraction module is the input slice image, which directly obtains high-resolution spatial information from the slice image and outputs it to the subsequent feature fusion module. The FlapHead branch of the boundary line detection module takes the 1 / 8 size feature map output by the second-level feature extraction unit (stage2) in the feature extraction trunk as the starting point, and outputs the result to the loss function calculation module to guide shallow feature extraction.

[0133] The final loss function of the network is as follows, L flap The network loss calculated by the feature output of the FlapHead guide head of the boundary detection module, with a weight of λ c .L seg Represents the network loss calculated from the inner valve region segmentation result and some semantic features.

[0134] L SEWS =λ c L flap +L seg

[0135] L seg The expression is:

[0136] L seg =L f +λ s L sep

[0137] It is caused by the focal loss L f and feature separation loss L sep Composition, where λ s is the weight of feature separation loss.

[0138] Focus loss L f The expression is:

[0139]

[0140] in, Represents the number of predicted categories, which is equivalent to the number of channels output by the network at the end. Represents the probability that the current element belongs to the cth category calculated by the network; with the help of the product term Focus loss L f Reduce the loss calculated from high-confidence prediction results, so that the network pays more attention to difficult-to-recognize examples. ζ is the adjustment factor that controls the tilt of the scheme.

[0141] Feature separation loss L sep Different features are separated early by calculating the feature differences of elements belonging to different categories. The expression is shown in the following formula:

[0142]

[0143] in, In the expression and Represents the number of elements of different categories, N M represents the number of channels of the feature map, and Represents the eigenvalues ​​of different categories of elements on the nth channel of feature map m, μ n is the mean of the channel characteristic values. With this formula, The difference between the features in different categories of elements in the feature map is calculated. sep Based on the formula The feature difference between the 1 / 16 feature map obtained by the feature extraction backbone at the corresponding positions of the two types of labels 0 and 1 is calculated and used as part of the loss function to encourage the network to distinguish them in the early stage.

[0144] It should be noted that the above processes are all executed in a computer or other program execution module.

[0145] In a second aspect, the present invention provides an aortic dissection detection system, comprising: a memory and a processor, wherein the memory stores a computer program, and the processor executes the method provided in the first aspect of the present invention when executing the computer program.

[0146] The related technical solution is the same as the method provided in the first aspect of the present invention and will not be described in detail here.

[0147] In a third aspect, the present invention further provides a computer-readable storage medium, which includes a stored computer program, wherein when the computer program is executed by a processor, the device where the storage medium is located is controlled to execute the method provided in the first aspect of the present invention.

[0148] The related technical solution is the same as the method provided in the first aspect of the present invention and will not be described in detail here.

[0149] It will be easily understood by those skilled in the art that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for detecting aortic dissection, characterized in that: include: The first segmentation network is used to segment the CTA image to be detected, and the slice images of the aorta region in each layer of slices stacked in the CTA image and the corresponding aorta category are obtained; Slice images of each aorta region are sorted according to the order of the corresponding slice layers in the CTA image to be detected along the stacking direction to form a slice sequence; A sliding window is used to slide on the slice sequence to segment the slice sequence into slice subsequences of different aorta categories, and slice images that do not belong to the aorta category in each of the slice subsequences are removed and reassembled to obtain a new slice sequence; wherein, when the sliding window is at different positions, the number of slice images of different aorta categories contained in the sliding window is counted respectively, and when the aorta category begins to change when the number exceeds a preset number threshold, the changed aorta category is used as the current category, and the first slice image of the current category in the current sliding window is segmented from the previous slice image to segment the slice sequence; The second segmentation network is used to segment the inner valve region in each slice image in the new slice sequence, and the number of slice images containing the inner valve region obtained by segmentation under different aortic categories is counted respectively, so as to obtain the DeBakey classification of aortic dissection in the CTA image to be detected.

2. The aortic dissection detection method according to claim 1, characterized in that: The aorta categories include: aortic arch, ascending aorta, and descending aorta; Count the number of slice images N containing the inner valve area obtained by segmentation under the categories of aortic arch, ascending aorta and descending aorta respectively a 、N u and N d , and then calculate the ascending and descending aortic dissection scores; Among them, the ascending and descending aortic dissection scores S u , S d They are: Th is the preset threshold.

3. The aortic dissection detection method according to claim 2, characterized in that: When the ascending and descending aortic dissection scores are both less than or equal to the preset scores, the CTA image to be tested is judged to be normal; When the ascending and descending aortic dissection scores are both greater than the preset scores, the aortic dissection of the CTA image to be detected is determined to be type I according to the DeBakey classification; When the ascending aortic dissection score is greater than or equal to the preset score, and the descending aortic dissection score is less than or equal to the preset score, the aortic dissection of the CTA image to be detected is determined to be DeBakey type II dissection; When the ascending aortic dissection score is less than a preset score and the descending aortic dissection score is greater than a preset score, the aortic dissection of the CTA image to be detected is determined to be a DeBakey type III dissection.

4. The aortic dissection detection method according to claim 1, characterized in that: The first segmentation network includes: a cascaded encoding module, a dilation module and a decoding module; The encoding module includes M cascaded residual modules; M≥2; the input end of each residual module is also connected to its input end through a residual; the expansion module includes one or more cascaded hole convolution layers.

5. The aortic dissection detection method according to claim 1, characterized in that: The first segmentation network is a pre-trained segmentation network, and its training process includes: Inputting a first training set into the first segmentation network, and training the first segmentation network by minimizing a first loss function; The first training set includes: CTA images of normal people and CTA images of patients with different types of aortic dissection according to DeBakey classification; wherein the labels carried by each slice of the CTA image include: the true segmentation result of the aorta region in the slice image and the corresponding aorta category; The first loss function includes: the difference loss between the aorta region segmentation result of each layer slice in the CTA image in the first training set and the corresponding true segmentation result.

6. The method for detecting aortic dissection according to any one of claims 1 to 5, characterized in that: The second segmentation network includes: a feature extraction module, a semantic extractor, a spatial feature extraction module, an attention module, a feature fusion module and a segmentation head; The feature extraction module is used to extract the features of the slice image step by step, and obtain N feature images f1, f2, ..., f N ; N ≥ 3; The semantic extractor is used to transform f1,f2,…,f N spliced ​​together, and semantic extraction is performed based on the channel attention mechanism and the spatial attention mechanism to obtain global semantic encoding information; the global semantic encoding information includes: f1, f2, ..., f N The corresponding semantic encoding information b1, b2, ..., b N ; The spatial feature extraction module is used to extract the spatial feature f of the slice image; The feature fusion module includes N semantic injection units; The jth semantic injection unit is used to inject N-j+1 Upsampling is performed to make it consistent with f N-j+1 After the size is the same as f N-j+1 Based on the attention mechanism, the jth decoding feature d is obtained. j ; j = 1, 2, ..., N-2; The N-1th semantic injection unit is used to upsample b2 to make it the same size as f2, and then fuse it with the features extracted by the attention module of f2 based on the attention mechanism to obtain the N-1th decoding feature d N-1 ; The Nth semantic injection unit is used to upsample b1 to make it the same size as f, and then fuse it with f based on the attention mechanism to obtain the Nth decoding feature d N ; The segmentation head is used based on d1, d2, ..., d N The spliced ​​features are used to segment the inner valve region in the slice image.

7. The aortic dissection detection method according to claim 6, characterized in that: The semantic extractor includes: a cascaded channel attention module and a spatial attention module; The spatial feature extraction module includes: a plurality of cascaded CNN modules and a spatial attention module connected to the output end of the last-stage CNN module.

8. The method for detecting aortic dissection according to any one of claims 1 to 5, characterized in that: The second segmentation network is a pre-trained segmentation network, and its training process includes: Inputting the second training set into the second segmentation network, and training the second segmentation network by minimizing the second loss function; The second training set includes: slice images of aorta regions of different aorta categories; the labels carried by each slice image include: the true segmentation result and the true boundary line image of the inner valve region in the slice image; During the training process, the second segmentation network further includes a boundary line detection module, which is used to extract the boundary line of the inner valve region in the slice image based on f2 to obtain a boundary line image of the inner valve region in the slice image; The second loss function includes: the difference loss between the inner valve area segmentation result of the slice image in the second training set obtained by the second segmentation network and the corresponding true segmentation result, and the difference loss between the boundary line image obtained by the boundary line detection module and the corresponding true boundary line image.

9. An aortic dissection detection system, characterized in that: include: A memory and a processor, wherein the memory stores a computer program, and the processor executes the method according to any one of claims 1 to 8 when executing the computer program.

10. A computer-readable storage medium, characterized in that The computer-readable storage medium includes a stored computer program, wherein when the computer program is executed by a processor, the device where the storage medium is located is controlled to execute the method according to any one of claims 1 to 8.

Images