A Medical Image Segmentation Method and System Based on Prior Target Location Information

Abstract

This invention discloses a medical image segmentation method and system based on prior information about the target location. First, a medical image is acquired, and then a medical image segmentation neural network is used to obtain the segmented result image. The medical image segmentation neural network is a pre-trained network; specifically, it generates a distance map based on the label image, and uses the distance map as weights to weight the loss function during training. This application employs a medical image segmentation method based on prior information about the target location, guiding the network to focus more on indistinguishable edge regions in the medical image during training, which can significantly improve the segmentation accuracy of the medical image segmentation method. Furthermore, the method used in this application can be easily applied to various medical image segmentation methods, and consumes less time and storage resources during training, without requiring additional testing time.

Description

Technical Field

[0001] This application belongs to the field of medical image processing technology, and relates to an image segmentation method and system, particularly to a medical image segmentation method and system based on prior information of target location. Background Technology

[0002] Modern hospitals typically use four methods to acquire images of the human body that require diagnosis: X-rays, computed tomography (CT), magnetic resonance imaging (MRI), and ultrasound. These methods cause minimal damage to the patient's body and are therefore frequently used in disease diagnosis, surgical planning, and prognostic assessment. Medical image segmentation, as an important technique in computer-aided medicine, involves classifying each pixel in the acquired medical images, allowing doctors to more directly observe changes in the anatomical or pathological structures of organs, thereby assisting doctors in more efficient diagnostic work. With the continuous development and widespread use of these devices, medical image segmentation technology is playing an increasingly important role in the field of computer-aided medicine.

[0003] In recent years, with the rapid development of deep learning, especially convolutional neural networks, medical image segmentation technology has made significant progress. Improved segmentation accuracy and efficiency have made it possible to assist doctors in diagnosis using deep learning-based medical image segmentation techniques. The most classic and widely used network model in the field of medical image segmentation is U-Net. U-Net adds skip connections between the encoder and decoder structures, effectively fusing shallow texture features with deep semantic features by connecting features in the same layer of the encoder and decoder.

[0004] Although U-Net was proposed early on, it is limited by the inherent limitations of convolutional neural networks and does not pay enough attention to the structural information of the segmented target. Later, many other network structures based on U-Net emerged, but most of these methods modified the network structure, resulting in poor transferability. Furthermore, most segmentation methods did not pay attention to the prior information of the target location, which could be utilized, or did not make full use of it.

[0005] Meanwhile, the sheer volume of medical image data also hinders medical image segmentation methods from achieving more satisfactory results. Generally, more training data leads to better network performance; however, due to the high cost of acquiring and annotating medical images, finding a better-performing medical image segmentation method on limited medical image data is particularly important. Medical images contain a wealth of anatomical and structural prior information. If this additional information can be incorporated into the training process of medical image segmentation methods, segmentation results that better conform to anatomical structures can be obtained, thus improving the segmentation performance of medical image segmentation methods. Summary of the Invention

[0006] To address the issues of insufficient utilization of target location information and inadequate image segmentation data in existing medical image segmentation networks, this invention provides a medical image segmentation method and system based on target location prior information that fully utilizes structural prior information in medical images.

[0007] The technical solution adopted by the method of the present invention is: a medical image segmentation method based on prior information of target location, comprising the following steps:

[0008] Step 1: Acquire medical images;

[0009] Step 2: Obtain the segmentation result image using a medical image segmentation neural network;

[0010] The medical image segmentation neural network includes an image downsampling encoder, a feature upsampling encoder, a same-layer fusion module, and a segmentation result generation module;

[0011] The image downsampling encoder comprises five cascaded convolutional modules, each of which contains two cascaded convolutional layers. The kernel size of the two convolutional layers is 3, and the stride is 1. Each convolutional layer is followed by a normalization layer and a corrected linear unit layer.

[0012] The feature upsampling encoder includes four cascaded convolutional modules, each of which contains an upsampling module and a convolutional module. The upsampling module contains, in sequence, a bilinear interpolation upsampling layer, a convolutional layer, a normalization layer, and a corrected linear unit layer. The convolutional layer has a kernel size of 3 and a stride of 1. The convolutional module includes two convolutional layers, each with a kernel size of 3 and a stride of 1. Each convolutional layer is followed by a normalization layer and a corrected linear unit layer.

[0013] The same-layer fusion module is used to concatenate the output features F1, F2, F3, and F4 of the first four stages of the image downsampling encoder with the output features G1, G2, G3, and G4 of each stage of the feature upsampling encoder in the order of G4, G3, G2, and G1.

[0014] The segmentation result generation module is used to generate segmentation results of corresponding size based on the features output by each convolutional module of the feature upsampling encoder. This module consists of a convolutional layer with a kernel size of 1 and a stride of 1.

[0015] The technical solution adopted by the system of the present invention is: a medical image segmentation system based on prior information of target location, comprising:

[0016] One or more processors;

[0017] A storage device for storing one or more programs, which, when executed by one or more processors, enable the one or more processors to implement the medical image segmentation method based on prior information about the target location.

[0018] The beneficial effects of this invention include:

[0019] (1) This invention employs a medical image segmentation method based on prior information about the target location to segment medical images. The loss function is weighted using a distance map calculated from the labels, allowing the medical image segmentation network to pay different levels of attention to different regions during training, such as focusing more on difficult-to-distinguish edge regions, thereby integrating structural information into the segmentation network. Theoretically, this invention can be applied to various medical image segmentation networks and can improve the segmentation accuracy of medical images without increasing testing time.

[0020] (2) The loss function used in this invention can be applied to different network structures and is not limited to a certain method. Although it is just a weighting of the loss function, the weights include prior information about the target location, which helps to improve the segmentation effect of the network. Attached Figure Description

[0021] Figure 1 This is a diagram of the medical image segmentation neural network structure according to an embodiment of the present invention;

[0022] Figure 2 This is a flowchart illustrating the training process of a medical image segmentation neural network according to an embodiment of the present invention. Detailed Implementation

[0023] To facilitate understanding and implementation of the present invention by those skilled in the art, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0024] This invention provides a medical image segmentation method based on prior information about the target location, comprising the following steps:

[0025] Step 1: Acquire medical images;

[0026] Step 2: Obtain the segmentation result image using a medical image segmentation neural network;

[0027] Please see Figure 1 The medical image segmentation neural network of this embodiment includes an image downsampling encoder, a feature upsampling encoder, a same-layer fusion module, and a segmentation result generation module;

[0028] The image downsampling encoder consists of five cascaded convolutional modules, each containing two cascaded convolutional layers. The kernel size of each convolutional layer is 3, and the stride is 1. Each convolutional layer is followed by a normalization layer and a corrected linear unit layer.

[0029] The feature upsampling encoder includes four cascaded convolutional modules, each of which contains an upsampling module and a convolutional module. The upsampling module contains, in sequence, a bilinear interpolation upsampling layer, a convolutional layer, a normalization layer, and a corrected linear unit layer. The convolutional layer has a kernel size of 3 and a stride of 1. The convolutional module includes two convolutional layers, each with a kernel size of 3 and a stride of 1. Each convolutional layer is followed by a normalization layer and a corrected linear unit layer.

[0030] The same-layer fusion module is used to concatenate the output features F1, F2, F3, and F4 of the first four stages of the image downsampling encoder with the output features G1, G2, G3, and G4 of each stage of the feature upsampling encoder in the order of G4, G3, G2, and G1.

[0031] The segmentation result generation module is used to generate segmentation results of corresponding sizes based on the features output by each convolutional module of the feature upsampling encoder. This module consists of a convolutional layer with a kernel size of 1 and a stride of 1.

[0032] Please see Figure 2 The medical image segmentation neural network in this embodiment is a pre-trained medical image segmentation neural network; its training process includes the following steps:

[0033] Step 2.1: Obtain several training medical images, label them to obtain training label images, and downsample the training label images three times to obtain label data of four sizes: original image size, 1 / 2, 1 / 4, and 1 / 8.

[0034] Step 2.2: Calculate the distance map based on the training labels;

[0035] In this embodiment, the distance map is calculated based on the training labels. A maximum distance 'a' is set in the distance map. For the background region, the nearest distance from each pixel to the category boundary is calculated. For the foreground region, the nearest distance from each pixel to the category boundary is calculated. Then, this distance is subtracted from 'a'. Finally, the part of the map that is greater than 'a' is set to 'a', and the part that is less than zero is set to zero.

[0036] The distance map calculated based on the training labels in this embodiment is as follows:

[0037]

[0038] Where i represents the current i-th pixel, di represents the value of the current i-th pixel in the distance map, and k represents the category. Let p represent the edge region of the k-th category, j represent a pixel within the edge region, and p represent the edge region of the k-th category. i p represents the position of pixel i. j Ω represents the position of pixel j. background Indicates the background area, Ω foreground The foreground region is represented by 'a', and 'a' represents the maximum distance 'a' in the distance map.

[0039] Step 2.3: Train the medical image segmentation neural network based on the training label images and the distance map calculated based on the training labels. The training ends when the preset number of times is reached (200 in this embodiment). At the end of each training session, the network performance is tested using a validation set, and the network parameters with the best performance are saved as the final network parameters.

[0040] In this embodiment, the loss function used for training the medical image segmentation neural network model is a cross-entropy loss function weighted according to the target location.

[0041] In this embodiment, the cross-entropy loss function weighted according to the target location is:

[0042]

[0043] Where N represents the number of pixels, i represents the current i-th pixel, M represents the number of categories, c represents the current category, and d i y represents the value of the current i-th point in the distance graph. ic Let c and p represent the labels of the current i-th point. ic This represents the probability that the current i-th point belongs to category c.

[0044] The overall loss function in this embodiment is:

[0045] L total =L1+L 1 / 2 +L 1 / 4 +L 1 / 8 ;

[0046] Where L1 represents the cross-entropy loss function that weights the original image size according to the target position, L 1 / 2 Indicate the original Figure 1 A cross-entropy loss function of size / 2, weighted according to the target position, L 1 / 4 Indicate the original Figure 1 A cross-entropy loss function of size / 4, weighted according to the target position, L 1 / 8 Indicate the original Figure 1 / 8-sized cross-entropy loss function weighted according to target location.

[0047] The following experiments further illustrate the invention. This experiment uses the commonly used medical evaluation metrics Dicescore (Dice) and Hausdorff Distance (HD) as evaluation standards. The invention's performance was verified on three commonly used medical image segmentation datasets and three commonly used medical image segmentation methods.

[0048] The datasets used were the ACDC cardiac MRI dataset, the Synapse abdominal multi-organ CT dataset, and the colorectal polyp endoscopic dataset. The colorectal polyp endoscopic dataset included five datasets: Kvasir, CVC-ClinicDB, CVC-ColonDB, ETIS, and Endoscene.

[0049] The medical image segmentation methods used include U-Net, SAUNet, and TransUNet. The experimental results are shown in Tables 1, 2, and 3 below. The three tables represent the results on three different datasets. Taking Table 1 as an example, the U-Net row represents the results of the original U-Net on ACDC, U-Net+Ours represents the results of building the method proposed in the patent on ACDC on U-Net, SAUNet represents the results of the original SAUNet on ACDC, SAUNet baseline represents the results of the network with only the basic part of SAUNet retained on ACDC because SAUNet consists of two parts, and SAUNet baseline+Ours represents the results of building the method proposed in the patent on ACDC on the basic part of SAUNet.

[0050] Table 1

[0051]

[0052] Table 2

[0053]

[0054] Table 3

[0055]

[0056]

[0057] Experimental results show that the segmentation method of the present invention can be applied to various medical image segmentation methods and can significantly improve the segmentation effect of neural networks on medical images.

[0058] It should be understood that the above description of the preferred embodiments is quite detailed, but it should not be considered as a limitation on the scope of protection of this invention. Those skilled in the art, under the guidance of this invention, can make substitutions or modifications without departing from the scope of protection of the claims of this invention, and all such substitutions or modifications fall within the scope of protection of this invention. The scope of protection of this invention should be determined by the appended claims.

Claims

1. A medical image segmentation method based on prior information about the target location, characterized in that, Includes the following steps: Step 1: Acquire medical images; Step 2: Obtain the segmentation result image using a medical image segmentation neural network; The medical image segmentation neural network includes an image downsampling encoder, a feature upsampling encoder, a same-layer fusion module, and a segmentation result generation module; The image downsampling encoder comprises five cascaded convolutional modules, each of which contains two cascaded convolutional layers. The kernel size of the two convolutional layers is 3, and the stride is 1. Each convolutional layer is followed by a normalization layer and a corrected linear unit layer. The feature upsampling encoder includes four cascaded convolutional modules, each of which contains an upsampling module and a convolutional module. The upsampling module contains, in sequence, a bilinear interpolation upsampling layer, a convolutional layer, a normalization layer, and a corrected linear unit layer. The convolutional layer has a kernel size of 3 and a stride of 1. The convolutional module includes two convolutional layers, each with a kernel size of 3 and a stride of 1. Each convolutional layer is followed by a normalization layer and a corrected linear unit layer. The same-layer fusion module is used to concatenate the output features F1, F2, F3, and F4 of the first four stages of the image downsampling encoder with the output features G1, G2, G3, and G4 of each stage of the feature upsampling encoder in the order of G4, G3, G2, and G1. The segmentation result generation module is used to generate segmentation results of corresponding size based on the features output by each convolutional module of the feature upsampling encoder. This module consists of a convolutional layer with a kernel size of 1 and a stride of 1.

2. The medical image segmentation method based on prior information of target location according to claim 1, characterized in that: The medical image segmentation neural network is a trained medical image segmentation neural network; Its training process includes the following steps: Step 2.1: Obtain several training medical images, label them to obtain training label images, and downsample the training label images three times to obtain label data of four sizes: original image size, 1 / 2, 1 / 4, and 1 / 8. Step 2.2: Calculate the distance map based on the training labels; Step 2.3: Train the medical image segmentation neural network based on the training label images and the distance map calculated based on the training labels. The training ends after a preset number of iterations. At the end of each training iteration, use the validation set to test the network performance at that time, and save the network parameters with the best performance as the final network parameters.

3. The medical image segmentation method based on prior information of target location according to claim 2, characterized in that, In step 2.2, the distance map is calculated based on the training labels. This involves setting a maximum distance 'a' in the distance map, calculating the nearest distance from each pixel to the class boundary for the background region, and calculating the nearest distance from each pixel to the class boundary for the foreground region. Then, this distance is subtracted from 'a'. Finally, the portion of the image greater than 'a' is set to 'a', and the portion less than 'a' is set to 'zero'. The distance map calculated based on the training labels is as follows: Where i represents the current i-th pixel, d i This represents the value of the i-th point in the distance map, and j represents a pixel in the edge region. This represents the edge region of the k-th category, where k represents category p. i p represents the position of pixel i. j Ω represents the position of pixel j. background Indicates the background area, Ω foreground The foreground region is represented by 'a', and 'a' represents the maximum distance 'a' in the distance map.

4. The medical image segmentation method based on prior information of target location according to claim 2, characterized in that, In step 2.3, the loss function used for training is the cross-entropy loss function weighted according to the target position; The cross-entropy loss function weighted according to the target position is: Where N represents the number of pixels, i represents the current i-th pixel, M represents the number of categories, c represents the current category, and d i y represents the value of the current i-th point in the distance graph. ic Let c and p represent the labels of the current i-th point. ic This represents the probability that the current i-th point belongs to category c; The overall loss function is: L total =L1+L 1 / 2 +L 1 / 4 +L 1 / 8 ; Where L1 represents the cross-entropy loss function that weights the original image size according to the target position, L 1 / 2 L represents the cross-entropy loss function, which is a half-size representation of the original image, weighted according to the target location. 1 / 4 L represents the cross-entropy loss function, which is a 1 / 4-size representation of the original image, weighted according to the target location. 1 / 8 This represents the cross-entropy loss function, which is weighted according to the target position and represents 1 / 8 of the original image size.

5. A medical image segmentation system based on prior information about the target location, characterized in that, include: One or more processors; A storage device for storing one or more programs, which, when executed by one or more processors, cause the one or more processors to implement the medical image segmentation method based on prior information of target location as described in any one of claims 1 to 4.

Images