Deep Learning-Based CT Image Segmentation Method, Device and System for Cerebral Hemorrhage

Abstract

The present invention provides a method, device and system for CT image segmentation of intracerebral hemorrhage based on deep learning. The segmentation method includes: acquiring CT tomographic images; inputting the CT tomographic images into a trained CT image segmentation model, and the CT image segmentation model outputs a segmentation image; the CT image segmentation model includes N downsampling modules and N upsampling modules connected in sequence, and a first skip connection path is provided between each upsampling module and the previous downsampling module of the corresponding downsampling module, an AG module is provided on the first skip connection path, and a second skip connection path is provided between adjacent downsampling modules, and a Focus module is provided on the second skip connection path. The segmentation network can effectively retain the integrity of the overall features of the bleeding area by using the Focus module, and makes the segmentation network more focused on the important features related to the task through the AG module. The combination of the Focus module and the AG module enables the segmentation network of the CT image segmentation model to have excellent segmentation effects.

Description

Technical Field

[0001] The present invention relates to the technical fields of machine learning and image processing, and in particular, to a method, device and system for segmenting CT images of cerebral hemorrhage based on deep learning. Background Art

[0002] As the first choice for clinical diagnosis of cerebral hemorrhage, CT has the characteristics of convenience, speed, and definite effect. At present, the "Tada method" is mostly used clinically to measure the volume of cerebral hemorrhage. Its principle is to idealize the bleeding shape as an ellipsoid and calculate according to the formula V = A×B×D×1 / 2, where V is the volume of cerebral hemorrhage, A is the longest diameter of the lesion on the largest bleeding layer of CT, B is the maximum width perpendicular to A on this layer, and D is the number of bleeding layers × slice thickness. The advantage of this method is simplicity and speed. When the bleeding shape is regular, its accuracy is acceptable. However, in clinical practice, the shapes of the vast majority of cerebral hemorrhages are irregular, and limited by the experience of the film readers, there are large errors in the measurement accuracy of the volume of cerebral hemorrhage, causing a certain degree of uncertainty in the formulation of clinical decisions. The method of accurately delineating the bleeding site layer by layer on CT images and calculating the volume is called the CT quantitative method, which can be considered the gold standard for non-invasive measurement of the volume of cerebral hemorrhage. However, its operation is complex and time-consuming, making it difficult to be applied clinically.

[0003] In recent years, with the vigorous development of artificial intelligence technologies represented by deep learning, segmentation methods based on deep learning have been widely used in the segmentation and measurement of brain tissues and have achieved certain results. In terms of cerebral hemorrhage, Cho et al. constructed a cascaded deep learning model for both the detection and segmentation of cerebral hemorrhage, but there are still problems with insufficient accuracy in the segmentation of the cerebral hemorrhage area. Summary of the Invention

[0004] The present invention aims to at least solve the technical problems existing in the prior art, and is capable of accurately segmenting the bleeding area, and provides a method, device and system for segmenting CT images of cerebral hemorrhage based on deep learning.

[0005] To achieve the above object of the present invention, according to the first aspect of the present invention, a method for segmenting CT images of cerebral hemorrhage based on deep learning is provided, including: obtaining CT tomographic images; inputting the CT tomographic images into a trained CT image segmentation model, and the CT image segmentation model segments the hemorrhage area in the CT tomographic images and outputs a segmented image; the segmentation network of the CT image segmentation model is constructed based on the Unet network architecture, including N downsampling modules and N upsampling modules connected in sequence, the downsampling modules and the upsampling modules correspond one by one, and there is a first skip connection path between each upsampling module and the previous downsampling module of the corresponding downsampling module, and an AG module is provided on the first skip connection path, and there is a second skip connection path between adjacent downsampling modules, and a Focus module is provided on the second skip connection path, where N is a positive integer.

[0006] To achieve the above object of the present invention, according to the second aspect of the present invention, a device for segmenting CT images of cerebral hemorrhage based on deep learning is provided, including: an image acquisition module for obtaining CT tomographic images; a CT image segmentation module that inputs the CT tomographic images obtained by the image acquisition module into a trained CT image segmentation model, and the CT image segmentation model segments the hemorrhage area in the CT tomographic images and outputs a segmented image; the segmentation network of the CT image segmentation model is constructed based on the Unet network architecture, including N downsampling modules and N upsampling modules connected in sequence, the downsampling modules and the upsampling modules correspond one by one, and there is a first skip connection path between each upsampling module and the previous downsampling module of the corresponding downsampling module, and an AG module is provided on the first skip connection path, and there is a second skip connection path between adjacent downsampling modules, and a Focus module is provided on the second skip connection path, where N is a positive integer.

[0007] The above segmentation method and segmentation device: construct a CT image segmentation model capable of accurately segmenting the hemorrhage area. In the segmentation network of the CT image segmentation model, the Focus module is used to integrate the image information in the downsampling into the channel space for convolution, avoiding regarding the small hemorrhage points with low signal around the hemorrhage area as redundant information and removing them during the dimensionality reduction operation of the downsampling module, and effectively retaining the integrity of the overall characteristics of the hemorrhage area during the encoding process of the downsampling; the two-dimensional implementation of the AG module is carried out, and the AG module effectively suppresses the low-level redundant features output by the Focus module, making the segmentation network more focused on the important features related to the task. The combination of the Focus module and the AG module enables the segmentation network of the CT image segmentation model to have excellent segmentation effects, and good segmentation effects can be obtained whether for large hemorrhage areas or small hemorrhage points; the CT image segmentation model is a fully automatic segmentation model, without manual participation, and can significantly reduce the workload.

[0008] To achieve the above object of the present invention, according to the third aspect of the present invention, there is provided a cerebral hemorrhage volume measurement system, including: a CT image acquisition module for acquiring cerebral hemorrhage CT images, where the cerebral hemorrhage CT images include L layers of CT tomographic images, and L is a positive integer; a cerebral hemorrhage CT image segmentation device as described in the second aspect of the present invention; an execution module for inputting the L layers of CT tomographic images into the cerebral hemorrhage CT image segmentation device respectively to obtain corresponding segmented images; a bleeding area size acquisition module for counting the number of voxels in the bleeding area in the segmented images corresponding to each CT tomographic image; a cerebral hemorrhage volume calculation module for calculating the cerebral hemorrhage volume V according to the following formula: where i' represents the CT tomographic image index, 1 ≤ i' ≤ L; X represents the voxel spacing in the horizontal direction; Y represents the voxel spacing in the vertical direction; T represents the CT scan layer thickness; P i' represents the number of voxels in the bleeding area in the segmented image corresponding to the i'-th CT tomographic image.

[0009] The above cerebral hemorrhage volume measurement system: In addition to having the beneficial technical effects of the above segmentation method and segmentation device, the system also has the beneficial technical effect of being able to accurately measure the volume of irregular bleeding areas. The accuracy and efficiency of the cerebral hemorrhage volume measurement results are much higher than those of the Tada method, realizing fully automatic volume measurement and significantly reducing the workload of clinical cerebral hemorrhage volume measurement. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] Figure 1 is a flowchart of the deep learning-based cerebral hemorrhage CT image segmentation method in Embodiment 1 of the present invention;

[0011] Figure 2 is a schematic diagram of the segmentation network structure of the CT image segmentation model in Embodiments 1 and 2 of the present invention;

[0012] Figure 3 is a schematic diagram of the Focus module structure in Embodiments 1 and 2 of the present invention;

[0013] Figure 4 is a schematic diagram of the AG module structure in Embodiments 1 and 2 of the present invention;

[0014] Figure 5 is the segmented image obtained by the present invention and the segmented image obtained by the Tada method when the bleeding area is relatively regular in Embodiment 1 of the present invention;

[0015] Figure 6 is the segmented image obtained by the present invention and the segmented image obtained by the Tada method when the bleeding area is irregular and varies greatly in Embodiment 1 of the present invention;

[0016] Figure 7 is the three-dimensional reconstruction effect diagram of cerebral hemorrhage in Embodiment 1 of the present invention;

[0017] Figure 8 It is the structural block diagram of the cerebral hemorrhage volume measurement system in Embodiment 3 of the present invention;

[0018] Figure 9 It is the schematic diagram for judging the consistency of the measurement results of the cerebral hemorrhage volume measurement system in Embodiment 3 of the present invention. Detailed implementation manners

[0019] The embodiments of the present invention will be described in detail below. The examples of the embodiments are shown in the drawings, where the same or similar reference numerals represent the same or similar elements or elements with the same or similar functions throughout. The embodiments described below by referring to the drawings are exemplary and are only used to explain the present invention and should not be construed as a limitation to the present invention.

[0020] In the description of the present invention, it should be understood that the orientation or positional relationships indicated by the terms "longitudinal", "transverse", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc. are based on the orientation or positional relationships shown in the drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and thus should not be construed as a limitation to the present invention.

[0021] In the description of the present invention, unless otherwise specified and defined, it should be noted that the terms "installation", "connection", and "connection" should be understood in a broad sense. For example, it can be a mechanical connection or an electrical connection, or it can be the communication inside two elements. It can be directly connected or indirectly connected through an intermediate medium. For those of ordinary skill in the art, the specific meanings of the above terms can be understood according to specific situations.

[0022] Embodiment 1

[0023] This embodiment discloses a cerebral hemorrhage CT image segmentation method based on deep learning. The flow schematic diagram of this segmentation method is as Figure 1 shown and includes:

[0024] Step S1, obtaining CT tomographic images. General brain CT images are obtained by continuous tomographic scanning, that is, they include multiple layers of CT tomographic images.

[0025] To improve the efficiency of subsequent image segmentation processing and ensure the segmentation effect of the bleeding area, further preferably, before inputting the CT tomographic image into the segmentation network of the CT image segmentation model, windowing processing is performed on the CT tomographic image according to the CT value range of the bleeding area, the image after windowing processing is normalized, and the normalized image is input into the CT image segmentation model. The CT value range is a preset range, which can be set according to actual experience. The CT value range is preferably but not limited to [-60, 140]. Windowing processing means only retaining the pixel points with CT values within the range [-60, 140] in the CT tomographic image and removing the remaining pixel points. The normalization processing is preferably but not limited to normalizing through the maximum CT value and the minimum CT value in the image after windowing processing, which is a prior art and will not be elaborated here.

[0026] Step S2, input the CT tomographic image into the trained CT image segmentation model, and the CT image segmentation model segments the bleeding area in the CT tomographic image and outputs the segmented image.

[0027] In this embodiment, preferably, as Figure 2 shown, the segmentation network of the CT image segmentation model is constructed based on the Unet network architecture, including N layers of downsampling modules and N layers of upsampling modules connected in sequence. The downsampling modules and the upsampling modules correspond one by one, that is, the k-th downsampling module corresponds to the j-th upsampling module, k = 1, 2,..., N, and j represents the layer index. There is a first skip connection path between each upsampling module and the previous downsampling module of the corresponding downsampling module. An AG module is provided on the first skip connection path, and there is a second skip connection path between adjacent downsampling modules. A Focus module is provided on the second skip connection path. N is a positive integer, and N is preferably but not limited to 3 to 15, and preferably but not limited to 5.

[0028] In this embodiment, AG is the abbreviation of Attention Gate, which represents the attention gate model. In this embodiment, the AG module fuses the feature information of the same size, ensuring that the upsampling process can incorporate the image features in the downsampling process, reducing feature loss, and at the same time can more fully extract the transmitted features. The output image of each upsampling module has the same size as the output image of the previous downsampling module of the corresponding downsampling module. The Focus module is used for slice fusion of the image. The specific structure is preferably but not limited to referring to the Focus module disclosed in the Chinese patent with the publication number CN114663769A or the Focus layer structure in YOLOv5, which will not be elaborated here.

[0029] In this embodiment, in order to maintain high segmentation accuracy while maintaining efficient processing, further preferably, N is 5. As Figure 2As shown in the figure, the segmentation network of the CT image segmentation model includes a first downsampling module, a second downsampling module, a third downsampling module, a fourth downsampling module, a fifth downsampling module, a fifth upsampling module, a fourth upsampling module, a third upsampling module, a second upsampling module, and a first upsampling module connected in sequence; and also includes four second skip connection paths respectively connecting the first downsampling module and the second downsampling module, the second downsampling module and the third downsampling module, the third downsampling module and the fourth downsampling module, and the fourth downsampling module and the fifth downsampling module; and also includes four first skip connection paths respectively connecting the first downsampling module and the second upsampling module, the second downsampling module and the third upsampling module, the third downsampling module and the fourth upsampling module, and the fourth downsampling module and the fifth upsampling module.

[0030] As Figure 2 shown, the first downsampling module includes an encoding convolution D1, and the second downsampling module, the third downsampling module, the fourth downsampling module, and the fifth downsampling module each include a max pooling layer and an encoding convolution connected in sequence. The second skip connection path between adjacent downsampling modules connects the encoding convolution of the previous downsampling module and the max pooling layer of the next downsampling module.

[0031] Specifically, as Figure 2 shown, the second downsampling module includes a max pooling layer MP2 and an encoding convolution D2, the third downsampling module includes a max pooling layer MP3 and an encoding convolution D3, the fourth downsampling module includes a max pooling layer MP4 and an encoding convolution D4, and the fifth downsampling module includes a max pooling layer MP5 and an encoding convolution D5; the second skip connection path between the adjacent first downsampling module and the second downsampling module connects the encoding convolution D1 and the max pooling layer MP2, the second skip connection path between the adjacent second downsampling module and the third downsampling module connects the encoding convolution D2 and the max pooling layer MP3, the second skip connection path between the adjacent third downsampling module and the fourth downsampling module connects the encoding convolution D3 and the max pooling layer MP4, and the second skip connection path between the adjacent fourth downsampling module and the fifth downsampling module connects the encoding convolution D4 and the max pooling layer MP5.

[0032] Specifically, as Figure 2 shown, the first upsampling module includes a decoding convolution U1, and the convolution kernel size of the decoding convolution U1 is preferably but not limited to 1*1. The fifth upsampling module, the fourth upsampling module, the third upsampling module, and the second upsampling module each include an upsampling layer and a decoding convolution connected in sequence. The upsampling layer is preferably but not limited to a transposed convolution layer. The first skip connection path connected to each upsampling module connects the upsampling layer of this upsampling module and the encoding convolution of the previous downsampling module corresponding to this upsampling module.

[0033] Among them, as Figure 2As shown, the second upsampling module includes an upsampling layer US2 and a decoding convolution U2, the third upsampling module includes an upsampling layer US3 and a decoding convolution U3, the fourth upsampling module includes an upsampling layer US4 and a decoding convolution U4, and the fifth upsampling module includes an upsampling layer US5 and a decoding convolution U5. The first skip connection path between the first downsampling module and the second upsampling module connects the encoding convolution D1 and the upsampling layer US2, the first skip connection path between the second downsampling module and the third upsampling module connects the encoding convolution D2 and the upsampling layer US3, the first skip connection path between the third downsampling module and the fourth upsampling module connects the encoding convolution D3 and the upsampling layer US4, and the first skip connection path between the fourth downsampling module and the fifth upsampling module connects the encoding convolution D4 and the upsampling layer US5. The convolution kernel sizes of all encoding convolutions and decoding convolutions except the decoding convolution U1 are preferably but not limited to 3*3. In Figure 2 where, the symbol represents the fusion of two features, and the fusion method is preferably but not limited to addition.

[0034] In this embodiment, in the image segmentation method based on deep learning, most encoders use conventional pooling operations for feature dimensionality reduction. Since some small bleeding points around cerebral hemorrhage appear as low signals on CT images, they may be regarded as redundant information and removed during feature extraction. By introducing the Focus module to integrate image information into the channel space for convolution, the complete image features without pooling operations are fused with the features after conventional pooling operations, so as to effectively retain the integrity of the overall features of cerebral hemorrhage during the encoding process. Therefore, further preferably, the schematic diagram of the execution process of the Focus module is as shown in Figure 3 shown, the Focus module on the second skip connection path between adjacent downsampling modules executes:

[0035] Step A, sample the output image of the encoding convolution in the previous downsampling module with a stride of 2, so that the features of the output image are evenly distributed into 4C channels, and the size of each channel is C×H / 2×W / 2. The size of the output image is C×H×W, where H is the height and W is the width. The number of channels of the output image of the encoding convolution in the previous downsampling module is C, and C is a positive integer.

[0036] Step B, perform a convolution operation on the sampled image to output a feature map including C / 2 channels; preferably, for subsequent processing convenience, after the convolution operation, steps of batch normalization and Relu activation function processing are also performed, and finally a feature map with a size of C / 2×H / 2×W / 2 is output.

[0037] Step C, perform feature fusion on the feature map and the output image of the max-pooling layer of the next downsampling module, and input the obtained feature fusion image into the encoding convolution processing of the next downsampling module.

[0038] In this embodiment, the output of the Focus module will have low-level redundant features. To effectively suppress them, it can be achieved through the AG module, which makes it more focused on the important features related to the task. As Figure 4 shown, the two-dimensional implementation process of the AG module on the second skip connection path connected to each upsampling module is executed as follows:

[0039] Step a: Perform a convolution operation on the output image Up-Input of the upsampling layer of the current upsampling module to obtain a first convolutional image, and perform a convolution operation on the encoded convolutional output image Skip-Input of the previous downsampling module corresponding to the current upsampling module's downsampling module to obtain a second convolutional image; assume that the current upsampling module is the i-th upsampling module, the feature sizes of Skip-Input and Up-Input are the same, and Skip-Input is the feature output by the encoded convolutional layer Dj of the j-th downsampling module, where j = i - 1. In step a, the convolution kernel size in the convolution operation is preferably but not limited to 1*1.

[0040] Step b: Combine the first convolutional image and the second convolutional image, input the combined image into the first activation layer. The first activation layer preferably but not limited to uses the Relu activation function. After performing a convolution operation on the output image of the first activation layer, input it into the second activation layer, and the second activation layer outputs an attention coefficient matrix. The second activation layer preferably but not limited to uses the Sigmoid activation function. In step b, the convolution kernel size in the convolution operation is preferably but not limited to 1*1.

[0041] Step c: Multiply the attention coefficient matrix by the encoded convolutional output image Skip-Input of the previous downsampling module corresponding to the current upsampling module to obtain an AG-processed image, which is the output of the two-dimensional AG module.

[0042] Step d: The second skip connection path combines the AG-processed image with the output image of the upsampling layer of the current upsampling module and then inputs it into the decoded convolutional processing of the current upsampling module.

[0043] In this embodiment, preferably, the construction process of the CT image segmentation model includes:

[0044] Step 1: Construct the segmentation network of the CT image segmentation model according to the network structure shown in Figure 2 the figure.

[0045] Step 2: Obtain the CT image dataset of cerebral hemorrhage. Based on the CT image dataset of cerebral hemorrhage, obtain the CT sectional image set of cerebral hemorrhage, and randomly divide the CT sectional image set of cerebral hemorrhage into training set, validation set, and test set according to the ratio of 7:1:2. The CT image set of cerebral hemorrhage can be from the First Affiliated Hospital of Army Medical University, and all are the cranial CT images of cerebral hemorrhage patients within 24 hours of admission, with a total of 1027 cases. The CT slice thickness is 4.0 mm. The true cerebral hemorrhage area GroundTruth (GT) of each sectional CT image is manually segmented by clinicians. On this basis, the true cerebral hemorrhage volume V-GT (Volume-GT) can be further obtained.

[0046] Step 3: The CT image segmentation model is represented by AttFocusNet. AttFocusNet is implemented based on the Pytorch deep learning framework. Use the training set to train AttFocusNet, use the validation set to verify the trained AttFocusNet, and use the test set to test the performance of AttFocusNet. Finally, obtain the trained CT image segmentation model AttFocusNet that meets the accuracy requirements. During the training process, use a workstation equipped with NVIDIA Quadro RTX5000 video memory for model training, the video memory is 16G, the optimizer uses the RMSprop optimization function, the initial learning rate is 0.00001, the epoch is set to 40, and the batchsize is set to 2. The adjustment of the learning rate adopts a conditional trigger strategy, which is triggered when the model does not converge for 2 consecutive epochs. Resample the resolution of all CT images to 512×512, remove the scanning layers of the chest and abdomen, and retain 40 layers of CT images for each patient.

[0047] In an application scenario of this embodiment, the segmentation performance of the CT image segmentation model AttFocusNet is tested and verified. Dice, intersection over union (Iou), sensitivity (Sensitivity), positive predictive value (PPV), and 95% Hausdorff Distance (HD) are used as the evaluation indicators for cerebral hemorrhage segmentation performance. If True Positive (TP) represents the number of true positive samples predicted as positive samples, False Positive (FP) represents the number of true negative samples predicted as positive samples; False Negative (FN) represents the number of true positive samples predicted as negative samples, and the calculation formulas for each indicator are as follows:

[0048]

[0049]

[0050]

[0051]

[0052] HD represents calculating the distance between the surface point sets of the real sample and the predicted sample, and its expression is

[0053]

[0054] where G' represents the Ground truth, P' represents the predicted value. The purpose of taking 95% HD is to exclude the influence of very small outlier clusters and maintain the stability of the overall result.

[0055] In this application scenario, the AttFocusNet constructed in this embodiment is compared with Unet++, AttUnet, PraNet, 3DUnet, and UNETR in terms of segmentation performance. As shown in Table 1, where Unet++, AttUnet, and PraNet are two-dimensional segmentation networks, and their input and output are both one layer of CT; 3DUnet and UNETR are three-dimensional segmentation networks. They regard the CT data of each patient as a whole, and their input and output are both a CT sequence. It can be seen from Table 1 below that AttFocusNet has the best comprehensive performance, being superior to other networks in terms of Dice, Iou, Sensitivity, and PPV, and is more prominent in HD95, fully demonstrating the effectiveness of AttFocusNet. In addition, AttFocusNet combines the Focus structure on the basis of AttUNet. It can be seen from Table 1 that compared with AttUNet, AttFocusNet has a certain degree of improvement in all 5 indicators, fully illustrating the effectiveness of the Focus structure in feature preservation.

[0056] Table 1 Performance comparison of various segmentation methods in the intracerebral hemorrhage segmentation task

[0057]

[0058] To verify the efficiency of intracerebral hemorrhage volume measurement based on AttFocusNet, its efficiency is compared with that of the multi-tada method and the efficiency of manual segmentation by doctors (based on Mimics software). For each patient, on average, it takes 170.1 seconds in the manual contouring mode by clinicians, 47.7 seconds for the multi-tada method, and only 5.6 seconds for the method based on AttFocusNet. Therefore, the segmentation method provided in this embodiment can not only provide high accuracy in intracerebral hemorrhage volume measurement but also significantly reduce the workload of clinicians.

[0059] It can be seen that there are obvious differences between the segmentation method provided in this embodiment and the measurement results of the Tada method in most cases. Figure 5 It is a schematic diagram of the segmentation results of the Tada method GT and the segmentation method AttFocusNet provided in this embodiment when the bleeding area is relatively regular. From Figure 5 It can be seen that when the bleeding shape is relatively regular, the difference in the bleeding area segmentation results between the Tada method and the segmentation method provided in this embodiment is small, and the difference in the bleeding volume measurement results obtained based on the segmentation results is small. From Figure 6 It can be seen that when the blood shape is irregular and varies greatly, the difference in the bleeding area segmentation results between the Tada method and the segmentation method provided in this embodiment is large, and the volume accuracy measured by the Tada method is low. However, in most cases, the actual intracerebral hemorrhage area has an irregular shape. Therefore, the segmentation method provided in this embodiment has higher application value in clinical intracerebral hemorrhage volume measurement.

[0060] In this embodiment, in order to provide more information about intracerebral hemorrhage for clinical use and provide technical support for subsequent clinical treatment, after segmenting the bleeding area of the CT tomographic images in the brain CT images, three-dimensional visualization of intracerebral hemorrhage is realized based on the segmented images using the Python programming language. The visualization processing results are as Figure 7 shown, where the dark black area represents the intracerebral hemorrhage area.

[0061] Embodiment 2

[0062] This embodiment discloses a device for segmenting intracerebral hemorrhage CT images based on deep learning. This segmentation device corresponds to the segmentation method in Embodiment 1 and includes: an image acquisition module for acquiring CT tomographic images; a CT image segmentation module that inputs the CT tomographic images obtained by the image acquisition module into a trained CT image segmentation model. The CT image segmentation model segments the bleeding area in the CT tomographic images and outputs a segmented image. The segmentation network of the CT image segmentation model is constructed based on the Unet network architecture and includes N layers of downsampling modules and N layers of upsampling modules connected in sequence. The downsampling modules and the upsampling modules correspond one by one. There is a first skip connection path between each layer of upsampling module and the previous layer of downsampling module corresponding to the upsampling module. An AG module is provided on the first skip connection path. There is a second skip connection path between adjacent layers of downsampling modules. A Focus module is provided on the second skip connection path, where N is a positive integer.

[0063] In this embodiment, as Figure 2As shown, the segmentation network of the CT image segmentation model includes a first downsampling module, a second downsampling module, a third downsampling module, a fourth downsampling module, a fifth downsampling module, a fifth upsampling module, a fourth upsampling module, a third upsampling module, a second upsampling module, and a first upsampling module connected in sequence. The connection relationships of each module can be referred to the description in Embodiment 1 and will not be elaborated here.

[0064] In this embodiment, as Figure 2 shown, the first downsampling module includes an encoding convolution. The second downsampling module, the third downsampling module, the fourth downsampling module, and the fifth downsampling module all include a max pooling layer and an encoding convolution connected in sequence. The second skip connection path between adjacent downsampling modules connects the encoding convolution of the upper-layer downsampling module and the max pooling layer of the lower-layer downsampling module. The detailed introduction of each module can be referred to the description in Embodiment 1 and will not be elaborated here.

[0065] In this embodiment, as Figure 2 shown, the first upsampling module includes a decoding convolution. The fifth upsampling module, the fourth upsampling module, the third upsampling module, and the second upsampling module all include an upsampling layer and a decoding convolution connected in sequence. The first skip connection path connected to each upsampling module connects the upsampling layer of this upsampling module and the encoding convolution of the upper-layer downsampling module corresponding to this upsampling module. The detailed introduction of each module can be referred to the description in Embodiment 1 and will not be elaborated here.

[0066] In this embodiment, the Focus module on the second skip connection path between adjacent downsampling modules performs:

[0067] Sampling the output image of the encoding convolution in the upper-layer downsampling module with a stride of 2, and evenly distributing the features of the output image into 4C channels. The number of channels of the output image of the encoding convolution in the upper-layer downsampling module is C, and C is a positive integer;

[0068] Performing a convolution operation on the sampled image to output a feature map including C / 2 channels;

[0069] Performing feature fusion on the feature map and the output image of the max pooling layer of the lower-layer downsampling module, and inputting the obtained feature fusion image into the encoding convolution processing of the lower-layer downsampling module.

[0070] In this embodiment, the AG module on the second skip connection path connected to each upsampling module performs:

[0071] Performing a convolution operation on the output image of the upsampling layer of this upsampling module to obtain a first convolution image, and performing a convolution operation on the output image of the encoding convolution of the upper-layer downsampling module corresponding to this upsampling module to obtain a second convolution image;

[0072] Merge the first convolutional image and the second convolutional image, input the merged image into the first activation layer, perform a convolutional operation on the output image of the first activation layer and then input it into the second activation layer, and the second activation layer outputs an attention coefficient matrix;

[0073] Multiply the attention coefficient matrix by the encoded convolutional output image of the previous downsampling module corresponding to the downsampling module of the upsampling module of this layer to obtain an AG processed image;

[0074] The second skip connection path merges the AG processed image with the output image of the upsampling layer of the upsampling module of this layer and then inputs it into the decoded convolutional process of the upsampling module of this layer.

[0075] Embodiment 3

[0076] This embodiment discloses a cerebral hemorrhage volume measurement system, and its system block diagram is as Figure 8 shown, including: a CT image acquisition module for acquiring cerebral hemorrhage CT images, the cerebral hemorrhage CT images include L layers of CT tomographic images, L is a positive integer, L is preferably but not limited to being greater than or equal to 30, preferably, L is 40; the cerebral hemorrhage CT image segmentation device provided in Embodiment 2; an execution module that inputs the L layers of CT tomographic images into the cerebral hemorrhage CT image segmentation device respectively to obtain corresponding segmented images; a hemorrhage area size acquisition module that counts the voxel numbers of the hemorrhage areas in the segmented images corresponding to each CT tomographic image; a cerebral hemorrhage volume calculation module that calculates the cerebral hemorrhage volume V according to the following formula:

[0077]

[0078] where, i' represents the CT tomographic image index, 1≤i'≤L; X represents the voxel spacing in the horizontal direction; Y represents the voxel spacing in the vertical direction; T represents the CT scan layer thickness, T is preferably but not limited to 3 mm to 6 mm, preferably 4 mm; P i' represents the voxel number of the hemorrhage area in the segmented image corresponding to the i'-th CT tomographic image. The voxel spacing is a common parameter in medical images, and its meaning is how many millimeters the straight-line distance between the centers of two adjacent voxels (i.e., two points) in the image is converted into the real-world distance.

[0079] In an application scenario of this embodiment, the consistency of the cerebral hemorrhage volume of this system is evaluated, and the consistency of the cerebral hemorrhage volume can be evaluated by linear regression, as Figure 9As shown, the horizontal axis scale represents the true intracerebral hemorrhage volume V-GT, where GT is the abbreviation of Groundtruth, and the vertical axis represents the intracerebral hemorrhage volumes measured by various methods (AttFocusNet for this system and Coniglobus formula for the Tada method). Judging from the fitting of the linear regression, the volume distribution measured by the Tada method has a large degree of dispersion, and there are a large number of outliers, with poor consistency with V-GT. The ICC (intraclass correlation coefficient) is 0.776, indicating a low degree of linear correlation between the measurement results of the Tada method and V-GT. The fitting of the measurement results after segmentation by AttFocusNet is relatively ideal, with few outliers and high consistency with V-GT. The ICC reaches 0.997, indicating a high degree of linear correlation between the intracerebral hemorrhage volume measurement results of this system and V-GT. Compared with the Tada method, the measurement results of this system are closer to the true intracerebral hemorrhage volume.

[0080] In the description of this specification, the descriptions referring to terms such as "one embodiment", "some embodiments", "examples", "specific examples", or "some examples", etc. mean that the specific features, structures, materials, or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.

[0081] Although the embodiments of the present invention have been shown and described, those of ordinary skill in the art can understand that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and purposes of the present invention. The scope of the present invention is defined by the claims and their equivalents.

Claims

1. A method for CT image segmentation of intracerebral hemorrhage based on deep learning, characterized in that, it includes: Obtain CT tomographic images; input the CT tomographic images into a trained CT image segmentation model, and the CT image segmentation model segments the hemorrhage area in the CT tomographic images and outputs a segmentation image; The segmentation network of the CT image segmentation model is constructed based on the Unet network architecture, and includes N downsampling modules and N upsampling modules connected in sequence. The downsampling modules and the upsampling modules correspond one by one. There is a first skip connection path between each upsampling module and the previous downsampling module of the corresponding downsampling module. An AG module is provided on the first skip connection path. There is a second skip connection path between adjacent downsampling modules. A Focus module is provided on the second skip connection path, where N is a positive integer; Among them, the Focus module on the second skip connection path between adjacent downsampling modules performs: Sampling the output image of the encoding convolution in the previous downsampling module with a stride of 2. The number of channels of the output image is C, and C is a positive integer; Performing a convolution operation on the sampled image to output a feature map including C / 2 channels; Performing feature fusion on the feature map and the output image of the max pooling layer of the next downsampling module, and inputting the obtained feature fusion image into the encoding convolution processing of the next downsampling module; Among them, the AG module on the first skip connection path connected to each upsampling module performs: Performing a convolution operation on the output image of the upsampling layer of the current upsampling module to obtain a first convolution image, and performing a convolution operation on the output image of the encoding convolution of the previous downsampling module of the corresponding downsampling module of the current upsampling module to obtain a second convolution image; Merging the first convolution image and the second convolution image, inputting the merged image into the first activation layer, performing a convolution operation on the output image of the first activation layer and then inputting it into the second activation layer, and the second activation layer outputs an attention coefficient matrix; Multiplying the attention coefficient matrix by the output image of the encoding convolution of the previous downsampling module of the corresponding downsampling module of the current upsampling module to obtain an AG processed image; The first skip connection path merges the AG processed image and the output image of the upsampling layer of the current upsampling module and then inputs it into the decoding convolution processing of the current upsampling module.

2. The method for CT image segmentation of intracerebral hemorrhage based on deep learning according to claim 1, characterized in that, The segmentation network of the CT image segmentation model includes a first downsampling module, a second downsampling module, a third downsampling module, a fourth downsampling module, a fifth downsampling module, a fifth upsampling module, a fourth upsampling module, a third upsampling module, a second upsampling module, and a first upsampling module connected in sequence; The first downsampling module includes an encoding convolution. The second downsampling module, the third downsampling module, the fourth downsampling module, and the fifth downsampling module all include a max pooling layer and an encoding convolution connected in sequence. The second skip connection path between adjacent downsampling modules connects the encoding convolution of the previous downsampling module and the max pooling layer of the next downsampling module; The first upsampling module includes a decoding convolution. The fifth upsampling module, the fourth upsampling module, the third upsampling module, and the second upsampling module each include an upsampling layer and a decoding convolution connected in sequence. The first skip connection path connected to each upsampling module connects the upsampling layer of the current upsampling module and the encoding convolution of the previous downsampling module of the corresponding downsampling module of the current upsampling module.

3. The method for segmenting CT images of cerebral hemorrhage based on deep learning according to claim 1, characterized in that, after obtaining the CT tomographic image, windowing the CT tomographic image according to the CT value value range of the bleeding area, normalizing the windowed image, and inputting the normalized image into the CT image segmentation model.

4. A device for segmenting CT images of cerebral hemorrhage based on deep learning, characterized in that, comprising: an image acquisition module for acquiring CT tomographic images; a CT image segmentation module that inputs the CT tomographic image obtained by the image acquisition module into a trained CT image segmentation model. The CT image segmentation model segments the bleeding area in the CT tomographic image and outputs a segmentation image. The CT image segmentation model segments the bleeding area in the CT tomographic image and outputs a segmentation image, including N layers of downsampling modules and N layers of upsampling modules connected in sequence. The downsampling modules and the upsampling modules correspond one by one. There is a first skip connection path between each upsampling module and the previous downsampling module of the corresponding downsampling module. An AG module is provided on the first skip connection path. There is a second skip connection path between adjacent downsampling modules. A Focus module is provided on the second skip connection path. N is a positive integer; wherein, the Focus module on the second skip connection path between adjacent downsampling modules performs: sampling the output image of the encoding convolution in the previous downsampling module with a stride of 2. The number of channels of the output image is C, and C is a positive integer; performing a convolution operation on the sampled image to output a feature map including C / 2 channels; performing feature fusion on the feature map and the output image of the max-pooling layer of the next downsampling module, and inputting the obtained feature fusion image into the encoding convolution processing of the next downsampling module; wherein, the AG module on the first skip connection path connected to each upsampling module performs: performing a convolution operation on the output image of the upsampling layer of the current upsampling module to obtain a first convolution image, and performing a convolution operation on the output image of the encoding convolution of the previous downsampling module of the corresponding downsampling module of the current upsampling module to obtain a second convolution image; merging the first convolution image and the second convolution image, inputting the merged image into the first activation layer, performing a convolution operation on the output image of the first activation layer and then inputting it into the second activation layer, and the second activation layer outputs an attention coefficient matrix; multiplying the attention coefficient matrix by the output image of the encoding convolution of the previous downsampling module of the corresponding downsampling module of the current upsampling module to obtain an AG processed image; The first skip connection path merges the AG processed image and the output image of the upsampling layer of the current upsampling module and then inputs it into the decoding convolution processing of the current upsampling module.

5. The intracerebral hemorrhage CT image segmentation device based on deep learning according to claim 4, characterized in that, the segmentation network of the CT image segmentation model includes a first downsampling module, a second downsampling module, a third downsampling module, a fourth downsampling module, a fifth downsampling module, a fifth upsampling module, a fourth upsampling module, a third upsampling module, a second upsampling module, and a first upsampling module connected in sequence; the first downsampling module includes an encoding convolution, and the second downsampling module, the third downsampling module, the fourth downsampling module, and the fifth downsampling module each include a max pooling layer and an encoding convolution connected in sequence. A second skip connection path between adjacent downsampling modules connects the encoding convolution of the previous downsampling module and the max pooling layer of the next downsampling module; the first upsampling module includes a decoding convolution, and the fifth upsampling module, the fourth upsampling module, the third upsampling module, and the second upsampling module each include an upsampling layer and a decoding convolution connected in sequence. The first skip connection path connected to each upsampling module connects the upsampling layer of this upsampling module and the encoding convolution of the previous downsampling module corresponding to this upsampling module.

6. An intracerebral hemorrhage volume measurement system, characterized in that, it includes: a CT image acquisition module for acquiring intracerebral hemorrhage CT images, where the intracerebral hemorrhage CT images include L layers of CT tomographic images, and L is a positive integer; the intracerebral hemorrhage CT image segmentation device according to claim 4 or 5; an execution module that inputs the L layers of CT tomographic images into the intracerebral hemorrhage CT image segmentation device respectively to obtain corresponding segmentation images; a hemorrhage area size acquisition module for counting the number of voxels in the hemorrhage area in the segmentation images corresponding to each CT tomographic image; an intracerebral hemorrhage volume calculation module for calculating the intracerebral hemorrhage volume V according to the following formula: where, $i'$ represents the CT tomographic image index, where $1\leq i'\leq L$; $X$ represents the voxel spacing in the horizontal direction; $Y$ represents the voxel spacing in the vertical direction; $T$ represents the CT scan layer thickness; $P$ i' represents the number of voxels in the bleeding area of the segmentation image corresponding to the $i'$-th CT tomographic image.

Images