Non-enhanced CT image segmentation method based on cascade attention

Through the cascaded attention convolution network and the equalization joint loss function, the low contrast and fuzzy boundary problems of vascular segmentation in non-enhanced CT images are solved, and high-precision automatic vascular segmentation is achieved, which improves the accuracy and stability of segmentation. Support the development of personalized medical care.

CN119006483APending Publication Date: 2024-11-22THE FIRST MEDICAL CENT CHINESE PLA GENERAL HOSPITAL +1
View PDF 4 Cites 0 Cited by

Patent Information

Application Number
CN202411052119.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-01
Publication Date
2024-11-22

AI Technical Summary

Technical Problem

The prior art is difficult to effectively handle vascular segmentation tasks in non-enhanced CT images, especially due to segmentation inaccuracy caused by low contrast characteristics and blurred boundaries, and traditional methods increase patient radiation exposure and economic costs.

Method used

Using a non-enhanced CT image segmentation method based on cascade attention, the high-precision automatic segmentation of blood vessels is achieved through the cascade attention convolution network and the equalization joint loss function, combined with preprocessing and postprocessing strategies, and the non-enhanced CT is overcome. Limitations of images.

Benefits of technology

It achieves more accurate and safe vascular structure analysis without adding additional medical burden, improves the accuracy and stability of tissue segmentation on non-enhanced CT images, and supports the development of personalized medical care.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119006483A_ABST
    Figure CN119006483A_ABST
Patent Text Reader

Abstract

The invention discloses a non-enhanced CT image segmentation method based on cascade attention. The method comprises the following steps: acquiring a non-enhanced CT image and a corresponding enhanced CT image; obtaining a normalized non-enhanced CT image and a registered actual enhanced CT image; constructing a sample and an actual label mask; constructing a cascade attention convolutional network; constructing an equilibrium joint loss function; the cascade attention convolution network is trained; and obtaining a target tissue image corresponding to the to-be-processed non-enhanced CT image according to the trained cascaded attention convolutional network. According to the method, more accurate and safer automatic segmentation of the aorta is realized by constructing the cascade attention convolutional network and adopting a more reasonable multi-scale category equilibrium joint loss function.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the field of medical imaging and artificial intelligence technology, and in particular relates to a segmentation method of non-enhanced CT images based on cascade attention. Background Art

[0002] Blood vessels are the core components of systemic circulation. However, non-enhanced CT faces significant challenges in capturing subtle changes in vascular structure, and its low contrast characteristics limit the effectiveness of traditional analysis methods. Although contrast-enhanced CT angiography significantly improves the visualization of vascular structure, it increases the risk of radiation exposure for patients, accompanied by potential allergic reactions to contrast agents, and high economic costs. In view of this, exploring vascular segmentation technology based on non-enhanced CT images is in line with the safety and economic principles of medical practice. This paper proposes a non-enhanced CT image segmentation method based on a cascaded attention mechanism, aiming to overcome the limitations of non-enhanced CT images and achieve high-precision automatic segmentation of blood vessels through the innovative design of a deep learning model. This method not only takes into account the model structure itself and a more reasonable loss function, but also incorporates a strategy for post-processing of segmentation results, striving to provide a more accurate and safe vascular structure analysis tool without adding additional medical burden, so as to support clinical decision-making and promote the development of personalized medicine.

[0003] Recently, medical image segmentation models represented by nnU-Net [Isensee F, et al. Nature Methods. 2021; 18(2): 203-211] and MedNeXt [Roy S, et al. International Conference on Medical Image Computing and Computer-Assisted Intervention, 2023: 405-415.] have been developed. However, most of these advances focus on the processing of high-contrast images and fail to fully address the inherent difficulties of vascular segmentation in non-enhanced CT images, such as fuzzy boundaries and low-contrast features. This makes it impossible for existing methods to be directly applied to the vascular segmentation task of non-enhanced CT images, limiting the scope of application of non-enhanced CT. Summary of the invention

[0004] The purpose of the present invention is to provide a segmentation method for non-enhanced CT images based on cascaded attention in order to address the above-mentioned problems in the prior art.

[0005] The above-mentioned purpose of the present invention is achieved by the following technical means:

[0006] A segmentation method for non-enhanced CT images based on cascaded attention comprises the following steps:

[0007] Step 1: Acquisition of non-enhanced CT images and corresponding enhanced CT images;

[0008] Step 2, resampling and registering the non-enhanced CT image and the corresponding enhanced CT image in turn, respectively, to obtain a registered non-enhanced CT image and a registered actual enhanced CT image, and adjusting and normalizing the window width and window position of the registered non-enhanced CT image in turn, to obtain a normalized non-enhanced CT image;

[0009] Step 3: construct a sample based on the normalized non-enhanced CT image, and construct an actual label mask corresponding to the sample based on the registered actual enhanced CT image;

[0010] Step 4: Build a cascaded attention convolutional network;

[0011] Step 5: Construct a balanced joint loss function;

[0012] Step 6: Train the cascaded attention convolutional network;

[0013] Step 7: After image resampling, window width and window level selection, normalization, and block division of the non-enhanced CT image to be processed, a patch block of the non-enhanced CT image to be processed is obtained, and the patch block of the non-enhanced CT image to be processed is input into the trained cascade attention convolutional network to obtain the corresponding prediction mask, and the prediction masks corresponding to all the non-enhanced CT image patch blocks are recombined to generate a predicted target tissue mask, and the target tissue image is obtained according to the non-enhanced CT image and the predicted target tissue mask.

[0014] As mentioned above, constructing the sample and the corresponding actual label mask in step 3 includes the following steps:

[0015] Step 3.1, delineating the target tissue on the registered actual enhanced CT image to obtain the actual target tissue mask;

[0016] Step 3.2, based on the outlined actual target tissue mask, the registered non-enhanced CT image and the actual target tissue mask are cropped in the same way according to the set ratio range of the target segmentation region voxel-non-target segmentation region voxel, to obtain the non-enhanced CT image patch block and the corresponding actual target segmentation mask image patch block respectively;

[0017] Step 3.3: Expand the non-enhanced CT image patch and the corresponding actual target segmentation mask patch and use them as samples and corresponding actual label masks respectively.

[0018] As mentioned above, the cascaded attention convolutional network in step 4 includes an encoder, a bottleneck layer, a decoder, and a multi-level target output module.

[0019] The encoder includes 1 convolution input layer and multiple levels, each level includes a coding convolution layer and a downsampling convolution layer, there is no channel change between the coding convolution layer and the downsampling convolution layer of the same level, and the number of channels between adjacent levels is doubled; the coding convolution layer of each level receives the output features of the downsampling convolution layer of the previous level, the output features of the coding convolution layer are input to the downsampling convolution layer of the same level, the coding convolution layer of the first level receives the output features of the convolution input layer, and the output features of the downsampling convolution layer of the last level are input to the bottleneck layer.

[0020] As mentioned above, the bottleneck layer includes multiple MedNeXt convolution blocks, and the output features of the bottleneck layer are input to the decoder.

[0021] As mentioned above, the decoder includes multiple levels, each level includes a decoding convolution layer and an upsampling convolution layer, there is no channel change between the decoding convolution layer and the upsampling convolution layer of the same level, the channels between adjacent levels are halved, the output features of the bottleneck layer are input to the upsampling convolution layer of the first level of the decoder, the output features of the decoding convolution layer of the previous level of the decoder are input to the upsampling convolution layer of the next level, the output features of the encoding convolution layer of the fth level of the encoder and the output features of the upsampling convolution layer of the l-f+1th level of the decoder are added and input to the decoding convolution layer of the l-f+1th level of the decoder.

[0022] As mentioned above, the multi-level target output module includes a cascaded attention block and a 3D transposed convolution output layer, and the output features of the decoding convolution layer of each level of the decoder are respectively output as prediction masks of different sizes through the cascaded attention block and the 3D transposed convolution output layer.

[0023] As mentioned above, the balanced joint loss function is based on the following formula:

[0024] L total =w1*L seg +w2*L ban

[0025] Among them, L total To balance the joint loss function, w1 and w2 are weighted values, L seg is the target segmentation loss function, L ban is the category balance loss function;

[0026] The target segmentation loss function L seg Based on the following formula:

[0027] L seg =DSC_loss+CE_loss

[0028]

[0029] Among them, X is the predicted mask, Y is the actual mask, the actual mask Y and the predicted mask X are of the same size, the actual mask Y is obtained by downsampling the actual label mask, and X i is the i-th voxel point of the predicted mask, Y i is the i-th voxel point of the actual mask, i is the voxel point sequence number, N is the total number of voxel points of the sample, |X| and |Y| represent the number of voxel points of the predicted mask X and the actual mask Y respectively, |X∩Y| represents the number of voxel points of the intersection of X and Y, DSC_loss is the segmentation loss, and CE_loss is the cross entropy loss;

[0030] The class balancing loss function L ban Based on the following formula:

[0031] L ban =αL LDAM +βL FL

[0032]

[0033] Among them, L LDAM is the label distribution aware margin loss, L FL is the focusing loss, α and β are weighted values;

[0034]

[0035] in, is the margin for predicting category k, the constant n k To predict the number of voxels of category k, is the margin for predicting category j, the constant n j is the number of voxels of predicted category j; is the unnormalized probability of category k of the i-th voxel point of the predicted mask, where k∈{1,...,T}, T is the total number of categories; is the unnormalized probability of category j for the i-th voxel point of the predicted mask;

[0036]

[0037] in, α k is the weight of category k, r is the focus parameter, p k is the i-th voxel point X of the predicted mask i The probability value of belonging to category k after the Sigmoid function, is one-hot encoding.

[0038] As mentioned above, in step 6, based on the total model loss value L lossTrain the cascaded attention convolutional network,

[0039]

[0040] Where M is the total number of predicted mask sizes, h i is the weighted value of the balanced joint loss function value of the i-th predicted mask size, is the balanced joint loss function corresponding to the prediction mask size of the i-th level of the decoder.

[0041] As described above, after the non-enhanced CT image patch blocks are divided, they have a set degree of overlap. In the step of generating a predicted target tissue mask by recombining the predicted masks corresponding to all the non-enhanced CT image patch blocks, the masks of the overlapping voxel points of each predicted mask are averaged, and the mask with an average mask greater than 0.5 is the corresponding final valid mask of the overlapping voxel points.

[0042] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of the above-mentioned non-enhanced CT image segmentation method.

[0043] Compared with the prior art, the present invention has the following beneficial effects:

[0044] 1. The present invention designs a cascaded attention convolutional network, which enables the cascaded attention convolutional network to more fully learn the characteristics of different tissues from non-enhanced CT images, thereby achieving more accurate tissue segmentation;

[0045] 2. Construct a balanced joint loss function for multi-scale categories, so that the constructed cascaded attention convolutional network can overcome the problem of imbalanced segmentation targets and effectively improve the accuracy and stability of tissue segmentation on non-enhanced CT images;

[0046] 3. In the embodiment, a series of preprocessing, model training and post-processing steps are designed according to the characteristics of the organization to ensure the efficiency and automation level of the entire process. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 It is a schematic diagram of the process of the present invention.

[0048] Figure 2 Schematic diagram of the structure of the cascaded attention convolutional network. DETAILED DESCRIPTION

[0049] In order to facilitate those skilled in the art to understand and implement the present invention, the present invention is further described in detail below in conjunction with embodiments. The implementation examples described herein are only used to illustrate and explain the present invention and are not intended to limit the present invention.

[0050] Example:

[0051] like Figure 1 and Figure 2 As shown, a segmentation method of non-enhanced CT images based on cascade attention is used to obtain a vascular mask from the non-enhanced CT image segmentation. Taking the aorta segmentation of the non-enhanced CT image as an example, the method includes the following steps:

[0052] Step 1: Acquisition of non-enhanced CT images and corresponding enhanced CT images: Preferably, acquisition of non-enhanced CT images and corresponding enhanced CT images is performed according to the established exclusion rule. The exclusion rule may be based on one or more of the following standards:

[0053] 1) The non-enhanced CT image and the corresponding enhanced CT image were of the aortic section of the same subject, and the scanning interval was no more than 1 week;

[0054] 2) There are no obvious tissue differences between the non-enhanced CT images and the corresponding enhanced CT images except for brightness;

[0055] 3) Non-enhanced CT images and corresponding enhanced CT images are stored in the same way;

[0056] 4) The non-enhanced CT images and the corresponding enhanced CT images do not have serious artifacts, noise or deformation and can meet routine diagnostic requirements;

[0057] Step 2, resampling and registering the non-enhanced CT image and the corresponding enhanced CT image in sequence, respectively, to obtain a registered non-enhanced CT image and a registered actual enhanced CT image with the same spatial resolution and inter-slice correspondence, and further adjusting and normalizing the window width and window position of the registered non-enhanced CT image in sequence, to obtain a normalized non-enhanced CT image;

[0058] Step 2.1, image resampling: the preset cross-sectional resampling resolution is 0.67 mm, and the preset axial resampling resolution is 1.5 mm, that is, the preset resampling resolution is (1.5, 0.67, 0.67), and the non-enhanced CT image and the corresponding enhanced CT image are resampled to the preset resampling resolution;

[0059] Step 2.2, image registration: register the resampled enhanced CT image to the resampled non-enhanced CT image to obtain the registered non-enhanced CT image and the registered actual enhanced CT image, so as to facilitate the subsequent production of the actual target tissue mask of the non-enhanced CT image. This embodiment uses the Elastix registration software Python library PyElastix to perform registration processing and save the registration data without serious deformation;

[0060] Step 2.3, window width and window level selection: According to the modality of the registered non-enhanced CT image, the registered actual enhanced CT image and the target segmented tissue characteristics (in this embodiment, the aorta tissue characteristics), select a suitable window width and window level to highlight the contrast of the target segmented tissue. In this implementation example, the window level is set to 40 and the window width is set to 400, and the registered non-enhanced CT image and the registered actual enhanced CT image are adjusted based on the window width and window level.

[0061] Step 2.4, data normalization: Normalize the registered non-enhanced CT image after adjustment based on the window width and window level to obtain a normalized non-enhanced CT image for model training. There is no need to normalize the registered actual enhanced CT image, because the registered actual enhanced CT image is only used to obtain the actual target tissue mask of the registered non-enhanced CT image.

[0062] Step 3: Construct the sample and the corresponding actual label mask:

[0063] Step 3.1: Production of actual target tissue mask:

[0064] The target tissue is delineated on the registered actual enhanced CT image using the growth threshold method function module of the 3D slicer to obtain an actual target tissue mask. In this embodiment, the target tissue is the aorta lumen, and the actual target tissue mask is the aorta lumen mask.

[0065] Step 3.2: Cut by blocks:

[0066] According to the outlined actual target tissue mask, the registered non-enhanced CT image and the actual target tissue mask are cropped in the same way according to the set ratio range of the target segmentation area voxel-non-target segmentation area voxel. The cropping method is randomly selected to obtain the non-enhanced CT image patch block and the corresponding actual target segmentation mask map patch block, respectively. The target segmentation area voxel-non-target segmentation area voxel in the non-enhanced CT image patch block and the actual target segmentation mask map patch block is the set ratio.

[0067] The set ratio of the target segmentation region voxel to the non-target segmentation region voxel refers to the ratio of the number of voxels in the target segmentation region and the number of voxels in the non-target segmentation region in the image to be cropped (the registered non-enhanced CT image and the actual target tissue mask) to the set ratio range. In this embodiment, random cropping is performed according to the ratio of 1:1 of the target segmentation region voxel to the non-target segmentation region voxel to obtain a non-enhanced CT image patch block with an image depth × image width × image height of 96 × 160 × 160 and an actual target segmentation mask map patch block.

[0068] In this example, the number of sample batches is batch=1, the number of channels is 1, and the input of the module is a preprocessed non-enhanced CT image of size 1×1×96×160×160 (number of sample batches × number of channels × image depth × image width × image height). The multi-level target outputs are one-hot encoded in sizes of 1×1×6×10×10, 1×1×12×20×20, 1×1×24×40×40, 1×1×48×80×80, and 1×1×96×160×160, respectively.

[0069] Step 3.3, data perturbation expansion: Perform 4 times data perturbation expansion on the non-enhanced CT image patch by random rotation, flipping, jittering and scaling, and use the expanded non-enhanced CT image patch and the corresponding actual target segmentation mask map patch as samples and the corresponding actual label mask respectively;

[0070] Step 4, Cascaded Attention Convolutional Network Design: This embodiment uses Python language to build a cascaded attention convolutional network architecture based on the MONAI open source medical imaging deep learning framework and MedNeXt convolutional blocks [Roy S, et al. International Conference on Medical Image Computing and Computer-Assisted Intervention, 2023: 405-415.]. Through encoding processing of non-enhanced CT image samples, deep semantic feature extraction, high-resolution semantic feature reconstruction, and multi-level target mask supervised learning, the aorta is segmented from the non-enhanced CT image.

[0071] The cascaded attention convolution network includes an encoder, a bottleneck layer, a decoder, and a multi-level target output module. The encoder includes a convolution input layer, l (l ≥ 1) encoding convolution layers and l (l ≥ 1) downsampling convolution layers, and the encoding convolution layers and the one-to-one corresponding downsampling convolution layers are located at the same level; the bottleneck layer includes b (b ≥ 1) residual-like modules; the decoder includes l (l ≥ 1) decoding convolution layers and l (l ≥ 1) upsampling convolution layers, and the decoding convolution layers and the one-to-one corresponding upsampling convolution layers are located at the same level; the f-th downsampling convolution layer and the corresponding l-f+1-th upsampling convolution layer are jump-connected; the multi-level target output module includes cascaded attention and convolution layers.

[0072] In this embodiment, the specific structure is as follows:

[0073] (1) Encoder

[0074] The encoder includes 1 convolution input layer, 4 encoding convolution layers and 4 downsampling convolution layers in sequence; the encoding convolution layer includes multiple MedNeXt convolution blocks; the downsampling convolution layer includes a 3D convolution layer with a convolution kernel of 1×1×1 and a convolution step of 2×2×2.

[0075] In this embodiment, the convolution input layer of the encoder converts a single channel into 48 channels, each encoding convolution layer and the corresponding downsampling convolution layer are combined into one level, and there is no channel change between the encoding convolution layer and the downsampling convolution layer of the same level, and the number of channels between adjacent levels is doubled. Specifically, the channel change of the entire encoder is 1->48->96->192->384->768.

[0076] The number of MedNeXt convolution blocks used by the 4-layer encoding convolutional layer of the encoder is 3, 4, 8, and 8 respectively. The channel expansion rates used by the encoding convolutional layer are 2, 3, 4, and 4 respectively. In this embodiment, the inter-layer normalization strategy of the MedNeXt convolutional block adopts Group norm normalization, and the activation function used by the MedNeXt convolutional block is the GELU activation function.

[0077] The samples are input to the convolution input layer of the encoder, and the channels of the samples are expanded. After the samples after channel expansion pass through the encoding convolution layer and downsampling convolution layer of each level, the deep data features are obtained, and the deep data features are sent to the bottleneck layer for further abstract features. For each level of the encoding convolution layer, the output features of the downsampling convolution layer of the previous level are received, and the output features of the encoding convolution layer are input to the downsampling convolution layer of the same level. The encoding convolution layer of the first level receives the output features of the convolution input layer, and the output features of the downsampling convolution layer of the last level are input to the bottleneck layer. Then, the output features of the bottleneck layer are gradually restored to the pixel space through the transposed convolution and decoding convolution layers of the decoder.

[0078] (2) Bottleneck layer

[0079] The bottleneck layer includes multiple MedNeXt convolution blocks. In this example, the number of MedNeXt convolution blocks used in the bottleneck layer is 8, the channel expansion rate of the bottleneck layer is 8, and other parameters (MedNeXt convolution block inter-layer normalization strategy, activation function, convolution kernel size, convolution step) are consistent with the parameters used in the MedNeXt convolution block in the encoder. After the deep data features pass through the bottleneck layer, the output features (S5) of the bottleneck layer are input to the decoder.

[0080] (3) Decoder

[0081] The decoder corresponds to the encoder. The decoder includes 4 decoding convolution layers and 4 upsampling convolution layers. The decoding convolution layers of the encoder and the corresponding upsampling convolution layers are combined into one level. There is no channel change between the decoding convolution layers and upsampling convolution layers of the same level, and the channels between adjacent levels are halved. Specifically, the channel change of the entire encoder is 768->384->192->96->48.

[0082] The number of MedNeXt convolution blocks used in the decoder's decoding convolution layer is 8, 8, 4, and 3 respectively, and other parameter settings (channel expansion rate used in the convolution layer, the above parameters of the MedNeXt convolution block) are consistent with the encoding layer parameter settings of the encoder. The upsampling convolution layer of the decoder is consistent with the downsampling convolution layer parameter settings of the encoder, except that the downsampling convolution layer of the encoder is replaced by the upsampling convolution layer of the transposed convolution.

[0083] The output features of the bottleneck layer are input to the upsampling convolution layer of the first level of the decoder, the output features of the decoding convolution layer of the previous level of the decoder are input to the upsampling convolution layer of the next level, and the output features of the encoding convolution layer of the fth level of the encoder and the output features of the upsampling convolution layer of the l-f+1th level of the decoder are added and input to the decoding convolution layer of the l-f+1th level of the decoder.

[0084] Each level of the decoder performs decryption convolution and transposition convolution on the output features of the bottleneck layer to gradually restore the pixel space. The output features of the decoding convolution layer of each level of the decoder (corresponding to S1-S4 in this embodiment) are respectively input into the multi-level target output module.

[0085] (4) Multi-level target output module

[0086] The multi-level object output module consists of cascaded attention blocks and a 3D transposed convolutional output layer.

[0087] The convolution kernel size of the first layer of the cascaded attention block is 3×3×3, the convolution kernel size of the second layer is 1×1×1, the convolution step size is 1×1×1, the normalization used is BatchNorm3d normalization, and the activation function is the sigmod function. The convolution kernel size and convolution step size of the transposed convolution are both 1×1×1, there is no activation layer, and finally single-channel data is obtained. The output features of the decoding convolutional layers of each level of the decoder are respectively output as prediction masks of different sizes through the cascaded attention block and the 3D transposed convolution output layer. In this embodiment, the side length (width / height) of the output feature S1 of the decoding convolutional layer of the 4th level of the decoder = 2×the side length (width / height) of the output feature S2 of the decoding convolutional layer of the 3rd level of the decoder = 4×the side length (width / height) of the output feature S3 of the decoding convolutional layer of the 2nd level of the decoder = 8×the side length (width / height) of the output feature S4 of the decoding convolutional layer of the 1st level of the decoder = 16×the side length (width / height) of the output feature S5 of the bottleneck layer.

[0088] Step 5: Construct the balanced joint loss function L of the multi-scale category total , balanced joint loss function L total Associated with each level of the decoder, its mathematical expression is:

[0089] L total =w1*L seg +w2*L ban

[0090] Among them, w1=0.5 is the target segmentation loss function L seg The weighted value of w2 = 0.5 is the balanced loss function L for the non-enhanced target category ban Weighted value.

[0091] The target segmentation loss function L seg As shown in the following formula:

[0092] L seg =DSC_loss+CE_loss

[0093]

[0094] Among them, X is the predicted mask, Y is the actual mask, the actual mask Y and the predicted mask X are of the same size, the actual mask Y is obtained by downsampling the actual label mask, and X i is the i-th voxel point of the predicted mask, X i The value is 0 or 1, Y i is the ith voxel point of the actual mask, Y iThe value is 0 or 1, i is the voxel point number, N is the total number of voxel points of the sample, |X| and |Y| represent the number of voxel points of the predicted mask X and the actual mask Y respectively, |X∩Y| represents the number of voxel points of the intersection of X and Y, DSC_loss is the segmentation loss, and CE_loss is the cross entropy loss.

[0095] The class balancing loss function L ban As shown in the following formula:

[0096] L ban =αL LDAM +βL FL

[0097]

[0098] Among them, L LDAM is the label distribution aware margin loss, L FL is the focusing loss, and α=β=0.50 is the weighting value.

[0099]

[0100] in, To predict the margin of category k, is a constant, n k To predict the number of voxels of category k, To predict the margin of category j, is a constant, n j is the number of voxels of predicted category j; is the unnormalized probability of the i-th voxel point of the predicted mask, and the category k is the i-th voxel point Y of the actual mask Y. i The category to which it belongs, where k∈{1,...,T}, T=5 is the total number of categories; accordingly, is the unnormalized probability of the i-th voxel point of the predicted mask, i.e., the i-th voxel point Y of the actual mask. i Category.

[0101]

[0102] in, α k is the weight of category k [1,1,1.5,2,5], r=2 is the focus parameter, p k is the i-th voxel point X of the predicted mask i The probability value of belonging to category k after the Sigmoid function, is a one-hot encoding, and the sample belongs to k On the contrary

[0103] Step 6: The cascaded attention convolutional network training includes the following steps:

[0104] Step 6.1, loss function value calculation: the sample is sent to the constructed cascaded attention convolutional network, and the balanced joint loss function corresponding to the prediction mask size of the i-th level of the decoder is calculated according to the prediction mask output by each level of the decoder

[0105] The balanced joint loss function of each level is weighted and summarized to obtain the total model loss value L loss And pass it to the adaptive optimizer to optimize and update the model parameters of the cascaded attention convolutional network. The total model loss value L loss The mathematical expression is:

[0106]

[0107] Where M = 4 (M1-M4) is the total number of predicted mask sizes, h i is the weighted value of the balanced joint loss function value of the i-th predicted mask size, h1 = 1, h2 = 1 / 2, h3 = 1 / 4, h4 = 1 / 8. i = 1, which is the final aorta target mask output layer.

[0108] Step 6.2: Setting training hyperparameters and model optimization: Using the total model loss value L loss The model parameters are optimized by the preset AdamW optimizer, and the initial learning rate of the AdamW optimizer is 0.0001. The forward calculation and optimizer optimization are repeated for 40,000 iterations, and finally a trained cascaded attention convolutional network is obtained as a non-enhanced CT image vascular segmentation model.

[0109] Step 7. After image resampling, window width and window level selection, normalization, and block division of the non-enhanced CT image to be processed, a patch block of the non-enhanced CT image to be processed is obtained, and the patch block of the non-enhanced CT image to be processed is input into the trained cascade attention convolutional network to obtain the prediction mask output by the last layer of the decoder, and the prediction masks corresponding to all non-enhanced CT image patch blocks are recombined to generate a predicted target tissue mask, and the target tissue image is obtained according to the non-enhanced CT image and the predicted target tissue mask.

[0110] Preferably, the parameters of the above-mentioned image resampling, window width and window level selection, normalization processing, and blocking are respectively the same as the parameters of the image resampling in step 2.1, the window width and window level in step 2.3, the normalization processing in step 2.4, and the blocking in step 3.2.

[0111] In some embodiments, the overlap of the non-enhanced CT image patches after segmentation is 0.5, and the size of the non-enhanced CT image patches is 96×160×160 for model reasoning.

[0112] In the step of recombining and generating a predicted target tissue mask according to the prediction masks corresponding to all non-enhanced CT image patch blocks, the masks of the overlapping voxel points of each prediction mask are averaged, and the mask with an average mask greater than 0.5 is the final effective mask corresponding to the overlapping voxel point. In this embodiment, the target tissue image obtained by step 7 is the final aorta segmentation result.

[0113] In one embodiment, a computer device is further provided, including a memory and a processor, wherein a computer program is stored in the memory, and the processor implements the steps in the above method embodiments when executing the computer program.

[0114] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps in the above-mentioned method embodiments are implemented.

[0115] In one embodiment, a computer program product is provided, including a computer program, which implements the steps in the above method embodiments when executed by a processor.

[0116] A person of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be implemented by instructing related hardware through a computer program. 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.

[0117] The scope of the present invention is not limited to the above-described embodiments. These examples are only descriptions of preferred implementations of the present invention and are not intended to limit its core ideas. In fact, the various schemes in these embodiments can be further integrated, replaced, or adjusted and improved according to various adjustments and improvements made by those skilled in the art based on the technical ideas of the present invention. All these reasonable changes and optimizations fall within the scope of protection of the present invention.

Claims

1. A segmentation method for non-enhanced CT images based on cascaded attention, characterized in that: The following steps are involved: Step 1: Acquisition of non-enhanced CT images and corresponding enhanced CT images; Step 2, resampling and registering the non-enhanced CT image and the corresponding enhanced CT image in turn, respectively, to obtain a registered non-enhanced CT image and a registered actual enhanced CT image, and adjusting and normalizing the window width and window position of the registered non-enhanced CT image in turn, to obtain a normalized non-enhanced CT image; Step 3: construct a sample based on the normalized non-enhanced CT image, and construct an actual label mask corresponding to the sample based on the registered actual enhanced CT image; Step 4: Build a cascaded attention convolutional network; Step 5: Construct a balanced joint loss function; Step 6: Train the cascaded attention convolutional network; Step 7: After image resampling, window width and window level selection, normalization, and block division of the non-enhanced CT image to be processed, a patch block of the non-enhanced CT image to be processed is obtained, and the patch block of the non-enhanced CT image to be processed is input into the trained cascade attention convolutional network to obtain the corresponding prediction mask, and the prediction masks corresponding to all the non-enhanced CT image patch blocks are recombined to generate a predicted target tissue mask, and the target tissue image is obtained according to the non-enhanced CT image and the predicted target tissue mask.

2. The non-enhanced CT image segmentation method based on cascaded attention according to claim 1, characterized in that: Constructing the sample and the corresponding actual label mask in step 3 includes the following steps: Step 3.1, delineating the target tissue on the registered actual enhanced CT image to obtain the actual target tissue mask; Step 3.2, based on the outlined actual target tissue mask, the registered non-enhanced CT image and the actual target tissue mask are cropped in the same way according to the set ratio range of the target segmentation region voxel-non-target segmentation region voxel, to obtain the non-enhanced CT image patch block and the corresponding actual target segmentation mask image patch block respectively; Step 3.3: Expand the non-enhanced CT image patch and the corresponding actual target segmentation mask patch and use them as samples and corresponding actual label masks respectively.

3. The non-enhanced CT image segmentation method based on cascaded attention according to claim 1, characterized in that: The cascaded attention convolutional network in step 4 includes an encoder, a bottleneck layer, a decoder, and a multi-level target output module. The encoder includes 1 convolution input layer and multiple levels, each level includes a coding convolution layer and a downsampling convolution layer, there is no channel change between the coding convolution layer and the downsampling convolution layer of the same level, and the number of channels between adjacent levels is doubled; the coding convolution layer of each level receives the output features of the downsampling convolution layer of the previous level, the output features of the coding convolution layer are input to the downsampling convolution layer of the same level, the coding convolution layer of the first level receives the output features of the convolution input layer, and the output features of the downsampling convolution layer of the last level are input to the bottleneck layer.

4. The non-enhanced CT image segmentation method based on cascaded attention according to claim 3, characterized in that: The bottleneck layer includes multiple MedNeXt convolution blocks, and the output features of the bottleneck layer are input to the decoder.

5. The non-enhanced CT image segmentation method based on cascaded attention according to claim 4, characterized in that: The decoder includes multiple levels, each level includes a decoding convolution layer and an upsampling convolution layer, there is no channel change between the decoding convolution layer and the upsampling convolution layer of the same level, the channels between adjacent levels are halved, the output features of the bottleneck layer are input to the upsampling convolution layer of the first level of the decoder, the output features of the decoding convolution layer of the previous level of the decoder are input to the upsampling convolution layer of the next level, the output features of the encoding convolution layer of the fth level of the encoder and the output features of the upsampling convolution layer of the l-f+1th level of the decoder are added and input to the decoding convolution layer of the l-f+1th level of the decoder.

6. The non-enhanced CT image segmentation method based on cascaded attention according to claim 5, characterized in that: The multi-level target output module includes a cascaded attention block and a 3D transposed convolution output layer, and the output features of the decoding convolution layer of each level of the decoder are respectively output as prediction masks of different sizes through the cascaded attention block and the 3D transposed convolution output layer.

7. The non-enhanced CT image segmentation method based on cascaded attention according to claim 6, characterized in that: The balanced joint loss function is based on the following formula: Ltotal=W1*Lseg+w2*Lban Among them, Ltotal is the balanced joint loss function, w1 and w2 are weighted values, Lseg is the target segmentation loss function, Lba n is the category balance loss function; The target segmentation loss function Lseg is based on the following formula: Lseg =DSC_loss +CE_loss Among them, X is the predicted mask, Y is the actual mask, the actual mask Y and the predicted mask X are of the same size, the actual mask Y is obtained by downsampling the actual label mask, and X i is the i-th voxel point of the predicted mask, Y i is the i-th voxel point of the actual mask, i is the voxel point sequence number, N is the total number of voxel points of the sample, |X| and |Y| represent the number of voxel points of the predicted mask X and the actual mask Y respectively, |X∩Y| represents the number of voxel points of the intersection of X and Y, DSC_loss is the segmentation loss, and CE_loss is the cross entropy loss; The class balance loss function Lba n Based on the following formula: Lpan=αLLDAMM+βLpL Among them, L LDAM is the label distribution aware margin loss, L FL is the focusing loss, α and β are weighted values; in, is the margin for predicting category k, the constant nk is the number of voxels of predicted category k, is the margin for predicting category j, the constant nj is the number of voxels of predicted category j; is the unnormalized probability of category k of the i-th voxel point of the predicted mask, where k∈{1,...,T}, T is the total number of categories; is the unnormalized probability of category j for the i-th voxel point of the predicted mask; in, αk is the weight of category k, r is the focusing parameter, pk is the probability value of the i-th voxel point Xi of the predicted mask belonging to category k after the Sigmoid function, is one-hot encoding.

8. The non-enhanced CT image segmentation method based on cascaded attention according to claim 7, characterized in that: In step 6, based on the total model loss value Llos s Train the cascaded attention convolutional network, Where M is the total number of predicted mask sizes, h i is the weighted value of the balanced joint loss function value of the i-th predicted mask size, is the balanced joint loss function corresponding to the prediction mask size of the i-th level of the decoder.

9. The non-enhanced CT image segmentation method based on cascaded attention according to claim 1, characterized in that: The non-enhanced CT image patch blocks after the block division have a set overlap degree. In the step of recombining and generating a predicted target tissue mask according to the prediction masks corresponding to all the non-enhanced CT image patch blocks, the masks of the overlapping voxel points of each prediction mask are averaged, and the mask with an average mask greater than 0.5 is the corresponding final effective mask of the overlapping voxel points.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the non-enhanced CT image segmentation method according to any one of claims 1 to 9 are implemented.

Citation Information

Patent Citations

  • Non-contrast agent enhanced medical image segmentation method for target correction

    CN117522898A

  • System and method of camera video sharing for multi-camera applications

    KR102800135B1

  • Image segmentation and object detection using fully convolutional neural network

    US10304193B1

  • Methods and systems for biomedical image segmentation based on a combination of arterial and portal image information

    US20240169544A1