A method and system for measuring the distribution of pancreatic fat and iron deposition by magnetic resonance

Abstract

The present invention discloses a method and system for measuring the distribution of pancreatic fat and iron deposition by magnetic resonance. Multi-echo images of the abdomen are acquired by a magnetic resonance instrument; in-phase and opposed-phase dual-modal images are extracted and input into a neural network model to obtain a pancreatic segmentation mask; a mapping method is used to process and obtain information on the distribution of pancreatic fat deposition and iron deposition; the pancreas is divided into regions according to the pancreatic segmentation mask to obtain each pancreatic sub-region, and the distribution of pancreatic fat deposition and iron deposition in each pancreatic sub-region is obtained. The present invention can effectively remove the interference of signals from other parts of the abdomen, has strong adaptability to different disease populations and magnetic resonance equipment from different manufacturers, ensures the accuracy of pancreatic segmentation, realizes the visualization of the distribution of pancreatic fat deposition and iron deposition, and has high accuracy.

Description

Technical Field

[0001] The present application relates to a measurement method and system in the field of biomedical engineering technology, and particularly to a method and system for measuring the distribution of pancreatic fat and iron deposition applicable to multi-population and multi-center data. Background Art

[0002] Type 2 diabetes mellitus (T2DM) is a chronic heterogeneous disease caused by a complex set of metabolic disorders. As of 2003, worldwide, the number of diabetes patients was expected to rise to 439 million, more than 90% of which were T2DM. Compared with visceral adipose tissue, the accumulation of ectopic fat in normal tissues is considered to be a better predictor of the risk of T2DM.

[0003] The pancreas, as a digestive and endocrine organ, is closely associated with T2DM. However, compared with the liver, the pathological process and subsequent development of pancreatic fat deposition are often overlooked. Studies have shown that the accumulation of ectopic fat in the pancreas may cause or exacerbate the progression of acute pancreatitis, β-cell dysfunction, T2DM, and is even associated with pancreatic tumors. At the same time, pancreatic inflammation can lead to edema and fat deposition in the pancreatic parenchyma, and when combined with bleeding, local iron overload may occur due to the deposition of ferritin and hemosiderin, and long-term iron deposition may cause damage to tissues. Pancreatic iron deposition is also considered to have a certain impact on the function of pancreatic β-cells and is related to impaired glucose metabolism. At the same time, the distribution of cell clusters in the pancreas is uneven, for example, the islets of Langerhans are the most numerous in the pancreatic tail. Therefore, measuring the distribution of ectopic fat deposition and iron deposition in different parts of the pancreas can enhance the understanding of pancreatic function abnormalities / metabolic abnormalities diseases.

[0004] Since it is difficult to obtain a biopsy of the pancreas, in vivo quantitative techniques are generally used, and the chemical shift dual echo / multi-echo technique in magnetic resonance imaging (Magnetic resonance imaging, MRI) is often used as the best quantitative method. The T1 dual echo technique is an indispensable routine sequence for abdominal liver imaging. Based on the principle of chemical shift, through a dual echo acquisition with a single breath-hold, perfectly registered in-phase / opposite-phase images of water and fat can be obtained. Dual echo DIXON is a continuation of this technique, and in the post-processing stage, water-fat separation imaging is calculated through in-phase / opposite-phase images and phase information, and finally a fat fraction map of the entire abdomen is obtained. Currently, a multi-echo technique has also been developed. Through multi-point acquisition, water-fat separation is more complete, and fat and iron deposition can be accurately quantified.

[0005] Due to the variable shape, blurred edges of the pancreas, the presence of blood vessels and bile ducts passing through it, and the close proximity of the surrounding liver and duodenum, it is difficult for doctors to manually outline the entire pancreas in magnetic resonance images. Although using deep learning technology for magnetic resonance pancreas segmentation is effective and can achieve standardization, due to the difficulty of pancreas segmentation, the Dice Similarity Coefficient (DSC) of the test set for pancreas segmentation in current papers is rarely higher than 0.85. At the same time, since all the published papers and methods do not mention the segmentation problem of multi-center datasets, the default model is only applicable to datasets of a single population and a single device, and has certain requirements for the setting of sequence parameters, increasing the operational complexity for technicians and resulting in a small scope of model application. Summary of the Invention

[0006] Aiming at the defects in the prior art, the purpose of the present invention is to provide a method for measuring the distribution of pancreatic fat and iron deposition applicable to multi-population and multi-center data. For the purpose of accurately measuring the distribution of pancreatic fat and iron deposition in multiple regions, a dual-modal 3D integrated network model is established to segment the abdominal multi-echo dataset, and then a mapping method is used to obtain the distribution of pancreatic fat deposition and iron deposition.

[0007] The segmentation model of the present invention does not limit the equipment and acquisition parameters, greatly expanding the application scenarios.

[0008] As Figure 1 shown, the technical solution adopted by the present invention is as follows:

[0009] 1. A method for measuring the distribution of pancreatic fat and iron deposition, comprising:

[0010] S1. Collect multi-echo images of the abdomen through a magnetic resonance instrument;

[0011] S2. Extract in-phase and opposed-phase dual-modal images from the multi-echo images, and input the dual-modal images into a neural network model to obtain a pancreas segmentation mask;

[0012] S3. Process the pancreas segmentation mask using a mapping method to obtain information on the distribution of pancreatic fat deposition and iron deposition;

[0013] S4. Divide the pancreas regions according to the pancreas segmentation mask to obtain each pancreatic sub-region, and obtain the distribution of pancreatic fat deposition and iron deposition in each pancreatic sub-region.

[0014] The specific pancreatic sub-regions refer to four regions of the pancreas, namely the four pancreatic sub-regions of the pancreatic head, pancreatic neck, pancreatic body, and pancreatic tail.

[0015] The specific steps of S2 are as follows:

[0016] Select a magnetic resonance multi-echo sequence to acquire magnetic resonance in-phase images and magnetic resonance opposed-phase images of the human abdomen as dual-modal images. Obtain magnetic resonance in-phase / opposed-phase / water / fat / proton density fat fraction / iron deposition R2* images of the human abdomen through a single acquisition with the multi-echo sequence. Among them, the in-phase image and the opposed-phase image are used as dual-modal images to be input into the neural network model for training, and a pancreatic segmentation mask is obtained through the neural network model. The proton density fat fraction image and the iron deposition R2* image are used for mapping values and calculating the pancreatic fat distribution and iron deposition distribution.

[0017] The neural network model is pre-trained using a training set. During training, data augmentation is performed on each image in the training set. The data augmentation includes three operations: random brightness enhancement and weakening, linear contrast stretching, and non-linear contrast stretching, which are performed in sequence. Each operation is set with a probability of 15%. Additionally, the adjustment range of random brightness enhancement and weakening, the range of linear contrast stretching, and the range of non-linear contrast stretching are increased.

[0018] The neural network model is mainly formed by fusing a first 3D high-resolution model and a 3D cascade model. The 3D cascade model is formed by sequentially connecting a 3D low-resolution model and a second 3D high-resolution model. The bimodal image is input into the first 3D high-resolution model after data normalization and high-resolution resampling, and then the first pancreatic segmentation probability output result is obtained. At the same time, the bimodal image is input into the 3D cascade model after data normalization and low-resolution resampling. First, it is input into the 3D low-resolution model to obtain a preliminary pancreatic segmentation mask with a lower resolution. The preliminary pancreatic segmentation mask is used as the third modality together with the bimodal image after high-resolution resampling and input into the second 3D high-resolution model in the 3D cascade model, and then the second pancreatic segmentation probability output result is obtained. The first pancreatic segmentation probability output result and the second pancreatic segmentation probability output result are fused by taking the mean or voting method, and then binarized to obtain the final pancreatic segmentation mask. The structures of the first 3D high-resolution model, the 3D low-resolution model, and the second 3D high-resolution model are the same, and each includes eleven convolutional units. Five of the convolutional units are sequentially connected to form an input convolutional part, and a downsampling operation is set between adjacent convolutional units to connect to the other five convolutional units that are sequentially connected to form an output convolutional part. An anti-convolution operation is set between adjacent convolutional units. The output result of the i-th convolutional unit in the input convolutional part and the result after anti-convolution of the output of the (5 - i)-th convolutional unit in the input convolutional part are connected and then jointly input into the (6 - i)-th convolutional unit in the input convolutional part, where i = 1 - 4. The fifth convolutional unit in the input convolutional part is connected to the remaining one convolutional unit through downsampling. The output result of the fifth convolutional unit in the input convolutional part and the result after anti-convolution of the output of the remaining one convolutional unit are connected and then input into the first convolutional unit in the output convolutional part. The input of the first convolutional unit at the beginning of the input convolutional part is used as the input of the model, and the output of the fifth convolutional unit at the end of the output convolutional part is used as the output of the model. Each convolutional unit is mainly formed by sequentially connecting two convolutional modules, and each convolutional module is mainly formed by sequentially connecting a convolution operation, a normalization operation, and an activation function.

[0019] In specific implementation, the neural network model can be replaced by its own 3D cascade model. The 3D cascade model alone as an overall model can also obtain a pancreatic segmentation mask with good results.

[0020] Specifically, S3 is: mapping the pancreatic segmentation mask to the proton density fat fraction image and the iron deposition R2* image of the multi-echo image.

[0021] Specifically, S4 includes the following steps:

[0022] Obtain the skeleton of the pancreatic segmentation mask through morphological processing.

[0023] Calculate the full length of the pancreatic skeleton, calculate the ranges of the pancreatic head, neck, body, and tail according to a preset ratio, and partition the pancreas to obtain each pancreatic sub-region;

[0024] Calculate the pancreatic fat fraction and iron deposition content of each pancreatic sub-region according to the pancreatic fat deposition distribution information and iron deposition distribution information.

[0025] The present invention designs a specific neural network model applied to bimodal image objects, which can simultaneously detect the distribution of pancreatic fat and iron deposition. The present invention uses bimodal images for segmentation, extracts complementary information from different modes, and improves the network accuracy and operation efficiency.

[0026] The present invention is applicable to the determination and processing of the distribution of pancreatic fat and iron deposition in multi-population and multi-center data.

[0027] Multi-population means that it can process magnetic resonance images of different age groups and different pathological populations. Usually, due to the different morphology, position, water content, and fat content of the pancreas in different age groups and different pathological populations, the existing determination of pancreatic fat deposition distribution and iron deposition distribution can only be carried out under the condition of a fixed age group and similar pathological populations, and the determination methods across age groups and across pathological populations are difficult to apply. However, the present invention can be applicable to multi-population and solves this problem.

[0028] Multi-center data means that it can process magnetic resonance images collected by magnetic resonance instruments of different manufacturers. Usually, due to the fact that on machines of different manufacturers, the selection of parameters, pulse shape, and reconstruction algorithm may change the contrast of the generated images, thus affecting the model prediction effect, the existing determination of pancreatic fat deposition distribution and iron deposition distribution can only be carried out under the condition of images collected by instruments of the same fixed manufacturer, and the determination methods across manufacturers' instruments are difficult to apply. However, the present invention can be applicable to multi-center data and solves this problem.

[0029] Second, a system for determining the distribution of pancreatic fat and iron deposition by magnetic resonance, the system comprising:

[0030] An image acquisition module for acquiring magnetic resonance multi-echo images.

[0031] A pancreatic prediction module for obtaining a pancreatic segmentation mask by segmenting the multi-echo image;

[0032] A mapping module for obtaining pancreatic fat deposition distribution information and iron deposition distribution information according to the pancreatic segmentation mask.

[0033] A sub-region parameter calculation module for obtaining the pancreatic fat fraction and iron deposition content of the pancreatic sub-region by processing the distribution information and the pancreatic segmentation mask.

[0034] III. An electronic device:

[0035] It includes a memory, a processor, and computer instructions stored in the memory and running on the processor. When the computer instructions are run by the processor, the steps in the method according to any one of claims 1 - 6 are completed.

[0036] In the present invention, multi - echo images of the abdomen are collected by a magnetic resonance instrument; dual - modality images of in - phase and opposed - phase are extracted and input into a neural network model to obtain a pancreatic segmentation mask; proton density fat fraction images and iron deposition R2* images are extracted, and a mapping method is used to obtain information on pancreatic fat deposition distribution and iron deposition distribution in the pancreas; the pancreas is divided into regions according to the pancreatic segmentation mask, and the pancreatic fat deposition distribution and iron deposition distribution of each pancreatic sub - region are obtained.

[0037] The system of the present invention can effectively remove the interference of abdominal fat, splenic vein, inferior mesenteric artery / vein, and duodenum by adopting a dual - modality 3D integration network. It has strong adaptability to different disease populations and magnetic resonance devices from different manufacturers, ensures the accuracy of pancreatic segmentation, is used for visualizing pancreatic fat deposition distribution and iron deposition distribution, calculating multi - region pancreatic proton density fat fraction and iron deposition content, with high accuracy, and provides help for the horizontal and vertical comparative studies of metabolic diseases.

[0038] The present invention has the following advantages compared with the prior art:

[0039] (1) The segmentation model of the present invention has excellent segmentation accuracy, and the average DSC of the test set can reach 0.8895:

[0040] Dual - modality input: Since the magnetic resonance dual - echo / multi - echo acquisition technology can be carried out within a single breath - hold, perfect registration is theoretically achieved. The dual - modality input network structure adopted in this study can introduce additional contrast information and obtain better segmentation performance.

[0041] 3D integration model and multi - center prediction: It combines the advantages of 3D cascade network model and 3D high - resolution network model, and obtains better prediction performance than using 3D cascade network and 3D high - resolution network alone.

[0042] (2) The present invention is applicable to multi - population and multi - center dataset situations and has high application value:

[0043] Application value for multi - population: People of different age groups (teenagers, middle - aged people, elderly people, etc.), as well as different patient groups (obesity patients, diabetes patients, pancreatic disease patients, etc.) and normal people can use the method of the present invention.

[0044] Multi - center application value: Since the 3D integrated model (and the 3D cascaded network) uses a wide range of random contrast enhancement modes in the data enhancement step, it has a robust performance on datasets from different centers and different devices.

[0045] (3) The present invention can visualize and partition - measure the distribution of pancreatic fat deposition and iron deposition, improving the efficiency and value of data analysis:

[0046] 3D visualization: Through the visualization performance, doctors can clearly and intuitively observe the distribution of pancreatic fat and iron deposition in different parts of the subject by rotating the 3D interactive interface.

[0047] Partition - area measurement: Through morphological calculation and automatic partitioning according to proportion, the distribution of pancreatic fat deposition and iron deposition in different regions of the pancreas can be obtained, which can be used as a prediction index for type 2 diabetes risk. Brief Description of the Drawings

[0048] Figure 1 It is a flowchart of the method according to an embodiment of the present invention.

[0049] Figure 2 It is a schematic diagram of the 3D integrated segmentation network model according to an embodiment of the present invention.

[0050] Figure 3 It is a 3D schematic diagram of the distribution of pancreatic fat deposition and iron deposition of an obese patient according to an embodiment of the present invention. Among them, Figure 3 (a) is the distribution diagram of pancreatic fat deposition; Figure 3 (b) is the distribution diagram of pancreatic iron deposition. Detailed Embodiments

[0051] The present invention will be further described below in conjunction with the drawings and embodiments. The following embodiments will help those skilled in the art to further understand the present invention, but do not limit the present invention in any form. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present invention, several deformations and improvements can still be made. These all belong to the protection scope of the present invention.

[0052] Embodiments of the present invention are as follows:

[0053] As Figure 1 shown, first, obtain the abdominal magnetic resonance multi - echo data of the subject, extract the in - phase / out - of - phase images and input them into the neural network model as Figure 2 shown.

[0054] S1. Collect multi - echo images of the abdomen through a magnetic resonance instrument;

[0055] S2. Extract in-phase and opposed-phase bimodal images from multi-echo images, and input the bimodal images into a neural network model to obtain a pancreatic segmentation mask;

[0056] Single acquisition of a multi-echo sequence can obtain in-phase / opposed-phase / water / fat / proton density fat fraction / iron deposition R2* images of the human abdomen. The in-phase image and the opposed-phase image among them are used as bimodal images to input into the neural network model for training. After the neural network model obtains the pancreatic segmentation mask, the proton density fat fraction image and the iron deposition R2* image are used to map values and calculate the pancreatic fat distribution and iron deposition distribution.

[0057] As Figure 2 shown, the neural network model is mainly composed of the fusion of a first 3D high-resolution model and a 3D cascade model. The 3D cascade model is sequentially connected by a 3D low-resolution model and a second 3D high-resolution model;

[0058] The fusion steps are as follows: Input the bimodal images into the first 3D high-resolution model after data normalization and high-resolution resampling, and then output the first pancreatic segmentation probability output result; at the same time, input the bimodal images into the 3D cascade model after data normalization and low-resolution resampling. First, input them into the 3D low-resolution model to obtain a preliminary pancreatic segmentation mask with a lower resolution. After high-resolution resampling of the preliminary pancreatic segmentation mask, it is used as a third modality and input into the second 3D high-resolution model in the 3D cascade model together with the bimodal images, and then output the second pancreatic segmentation probability output result; The first pancreatic segmentation probability output result and the second pancreatic segmentation probability output result are fused by taking the mean or voting method, and after fusion, binaryzation is performed to obtain the final pancreatic segmentation mask;

[0059] The structures of the first 3D high-resolution model, 3D low-resolution model, and second 3D high-resolution model are the same, and each includes eleven convolutional units. Five of the convolutional units are sequentially connected to form the input convolutional part, and a downsampling operation is set between adjacent convolutional units, that is, the output of the previous convolutional unit is input to the next convolutional unit after downsampling. Another five convolutional units are sequentially connected to form the output convolutional part, and a transposed convolution operation is set between adjacent convolutional units, that is, the output of the previous convolutional unit is input to the next convolutional unit after transposed convolution. The output result of the i-th convolutional unit in the input convolutional part and the result after transposed convolution of the output of the (5 - i)-th convolutional unit in the input convolutional part are connected and jointly input to the (6 - i)-th convolutional unit in the input convolutional part, where i = 1 - 4; the 5th convolutional unit in the input convolutional part is connected to the remaining one convolutional unit through downsampling, and the output result of the 5th convolutional unit in the input convolutional part and the result after transposed convolution of the output of the remaining one convolutional unit are connected and input to the 1st convolutional unit in the output convolutional part; the input of the first convolutional unit at the beginning of the input convolutional part is used as the input of the model, and the output of the last 5th convolutional unit in the output convolutional part is used as the output of the model; each convolutional unit is mainly composed of two convolutional modules connected in sequence, and each convolutional module is mainly composed of a convolution operation, a normalization operation, and an activation function connected in sequence.

[0060] In specific implementation, the high resolution is resampled to a resolution of (1.99, 0.82, 0.82), and the low resolution is resampled to a resolution of (2.19, 1.11, 1.11).

[0061] After high-resolution resampling and low-resolution resampling, the operation of collecting patches can be performed. Under hardware limitations, the patch size input to the high-resolution network is 40×192×224, and the patch size of the low-resolution network is 56×192×224.

[0062] In specific implementation, the neural network model is pre-trained using a training set. During training, data augmentation is performed on each image in the training set. The data augmentation includes three operations: random brightness enhancement and weakening, linear contrast stretching, and non-linear contrast stretching, which are performed in sequence, and each operation is set with a probability of 15%. In addition, the adjustment range of random brightness enhancement and weakening, the range of linear contrast stretching, and the range of non-linear contrast stretching are increased, which can increase the adaptability of the neural network model to multi-center data.

[0063] S3. The pancreatic segmentation mask is processed by a mapping method to obtain the pancreatic fat deposition distribution and iron deposition distribution;

[0064] Multi-echo sequence single acquisition can obtain in-phase / opposite-phase / water / fat / proton density fat fraction / iron deposition R2* images. By directly adopting a mapping method, corresponding values are taken in the proton density fat fraction image and the iron deposition R2* image to obtain the pancreatic fat deposition distribution and the iron deposition distribution.

[0065] S4. Divide the pancreas into regions according to the pancreatic segmentation mask to obtain each pancreatic sub-region, and obtain the pancreatic fat deposition distribution and the iron deposition distribution of each pancreatic sub-region.

[0066] Obtain the skeleton of the pancreatic segmentation mask through morphological processing;

[0067] Calculate the total length of the skeleton, calculate the ranges of the pancreatic head, neck, body, and tail according to a preset ratio, and divide the pancreas into regions to obtain each pancreatic sub-region; in a specific implementation, the partitioning is set according to a ratio of 3.5:3:3.5:3.

[0068] Calculate the pancreatic fat deposition distribution and the iron deposition distribution of each pancreatic sub-region.

[0069] As Figure 3 shows the 3D visualization interface of the pancreatic fat deposition distribution and the iron deposition distribution of an obese patient in the embodiment. Among them, the ectopic fat deposition and iron deposition in the pancreatic head are relatively obvious.

Claims

1. A magnetic resonance pancreatic fat and iron deposition distribution measurement system, characterized in that, The system includes: An image acquisition module for acquiring magnetic resonance multi-echo images; A pancreas prediction module for obtaining a pancreas segmentation mask by segmenting the multi-echo images; A mapping module for obtaining pancreatic fat deposition distribution information and iron deposition distribution information according to the pancreas segmentation mask; A sub-region parameter calculation module for obtaining the pancreatic fat fraction and iron deposition content of sub-regions of the pancreas by processing using the distribution information and the pancreas segmentation mask; The system is used to implement a method for measuring the distribution of pancreatic fat and iron deposition by magnetic resonance, and the method includes: S1. Collect multi-echo images of the abdomen through a magnetic resonance instrument; S2. Extract in-phase and opposed-phase dual-modal images from the multi-echo images, and input the dual-modal images into a neural network model to obtain a pancreas segmentation mask; S3. Process the pancreas segmentation mask by a mapping method to obtain pancreatic fat deposition distribution information and iron deposition distribution information; S4. Divide the pancreas into regions according to the pancreas segmentation mask to obtain each sub-region of the pancreas, and obtain the pancreatic fat deposition distribution and iron deposition distribution of each sub-region of the pancreas; The neural network model is mainly composed of a first 3D high-resolution model and a 3D cascade model fused together, and the 3D cascade model is composed of a 3D low-resolution model and a second 3D high-resolution model connected in sequence; Input the dual-modal images into the first 3D high-resolution model after data normalization and high-resolution resampling, and then output the first pancreas segmentation probability output result; at the same time, input the dual-modal images into the 3D cascade model after data normalization and low-resolution resampling, first input them into the 3D low-resolution model to obtain a preliminary pancreas segmentation mask with a lower resolution, and use the preliminary pancreas segmentation mask after high-resolution resampling as a third modality and input it into the second 3D high-resolution model in the 3D cascade model together with the dual-modal images, and then output the second pancreas segmentation probability output result; fuse the first pancreas segmentation probability output result and the second pancreas segmentation probability output result by means of averaging or voting, and then perform binarization after fusion to obtain the final pancreas segmentation mask; The structures of the first 3D high-resolution model, 3D low-resolution model, and second 3D high-resolution model are the same, each including eleven convolutional units. Five of the convolutional units are sequentially connected to form the input convolutional part, and downsampling operations are set between adjacent convolutional units to connect the other five convolutional units that are sequentially connected to form the output convolutional part. Deconvolution operations are set between adjacent convolutional units. The output result of the i-th convolutional unit in the input convolutional part and the result after deconvolution of the output of the (5 - i)-th convolutional unit in the input convolutional part are connected and then jointly input into the (6 - i)-th convolutional unit in the input convolutional part, where i = 1 - 4; the 5th convolutional unit in the input convolutional part is connected to the remaining one convolutional unit through downsampling, and the output result of the 5th convolutional unit in the input convolutional part and the result after deconvolution of the output of the remaining one convolutional unit are connected and then input into the 1st convolutional unit in the output convolutional part; the input of the 1st convolutional unit at the beginning of the input convolutional part is used as the input of the model, and the output of the 5th convolutional unit at the end of the output convolutional part is used as the output of the model; each convolutional unit is mainly composed of two convolutional modules connected in sequence, and each convolutional module is mainly composed of a convolutional operation, a normalization operation, and an activation function connected in sequence; The high resolution is resampled to a resolution of (1.99, 0.82, 0.82), and the low resolution is resampled to a resolution of (2.19, 1.11, 1.11).

2. The magnetic resonance pancreas fat and iron deposition distribution measurement system according to claim 1, wherein: The specific steps of S2 are as follows: The in-phase / opposite-phase / water / fat / proton density fat fraction / iron deposition R2* images of the human abdomen are obtained by single-shot acquisition through a multi-echo sequence. The in-phase image and the opposite-phase image among them are used as bimodal images to be input into the neural network model for training, and the pancreatic segmentation mask is obtained through the neural network model.

3. The magnetic resonance pancreas fat and iron deposition distribution measurement system according to claim 1, characterized in that: The neural network model is pre-trained using a training set. During training, data augmentation processing is performed on each image in the training set. The data augmentation includes three operations of randomly enhancing and weakening brightness, linearly stretching contrast, and non-linearly stretching contrast, which are sequentially performed, and each operation is set with a probability of 15%. Additionally, the adjustment range of randomly enhancing and weakening brightness, the range of linearly stretching contrast, and the range of non-linearly stretching contrast are increased.

4. The magnetic resonance pancreatic fat and iron deposition distribution measurement system according to claim 1, wherein: S3 specifically is: mapping the pancreatic segmentation mask to the proton density fat fraction image and the iron deposition R2* image of the multi-echo image.

5. The magnetic resonance pancreas fat and iron deposition distribution measurement system according to claim 1, wherein: The specific steps of S4 include the following: Obtaining the skeleton of the pancreatic segmentation mask through morphological processing; Calculating the total length of the skeleton, calculating the ranges of the pancreatic head, neck, body, and tail according to a preset ratio, and partitioning the pancreas to obtain each pancreatic sub-region; Calculating the pancreatic fat fraction and iron deposition content of each pancreatic sub-region according to the pancreatic fat deposition distribution information and iron deposition distribution information.

6. An electronic device, characterized in that: It includes a memory, a processor, and computer instructions stored on the memory and running on the processor. When the computer instructions are run by the processor, the method for measuring the distribution of pancreatic fat and iron deposition according to any one of claims 1 - 5 is completed.

Images