Pancreatic focus CT image segmentation method based on CF-HRnet

Abstract

The invention discloses a pancreatic focus CT image segmentation method based on CF-HRnet. According to the method, HRNet with imported pre-training weight is used for training on an NIH Pancreas data set, a coarse-to-fine training framework in an RSTN algorithm is used for training, two steps of preprocessing and joint training are added on the basis of original RSTN coarse segmentation and fine segmentation steps, and the network performance is further improved by increasing learning levels and training batches, so that the network performance is improved. The key optimization point is to introduce a model parameter transfer function and refine the RSTN segmentation process, and more accurately guide the neural network to identify the position information and detail information of the pancreatic focus. According to the method, the number of model parameters is greatly reduced while the focus region segmentation accuracy is improved, and powerful support is provided for diagnosis of pancreatic diseases.

Description

Technical Field

[0001] The present invention relates to the technical field of image segmentation, and in particular to a pancreatic lesion CT image segmentation method based on CF-HRnet. Background Art

[0002] In the field of medical imaging, the study of pancreatic CT image segmentation has always attracted much attention. The rapid and accurate segmentation of pancreatic lesion areas through CT images and the determination of the nature of the disease are of great significance for the early detection and targeted treatment of pancreatic diseases. The recognized difficulties in the segmentation of pancreatic lesion areas are: the pancreas is small in area and complex in morphology. It presents various irregular shapes and structures in CT images and is easily interfered by noise and other tissues; the pancreas has a small difference in density with surrounding organs such as blood vessels and lymph nodes, and the contrast in CT images is low, making it difficult to accurately distinguish the pancreas from surrounding organs or blood vessels; the position of the pancreas is relatively fixed, but its shape and size may vary significantly in different individuals or at different time points. This makes it difficult to design a robust segmentation algorithm.

[0003] The above multiple morphological problems increase the difficulty of accurately segmenting pancreatic lesions from the background of CT images. With the development of image segmentation technology, deep learning methods have been applied in pancreatic segmentation in recent years. The mainstream method is to use convolutional neural networks (CNN) or U-Net to extract the spatial relationship features of the pancreas in a hierarchical mode, so that the neural network can understand, identify and segment the pancreatic lesion area. Of particular concern is that Yu et al. proposed a pancreatic segmentation method based on recursive saliency transform network (RSTN), which decomposes pancreatic segmentation into two stages: coarse segmentation and fine segmentation. The coarse segmentation stage uses low pancreatic pixel resolution images for coarse segmentation and uses Visual Geometry Group (VGG) for feature extraction. The fine segmentation stage uses high pancreatic pixel resolution images for fine segmentation to obtain more accurate pancreatic lesion area segmentation results. However, the fine segmentation results depend to a large extent on the coarse segmentation results. If the coarse segmentation network cannot accurately locate the pancreas, the pancreatic region in the fine segmentation input image will be incomplete or there will be too much background information, thereby reducing the final segmentation accuracy. At the same time, although the feature extraction network VGG used in the above method can learn the position, shape and edge features of the pancreatic organ in CT images and distinguish the pancreas from the surrounding vascular organs, the VGG network is still an older network design. The parameter volume is large after model training, and there are problems such as difficulty in transplantation to downstream applications.

[0004] In view of this, the present invention proposes a pancreatic lesion CT image segmentation method based on CF-HRnet. C is the abbreviation for coarse segmentation, and F is the abbreviation for precise segmentation. The present invention is based on exploring the potential advantages of neural networks in segmenting and locating pancreatic lesions, focusing on improving the High-Resolution Net (HR-Net) network structure and optimizing the RSTN segmentation process to enhance the network's adaptability to changes in pancreatic position and shape, improve the accuracy of pancreatic lesion segmentation and positioning, and obtain more reliable and robust segmentation results in complex medical image scenes. While improving the accuracy of network segmentation, this patent improves the backbone network structure and greatly reduces the backbone network parameters. For CT equipment without a graphics computing unit, the lighter model volume makes it easier for the CT computer to segment the pancreatic area during reconstruction without causing excessive pressure on the memory, and can assist physicians in diagnosis, providing a segmentation basis for downstream tasks such as pancreatic volume calculation. Summary of the invention

[0005] The purpose of the present invention is to provide a pancreatic lesion CT image segmentation method based on CF-HRnet to solve the problems raised in the above background technology.

[0006] To achieve the above object, the present invention provides the following technical solutions:

[0007] A pancreatic lesion CT image segmentation method based on CF-HRnet includes training on the NIH Pancreas dataset using HRNet with imported pre-trained weights, and training using the coarse-to-fine training framework in the RSTN algorithm. The method is characterized in that two steps, preprocessing and joint training, are added on the basis of the original RSTN coarse segmentation and fine segmentation steps, and the network performance is further improved by increasing the learning level and training batches. The specific steps are as follows:

[0008] Step 1: Preprocessing

[0009] First, the pancreatic region is extracted from the original image using the label set, allowing the network to learn the detailed features of the pancreas. Random noise is introduced to address the problem that the object box marked by the label may cause an imbalance in the distribution of target and background pixels, so as to ensure that the number of pixels in the pancreatic lesion area and the number of background pixels always maintain a 1:1 balance. This step is mainly trained using the improved HRNet and initialized with the parameters pre-trained on the VOC dataset, which can be expressed as:

[0010]

[0011]

[0012] in Represents the input image According to the label set Remove the pancreatic area by deducting A noise adding function can add random noise to the image where the pancreatic lesion area is larger than the background area, ensuring that the cross entropy function can effectively supervise the network to learn pancreatic features. is the data set after this processing step, is the neural network HRNet at this stage, are the parameters of the network, For the network The segmentation result of

[0013] Step 2: Rough segmentation

[0014] This step is initialized using the parameters trained by the preprocessing network. Its goal is to let the neural network fit the relationship between the pancreatic image cut out from the original CT image according to the maximum circumscribed rectangle of the original label and the label; this step can be expressed as:

[0015]

[0016]

[0017] in Represents the input image According to the label set Remove the pancreatic area by deducting is the processed data set, is the neural network HRNet for this step. Indicates that the network at this step is the neural network at the previous step The segmentation task of this step is performed under the initialization of weights. represents the parameters obtained by HRNet after learning in the preprocessing step, is the learning parameter of HRNet in this step, For the network The segmentation result of

[0018] Step 3: Fine segmentation

[0019] This step uses a new neural network C-HRNet to learn the pancreas position features of the original CT image. The network is initialized using the parameters of the coarse segmentation step and trained using the original CT pancreas image.

[0020]

[0021]

[0022] in is the unprocessed original CT data set, This is the neural network C-HRNet for this step. Indicates that the neural network in this step is the neural network in the previous step The segmentation task of this step is performed under the initialization of weights. represents the parameters obtained by HRNet after learning in the coarse segmentation step, is the learning parameter of HRNet in this step, For the network The segmentation result of

[0023] Step 4: Joint training

[0024] This step inputs the original CT image without any processing, and uses the segmentation results of the fine segmentation network to locate the pancreatic lesion area from the original CT image. This step sets the adaptive parameters , controls the number of batches of segmentation results that the precise segmentation network uses the coarse segmentation network. This step can be expressed as:

[0025]

[0026]

[0027]

[0028]

[0029] in is the unprocessed original CT data set, where is the input image, is the label set, Indicates that this step applies a coarse segmentation neural network The pancreas localization and segmentation task is performed under the weight initialization. represents the fine segmentation result, According to the fine segmentation results The pancreatic region of the original input CT image is subtracted, and the subtraction result is , Indicates that this step applies the fine segmentation network Weight Initialization of the pancreas region removed To split, Represents the final segmentation result.

[0030] Furthermore, in step 4: joint training, multiple loss functions are used for joint supervision, and the IOU loss function is used to penalize areas where the segmentation results do not overlap with the labels. The IOU loss function is expressed as:

[0031]

[0032] in is the result of network prediction, For labels;

[0033] At the same time, the cross entropy loss function is used to supervise the learning of the network. The cross entropy loss function is expressed as:

[0034]

[0035] in is the result of network prediction, For labels;

[0036] Finally, a new loss function is defined to supervise the boundary area segmentation results. The prediction results and labels are Gaussian blurred at the same time, and then the loss is calculated. The Gaussian blurring process can be expressed as:

[0037]

[0038] in represents a two-dimensional convolution operation, To output the image, the Gaussian kernel function It is expressed as:

[0039]

[0040] in and represents the coordinates of the center point of the convolution kernel, is the standard deviation of the Gaussian kernel;

[0041] Forecast results and input image At the same time, Gaussian blur processing can be performed to obtain

[0042]

[0043]

[0044] in The input image is passed through the Gaussian kernel The blurred result is The input image is passed through the Gaussian kernel The result after blurring,

[0045] The new Gaussian loss function is expressed as:

[0046]

[0047] in is the result of network prediction, For labels;

[0048] The final total loss function is:

[0049] .

[0050] Compared with the prior art, the present invention has the following beneficial effects:

[0051] (1) More accurate pancreatic lesion segmentation and positioning: Through the improvement of the HRNet network, the characteristics of the high-resolution channel are successfully retained, the location information of pancreatic pixels is captured to the greatest extent, and the accuracy of CT pancreatic lesion segmentation is effectively improved.

[0052] (2) Lightweight design adapted to devices with limited memory: The improved HRNet does not have an overly deep downsampling part, which reduces the number of parameters and makes it more suitable for medical devices with limited memory. This lightweight design not only helps the effective convergence of the model, but also reduces the requirements for hardware resources and improves the usability of the algorithm in actual medical devices.

[0053] (3) Breaking through the limitations of traditional network design: While abandoning the limitations of the traditional VGG network, the performance and generalization ability of the HRNet network are improved, which provides greater flexibility for deploying this technology on different medical devices and platforms.

[0054] (4) Expanding the learning level of the neural network: By adding preprocessing steps and increasing the network learning level, significant results have been achieved in improving network performance. When processing medical images, the method of the present invention can capture and express the features in the image more finely, improving the accuracy and stability of segmentation.

[0055] Combining these beneficial effects, it can be concluded that the present invention provides an advanced and efficient pancreatic lesion segmentation method for the field of medical imaging, which will become a reliable image analysis tool for doctors to diagnose pancreatic diseases and is expected to improve the accuracy of early diagnosis of pancreatic diseases. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 This is a schematic diagram of the improved HRNet network structure of the present invention;

[0057] Figure 2 It is the CF-HRNet segmentation process of the present invention;

[0058] Figure 3 This is a comparison chart of the results of five CT pancreatic image segmentation methods of the present invention. DETAILED DESCRIPTION

[0059] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0060] The implementation steps and key algorithms of the technical solution of the present invention are:

[0061] (1) Use the improved HRNet to replace the original VGG network.

[0062] Improve the structure of the feature extraction network HRNet. The feature map resolution of the highest resolution channel of HRNet is doubled to obtain a larger intermediate feature map so that the pancreas position information will not be lost during the downsampling process. In order to reduce the computational overhead of the convolutional layer caused by the enlarged feature map resolution, the overall architecture of the HRNet network is redesigned to achieve lightweight model performance.

[0063] The core structure of HRNet is a multi-branch parallel network, each branch is responsible for processing feature maps of different resolutions. Assume that HRNet contains residual network branches, each branch can be expressed as ,in . Then the deep residual network branch can be expressed as:

[0064]

[0065] in Indicates branches, represents the convolution operation, Represents a residual block, which contains multiple convolutional layers and batch normalization layers.

[0066] The high-resolution feature fusion module for fusing different branches can be expressed as:

[0067]

[0068] in Indicates branches, Indicates concatenating all upsampled feature maps. Indicates that the resolution is The feature map is upsampled to match the highest resolution feature map.

[0069] Feature pyramid network for fusing and upsampling feature maps to the original input image size It can be expressed as:

[0070]

[0071] in Indicates branches. Indicates that the resolution is The feature map is upsampled to match the highest resolution feature map.

[0072] The improved HRNet network structure is as follows: Figure 1 shown.

[0073] The present invention uses HRNet with imported pre-trained weights to train on the NIH Pancreas dataset. The coarse-to-fine training framework in the RSTN algorithm is used for training, and the control variable method is used to evaluate whether HRNet is effective in the pancreatic lesion segmentation task, and the optimal HRNet network structure suitable for pancreatic segmentation is obtained through comparative experiments. The experimental results are shown in Table 1. The horizontal row of Table 1 is the abbreviation of the HRNet network structure, and the numbers represent the number of network modules in the first stage-the number of network modules in the second stage-the number of network modules in the third stage. The vertical row represents the maximum resolution of the network feature map.

[0074] Table 1 Segmentation performance of HRNet with different network structures on the NIH pancreas dataset. Different columns represent different network structures of HRNet, and different rows represent the size of the maximum feature map.

[0075]

[0076] The above experimental results show the superiority of HRNet over VGG network. By improving the adaptability of HRNet and performing system tuning on the pancreatic dataset, we successfully reduced the number of parameters and significantly improved the segmentation effect.

[0077] (2) Use the improved HRNet to replace the original VGG network.

[0078] Based on the original RSTN rough segmentation and fine segmentation steps, the present invention adds two steps of preprocessing and joint training, and further improves the network performance by increasing the learning level and training batch. The key optimization point is to introduce the model parameter transfer function and further refine the RSTN segmentation process, so as to more accurately guide the neural network to identify the location information and detailed information of pancreatic lesions.

[0079] In view of the problem that the inaccurate positioning of the coarse segmentation network in the traditional RSTN method will affect the fine segmentation effect, the present invention sets four segmentation steps to enable the network to better learn the location and morphological characteristics of pancreatic lesions, and achieve stable training and effective convergence. Different from the RSTN method, the method of the present invention first allows the network to learn the detailed features of the pancreas rather than the location features. Then, through the knowledge transfer method, the network is gradually generalized and learns more complex location relationship features.

[0080] Step 1: Preprocessing

[0081] First, the pancreatic region is extracted from the original image using the label set, allowing the network to learn the detailed features of the pancreas. Random noise is introduced to address the problem that the object box marked by the label may cause an imbalance in the distribution of target and background pixels, so as to ensure that the number of pixels in the pancreatic lesion area and the number of background pixels always maintain a 1:1 balance. This step mainly uses the improved HRNet for training and is initialized with parameters pre-trained on the VOC dataset. This step can be expressed as:

[0082]

[0083]

[0084] in Represents the input image According to the label set Remove the pancreatic region by deducting it. A noise adding function can be used to add random noise to the images where the pancreatic lesion area is larger than the background area, ensuring that the cross entropy function can effectively supervise the network to learn pancreatic features. is the data set after this processing step. is the neural network HRNet at this stage, are the parameters of the network. For the network The segmentation result.

[0085] Step 2: Rough segmentation

[0086] This step is initialized using the parameters trained by the preprocessing network. The goal is to let the neural network fit the relationship between the pancreatic image and the label extracted from the original CT image according to the maximum enclosing rectangle of the original label. Since it is necessary to control the pixel ratio of the pancreatic area to the background area in the original image to be 1:1, this step is slightly more difficult than the preprocessing step. This step can be expressed as:

[0087]

[0088]

[0089] in Represents the input image According to the label set Remove the pancreatic region by deducting it. The processed data set. is the neural network HRNet for this step. Indicates that the network at this step is the neural network at the previous step The segmentation task of this step is performed under the initialization of weights. represents the parameters obtained by HRNet after learning in the preprocessing step, is the learning parameter of HRNet in this step. For the network The segmentation result.

[0090] Step 3: Fine Segmentation

[0091] This step uses a new neural network C-HRNet to learn the pancreas position features of the original CT image.

[0092] The network is initialized with the parameters of the step and trained using the original CT pancreatic images.

[0093]

[0094]

[0095] in It is the unprocessed original CT data set. This is the neural network C-HRNet for this step. Indicates that the neural network in this step is the neural network in the previous step The segmentation task of this step is performed under the initialization of weights. represents the parameters obtained by HRNet after learning in the coarse segmentation step, is the learning parameter of HRNet in this step. For the network The segmentation result.

[0096] Step 4: Joint training

[0097] This step takes the original CT image without any processing as input and uses the segmentation result of the fine segmentation network to locate the pancreatic lesion area from the original CT image. This step sets the adaptive parameters , controls the number of batches of segmentation results that the precise segmentation network uses for the coarse segmentation network. This setting is to deal with the low accuracy results that the coarse segmentation network may produce at the beginning of training, and avoid fluctuations in the previously trained precise segmentation network. Then, the coarse segmentation and fine segmentation are jointly trained to achieve a more stable and effective model optimization. This step can be expressed as:

[0098]

[0099]

[0100]

[0101]

[0102] in is the unprocessed original CT data set, where is the input image, For label sets. Indicates that this step applies a coarse segmentation neural network The pancreas localization and segmentation task is performed under the weight initialization. Represents the refined segmentation result. According to the fine segmentation results The pancreatic region of the original input CT image is subtracted, and the subtraction result is . Indicates that this step applies the fine segmentation network Weight Initialization of the pancreas region removed to split. Represents the final segmentation result.

[0103] The present invention uses multiple loss functions for joint supervision and uses the IOU (intersection-over-union) loss function to penalize areas where the segmentation results do not overlap with the labels. The IOU loss function is expressed as:

[0104]

[0105] in is the result of network prediction, For label.

[0106] At the same time, the cross entropy loss function is used to supervise the learning of the network. The cross entropy loss function is expressed as:

[0107]

[0108] in is the result of network prediction, For label.

[0109] Finally, considering the fact that the pancreas has unclear boundaries and is easily mis-segmented, a new loss function is defined to supervise the boundary area segmentation results. Gaussian blur processing is performed on both the prediction results and the labels, and then the loss is calculated. The Gaussian blur processing process can be expressed as:

[0110]

[0111] in represents a two-dimensional convolution operation, To output the image, the Gaussian kernel function It is expressed as:

[0112]

[0113] in and represents the coordinates of the center point of the convolution kernel, is the standard deviation of the Gaussian kernel.

[0114] Forecast results and input image At the same time, Gaussian blur processing can be performed to obtain

[0115]

[0116]

[0117] in The input image is passed through the Gaussian kernel The blurred result is The input image is passed through the Gaussian kernel The result after blurring,

[0118] The new Gaussian loss function is expressed as:

[0119]

[0120] in is the result of network prediction, For label.

[0121] The final total loss function is:

[0122]

[0123] The optimized network segmentation process is as follows: Figure 2 shown.

[0124] Comparison of experimental data with evaluation results

[0125] The effectiveness of the method of the present invention is verified by comparing it with four other classic algorithms for pancreas segmentation. The four comparison algorithms include DeepOrgan algorithm, RSTN algorithm, U-Net algorithm, and Attention U-Net algorithm. The Dice coefficient is used to evaluate the segmentation performance. The maximum value, minimum value, and standard deviation are calculated by statistically analyzing the Dice coefficient of the segmentation results. Our test results on the NIHPancreas dataset are shown in Table 2.

[0126] Table 2: Comparison of test results of NIH Pancreas dataset

[0127]

[0128] The results of comparative experiments show that the present invention successfully solves the limitations of large parameters and difficult transplantation of pancreatic lesion segmentation models by using an improved High Resolution Net (HRNet) to replace the traditional VGG network. Since the improved HRNet retains the high-resolution channel, the position information of pancreatic pixels can be retained to the greatest extent during the downsampling process, thereby ensuring the accuracy of pancreatic lesion segmentation and positioning. In addition, the improved HRNet has fewer parameters, and according to the optimized segmentation process, the method of the present invention achieves better results in pancreatic lesion segmentation.

[0129] The PASCAL VOC dataset is used for model pre-training. This dataset is a classic dataset in the field of object detection, organized and released by the PASCAL (Pattern Analysis, Statistical Modelling and Computational Learning) visual challenge. The dataset contains 20 object categories and a total of about 11,000 images, each of which is annotated with the category and location of the object. The PASCAL VOC dataset is divided into versions from 2007 to 2012 according to the year, and each version is divided into a training set, a validation set, and a test set. The training set and the validation set are used to train and evaluate the object detection model, and the test set is used to evaluate the performance of the model on unknown data. In order to increase the amount of training set data, we use all datasets from 2007 to 2012 in pre-training, and expand the amount of training set data through data augmentation methods such as rotation and random subtraction. In the pre-training process, we use a variety of training techniques. The first is to freeze the decoding layer training, that is, freeze the parameter update of the decoding layer in the first 50 batches of training. After training 50 batches, the encoding layer has the initial feature extraction capability. After that, the decoding layer is allowed to update parameters, and the entire network completes training under the condition that parameter updates are fully allowed. Secondly, we set a parameter weight decay method, that is, when the training loss value exceeds 10 training batches and decays less than the preset value, training is stopped to prevent overfitting.

[0130] After HRNet was pre-trained on the Augmented VOC dataset, we deployed HRNet in the improved coarse-fine segmentation framework and initialized the network according to the initialization method described in the detailed description of the technical solution. We then began to fit the improved HRNet on the NIH pancreas dataset.

[0131] The NIH Pancreas dataset is a commonly used dataset for pancreatic image segmentation tasks, provided by the National Institutes of Health (NIH). The dataset contains abdominal CT scan images of 82 patients, 53 of whom are healthy subjects and 29 are patients with pancreatic diseases. All images are anonymized and annotated for pancreatic segmentation by experienced radiologists. Since the CT data of each patient is a three-dimensional stereoscopic CT image set, we divide the original CT image into three planes, X, Y and Z, according to the sagittal, coronal and transverse planes in medicine, and perform segmentation on the X, Y and Z planes respectively. The segmented image uses a voting mechanism to splice the segmentation results on the three planes, and finally compares them with the Ground Truth in the label set, and calculates the Dice coefficients on the X, Y, and Z planes and the Dice coefficient after splicing as a whole as the criterion for segmentation performance. The Dice coefficient is calculated as follows:

[0132] All experiments were run on a server with Linux Ubuntu system. The GPU model is NVIDIARTX 3090 (24GB) and the CPU model is AMD Ryzen Threadripper 5990X. All experiments were distributed and run on three GPUs.

[0133] It is obvious to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential features of the present invention. Therefore, the embodiments should be regarded as exemplary and non-limiting from any point of view, and the scope of the present invention is defined by the appended claims rather than the above description, and it is intended that all changes falling within the meaning and scope of the equivalent elements of the claims are included in the present invention, and any reference numerals in the claims should not be regarded as limiting the claims involved.

[0134] In addition, it should be understood that although the present specification is described according to implementation modes, not every implementation mode contains only one independent technical solution. This description of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment may also be appropriately combined to form other implementation modes that can be understood by those skilled in the art.

Claims

1. A pancreatic lesion CT image segmentation method based on CF-HRnet, comprising training on the NIH Pancreas dataset using HRNet with imported pre-trained weights, and training using the coarse-to-fine training framework in the RSTN algorithm, characterized in that: Based on the original RSTN coarse segmentation and fine segmentation steps, two steps of preprocessing and joint training are added. By increasing the learning level and training batches, the network performance is further improved. The specific steps are as follows: Step 1: Preprocessing First, the pancreatic region is extracted from the original image using the label set, allowing the network to learn the detailed features of the pancreas. Random noise is introduced to address the problem that the object box marked by the label may cause an imbalance in the distribution of target and background pixels, so as to ensure that the number of pixels in the pancreatic lesion area and the number of background pixels always maintain a 1:1 balance. This step is mainly trained using the improved HRNet and initialized with the parameters pre-trained on the VOC dataset, which can be expressed as: in Represents the input image According to the label set Remove the pancreatic area by deducting A noise adding function can add random noise to the image where the pancreatic lesion area is larger than the background area, ensuring that the cross entropy function can effectively supervise the network to learn pancreatic features. is the data set after this processing step, is the neural network HRNet at this stage, are the parameters of the network, For the network The segmentation result of Step 2: Rough segmentation This step is initialized using the parameters trained by the preprocessing network. Its goal is to let the neural network fit the relationship between the pancreatic image cut out from the original CT image according to the maximum circumscribed rectangle of the original label and the label; this step can be expressed as: in Represents the input image According to the label set Remove the pancreatic area by deducting is the processed data set, HRNet is the neural network of this step; Indicates that the network at this step is the neural network at the previous step The segmentation task of this step is performed under the initialization of weights. represents the parameters obtained by HRNet after learning in the preprocessing step, is the learning parameter of HRNet in this step, For the network The segmentation result of Step 3: Fine Segmentation This step uses a new neural network C-HRNet to learn the pancreas position features of the original CT image. The network is initialized using the parameters of the coarse segmentation step and trained using the original CT pancreas image. in is the unprocessed original CT data set, The neural network C-HRNet for this step; Indicates that the neural network in this step is the neural network in the previous step The segmentation task of this step is performed under the initialization of weights. represents the parameters obtained by HRNet after learning in the coarse segmentation step, is the learning parameter of HRNet in this step, For the network The segmentation result of Step 4: Joint training This step inputs the original CT image without any processing, and uses the segmentation results of the fine segmentation network to locate the pancreatic lesion area from the original CT image. This step sets the adaptive parameters , controls the number of batches of segmentation results that the precise segmentation network uses the coarse segmentation network. This step can be expressed as: in is the unprocessed original CT data set, where is the input image, is the label set, Indicates that this step applies a coarse segmentation neural network The pancreas localization and segmentation task is performed under the weight initialization. represents the fine segmentation result, According to the fine segmentation results The pancreatic region of the original input CT image is subtracted, and the subtraction result is , Indicates that this step applies the fine segmentation network Weight Initialization of the pancreas region removed To split, Represents the final segmentation result.

2. The pancreatic lesion CT image segmentation method based on CF-HRnet according to claim 1, characterized in that: In step 4, multiple loss functions are used for joint supervision in the joint training phase, and the IOU loss function is used to penalize the areas where the segmentation results do not overlap with the labels. The IOU loss function is expressed as: in is the result of network prediction, For labels; At the same time, the cross entropy loss function is used to supervise the learning of the network. The cross entropy loss function is expressed as: in is the result of network prediction, For labels; Finally, a new loss function is defined to supervise the boundary area segmentation results. The prediction results and labels are Gaussian blurred at the same time, and then the loss is calculated. The Gaussian blurring process can be expressed as: in represents a two-dimensional convolution operation, To output the image, the Gaussian kernel function It is expressed as: in and represents the coordinates of the center point of the convolution kernel, is the standard deviation of the Gaussian kernel; Forecast results and input image At the same time, Gaussian blur processing can be performed to obtain in The input image is passed through the Gaussian kernel The blurred result is The input image is passed through the Gaussian kernel The result after blurring, The new Gaussian loss function is expressed as: in is the result of network prediction, For labels; The final total loss function is: 。

Images