Pancreatic fat automatic measurement method and system based on deep convolutional neural network

Abstract

The invention relates to the technical field of medical image analysis and artificial intelligence, and solves the technical problems that manual pancreatic fat segmentation of an MRI image is time-consuming clinically at present, obvious variability between observers and in the observers exists, and large-scale clinical application and standardization are difficult to realize. The pancreatic fat automatic measurement method based on the deep convolutional neural network comprises the following steps: processing input magnetic resonance image data to form a superpixel region with a clearer boundary; then, the image is sent to a DCNN processing model, and a pancreas area is accurately identified in a complex background; and finally, accurately calculating the pancreas volume and the fat volume in the pancreas based on a segmentation result. The pancreatic fat and the pancreatic volume can be automatically and precisely measured, a reliable and standardized quantitative tool is provided for early diagnosis, risk stratification and treatment selection of metabolic diseases such as type 2 diabetes, pancreatic cancer and pancreatitis, and the pancreatic fat and pancreatic volume measuring device has important clinical application value.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of medical image analysis and artificial intelligence technology, and particularly relates to a method for processing magnetic resonance imaging (MRI) images, CT images and ultrasound images by using deep learning technology to realize automatic measurement of pancreatic fat and pancreatic volume. BACKGROUND

[0002] Pancreatic fat (PF) content and pancreatic volume are considered as key prognostic indicators for metabolic diseases such as type 2 diabetes (T2D), pancreatic cancer and pancreatitis. Early identification and quantification of pancreatic fat are of great significance for early diagnosis, risk stratification and treatment selection of these diseases. However, there are many challenges in the current clinical methods for quantification of pancreatic fat.

[0003] Traditional medical imaging techniques such as ultrasound (US) and computed tomography (CT) have limitations in quantifying pancreatic fat. Ultrasound examination cannot quantify the length of the entire pancreas, especially for obese patients; while CT has low sensitivity to the gradient response of pancreatic adipose tissue. Magnetic resonance imaging (MRI) is considered as an ideal method for quantifying pancreatic fat due to its higher soft tissue contrast and non-radiation advantage, especially the proton density fat fraction mapping (PDFF) technique. However, even MRI-PDFF still faces challenges in quantifying the internal fat deposition of the pancreas, which is a relatively small organ. The softness of the pancreas makes it susceptible to compression by surrounding organs, resulting in blurred boundaries and difficulty in distinguishing it from non-pancreatic soft tissues such as small intestine, blood vessels and abdominal visceral fat tissue. This leads to highly variable results in measuring pancreatic volume and fat deposition by MRI.

[0004] Currently, manual segmentation of MRI images for pancreatic fat is considered as the "gold standard" in clinic, but this method is extremely time-consuming and has significant inter-observer and intra-observer variability, making it difficult to achieve large-scale clinical application and standardization. In order to reduce this variability, some manual pancreatic fat measurement methods quantify the region of interest (ROI) within the entire pancreas and exclude pixels with fat percentage values less than 1% and greater than 20% (representing histologically verified blood vessels, ducts or visceral fat), but the time-consuming nature of manual segmentation remains a major bottleneck.

[0005] In recent years, deep convolutional neural networks (DCNNs) have made significant progress in medical image analysis, especially in the automatic segmentation of organs such as the heart, liver, kidney, and spleen. However, due to the special nature of the pancreas as a retroperitoneal organ and its high inter-observer variability, research on automatic pancreas segmentation is relatively limited. In addition, while 3D MRI scans provide rich additional information, they also make it more complex to use classical atlas models to establish 3D models for context learning, limited by current GPU memory. Although newer 2D and 3D networks have been developed and strategies such as the coarse-to-fine framework, recurrent learning, and ensemble learning have been applied to improve the accuracy of segmentation results, there is still no research report on the automatic quantification of pancreatic fat.

[0006] Therefore, there is still a gap in the existing technology for accurate and automated measurement of pancreatic fat, especially the lack of an automated tool that can effectively process multi-protocol MRI images, overcome the problem of ambiguous pancreatic boundaries, and achieve high-precision fat quantification and distribution tracking. The present invention aims to solve the above technical problems and provide an efficient, accurate, and generalizable method and system for automatic measurement of pancreatic fat. SUMMARY

[0007] To address the shortcomings of the prior art, the present invention provides a method and system for automatic measurement of pancreatic fat based on deep convolutional neural networks, which solves the technical problems of time-consuming manual segmentation of MRI images for pancreatic fat in current clinical practice, significant inter-observer and intra-observer variability, and difficulty in achieving large-scale clinical application and standardization.

[0008] The present invention can be applied to early auxiliary diagnosis, risk assessment, and treatment effect monitoring of metabolic diseases such as type 2 diabetes, pancreatic cancer, and pancreatitis. Specifically, a method for automatic measurement of pancreatic fat based on deep convolutional neural networks, the method comprising the following steps:

[0009] S1, acquiring clinical multi-protocol magnetic resonance image data and performing preprocessing to generate visualizable visual summary images, the preprocessing including image format conversion, size cropping, and enhancing tissue boundary contrast using a superpixel segmentation method;

[0010] S2, constructing a DCNN processing model for learning and imitating the expert manual pancreas segmentation process based on a deep convolutional neural network model to identify the segmented pancreas region and the fat tissue within the pancreas;

[0011] S3, using the trained DCNN processing model to automatically segment new magnetic resonance images and identify the pancreas region and its internal fat tissue to generate an automatic segmentation result;

[0012] S4、based on the automatic segmentation result, automatically calculate the volume of the pancreas and the volume of fat in the pancreas, and the fat fraction and fat deposition distribution of the pancreas.

[0013] Further, the superpixel segmentation method adopts a linear spectral clustering method to segment the superpixels of the magnetic resonance image data, so as to enhance the contrast of the pancreas tissue boundary;

[0014] Then, pixels with similar color, brightness and texture features are clustered to generate a visual summary image, thereby significantly improving the contrast of different tissue boundaries and greatly improving the accuracy of pancreas recognition.

[0015] Further, the ratio of spatial proximity and color similarity in the linear spectral clustering method is determined by using a simulated annealing algorithm.

[0016] Further, the visual summary image is used for dimensionality reduction processing of the magnetic resonance image data, and the pixel value in the visual summary image is determined by the average gray value of the pixels in the corresponding superpixel.

[0017] Further, the deep convolutional neural network model is a U-Net model, which integrates the feature kernel of a VGG-16 network, thereby effectively reducing the amount of calculation and improving the segmentation efficiency while ensuring the segmentation accuracy.

[0018] Further, the DCNN processing model is configured to only perform convolution classification on the center superpixel in the visual summary image.

[0019] Further, an erosion algorithm is used to remove the boundary pixels of the fat tissue identified by the DCNN processing model, and the volume of fat in the pancreas is calculated based on the fat tissue after removing the boundary pixels.

[0020] Further, the volume of the pancreas is calculated by multiplying the area of each segmented pancreas slice by the preset slice thickness and summing them up.

[0021] By the above technical solutions, the present application provides a pancreas fat automatic measurement method and system based on a deep convolutional neural network, which has at least the following beneficial effects:

[0022] High precision and automation: accurate and automatic segmentation and measurement of the pancreas and pancreatic fat, Dice score up to 91.2%, regression R 2 value (pancreas 0.9764, pancreatic fat volume 0.9675) verifies its high consistency with manual operation, significantly improving the diagnostic efficiency and accuracy.

[0023] Addressing the pain points of existing technologies: Effectively solves the problems of time-consuming manual segmentation, large variability among observers, and insufficient accuracy of existing automated methods.

[0024] First achievement: The first framework for measuring pancreatic fat volume and fat deposition has been established, providing a new tool for the quantitative study of pancreatic fat.

[0025] Wide range of clinical applications: It can serve as an effective tool for the early diagnosis, risk assessment, and treatment efficacy monitoring of metabolic diseases such as type 2 diabetes, pancreatic cancer, and pancreatitis, and has significant clinical application value.

[0026] Economic benefits: It is expected to save significant costs for the healthcare system and provide new growth opportunities for medical device manufacturers through commercial deployment.

[0027] In summary, the method proposed in this invention achieves automated and high-precision measurement of pancreatic fat and pancreatic volume. It significantly improves measurement efficiency and accuracy, reduces inter- and intra-observer variability, and provides a reliable and standardized quantitative tool for the early diagnosis, risk stratification, and treatment selection of metabolic diseases such as type 2 diabetes, pancreatic cancer, and pancreatitis, thus possessing significant clinical application value. Attached Figure Description

[0028] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0029] Figure 1 This is a network architecture diagram of the DCNN processing model in this invention;

[0030] Figure 2 This is a flowchart illustrating the pancreatic tissue scanning process and the acquisition of its internal fat volume in this invention.

[0031] Figure 3 This is a comparison chart of the pancreatic fat ratio between the normal blood glucose control group and type 2 diabetes patients in this invention;

[0032] Figure 4 The figure shows the comparison results of the differences in the number of superpixels and their MRI image segmentation effects in the improved LSC algorithm of this invention.

[0033] Figure 5 This refers to the average Dice score of the DCNN processing model when using preprocessed images and when not using preprocessed images in this invention;

[0034] Figure 6 This is a comparison chart of the performance of superpixel centering and non-centering processing in this invention;

[0035] Figure 7 This is a diagram showing the results of verifying the integrity of the DCNN processing model after training in this invention. Detailed Implementation

[0036] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. This will allow for a full understanding of how the present application uses technical means to solve technical problems and achieve technical effects, and to facilitate its implementation.

[0037] Example 1:

[0038] This embodiment aims to overcome the challenges faced by existing automated pancreatic fat measurement technologies, proposing an automated pancreatic fat measurement method based on deep convolutional neural networks. This method achieves accurate, efficient, and automated quantification of pancreatic fat and pancreatic volume in clinical multi-protocol magnetic resonance imaging, providing a reliable basis for early auxiliary diagnosis and risk assessment of metabolic diseases. The specific process is as follows:

[0039] 1. Image preprocessing

[0040] Researchers included abdominal fluid and fat magnetic resonance imaging (MR) images of the subjects in their analysis. These images were converted to NiFTI format using MRICroGL software and resized to 512×512 pixels with an average thickness of 3 mm after image processing. This example selected 2364 MRI fluid imaging images (1020 from the patient group and 1344 from the control group). All these images contained identifiable pancreatic images, and clinical medical experts manually labeled them using LabelMe to construct a DCNN. Images from 315 subjects were used for training, and 79 datasets were used for testing (approximately an 80:20 ratio).

[0041] Magnetic resonance imaging (MRI) provides clearer details of soft tissues due to its slower imaging speed, but its relatively low resolution often results in numerous artifacts at the pancreatic boundaries. Therefore, image preprocessing to enhance the contrast of organ and tissue boundaries is crucial. Superpixel segmentation technology generates a visual summary image by clustering pixels with similar color, brightness, and texture features. This method significantly improves the contrast of different tissue boundaries, thereby greatly enhancing the accuracy of pancreatic identification. The image preprocessing in this model consists of two stages: superpixel segmentation and image dimensionality reduction.

[0042] 1.1 Superpixel Segmentation

[0043] This embodiment employs the latest optimized superpixel segmentation method—Linear Spectral Clustering (LSC)—for preprocessing pancreatic images. This algorithm is widely recognized as the best-performing superpixel segmentation method for processing three-channel natural images. This embodiment adapts the LSC algorithm for single-channel medical imaging tasks and compares it with Simple Linear Iterative Clustering (SLIC0), a state-of-the-art method for medical image superpixel segmentation. The LSC algorithm for three-channel natural images uses the following pixel distance calculation formula:

[0044] (1)

[0045] in, Represents pixels Pixel spacing between them; Spatial proximity; For color similarity; , Represents pixels Pixel space coordinates; Represents pixels The pixel grayscale value.

[0046] Compare this with the pixel pitch measurement formula in the original LSC, namely:

[0047] (2)

[0048] Then keep the same pixel space distance value At the same time, the pixel color distance value is changed from Turn ,Right now:

[0049] (3)

[0050] (4)

[0051] To improve pancreas segmentation performance under limited GPU conditions, the variables in the formula... The pixel color value is now replaced with a grayscale value. This is achieved by adjusting the spatial proximity. Similarity to color ratio The improved LSC method can generate superpixels with higher compactness (CP) and boundary recall (BR).

[0052] (5).

[0053] ratio A higher value indicates a higher degree of spatial pixel clustering, resulting in superpixels with high compactness (CP) but low boundary recall (BR). Conversely, a lower ratio results in lower clustering. This will result in pixels with low compactness (CP) but high boundary recall (BR). To obtain superpixels that simultaneously possess high compactness (CP) and high boundary recall (BR), this embodiment proposes the following new metrics:

[0054] (6)

[0055] When the minimum positive value is reached, compactness (CP) and boundary recall (BR) will tend to balance.

[0056] In linear spectral clustering, the ratio of spatial proximity to color similarity is fixed. This leads to inconsistent superpixel segmentation performance across different magnetic resonance imaging (MRI) datasets, limiting segmentation accuracy. To address this, this embodiment employs simulated annealing (SA) to determine this ratio. The value is adjusted to optimize the superpixel segmentation effect and improve the segmentation accuracy.

[0057] This embodiment uses a simulated annealing algorithm to search for the optimal ratio of spatial proximity to color similarity, minimizing the energy function. In each iteration of the simulated annealing algorithm, the ratio of spatial proximity to color similarity is randomly changed, and a new energy function value is calculated. If the new energy function value is less than the current energy function value, the change is accepted; otherwise, the change is accepted with a certain probability, which decreases as the temperature decreases. Through multiple iterations, a better ratio of spatial proximity to color similarity can eventually be found, resulting in better superpixel segmentation.

[0058] By employing simulated annealing to dynamically determine the ratio of spatial proximity to color similarity in linear spectral clustering, the parameters of superpixel segmentation can be adaptively adjusted, thereby improving segmentation accuracy and robustness. Furthermore, optimizing superpixel segmentation provides more accurate input for subsequent DCNN processing models, thus enhancing the accuracy of automated pancreatic fat measurement.

[0059] This embodiment uses superpixel segmentation evaluation metrics to compare the performance of the improved LSC algorithm and the SLIC0 algorithm. The evaluation metrics include segmentation error (UE), boundary recall (BR), and achievable segmentation precision (ASA), all of which can be calculated using the following formula:

[0060] (7)

[0061] In this formula, Os Indicates the number of observations for segmented pixels that should not be displayed; Rs This is the theoretical number of pixels; Us This represents the number of pixels not displayed in the actual segmented image, which should theoretically be displayed.

[0062] The formula for calculating BR is:

[0063] (8)

[0064] in, G and S Let represent the reference boundary and the actual boundary of the superpixel segmented image, respectively, given the true values. According to formulas (2) and (4), the maximum pixel distance is 1. TP Represents the actual boundary S Center and reference boundary G In pixel space distance value d The true value of the number of overlapping pixels within the range. FN That is TP The opposite value of .

[0065] The formula for calculating ASA is:

[0066] (9)

[0067] in , Rs This indicates a reference area manually marked by experts. Ts This refers to the region generated by the trained DCNN processing model.

[0068] 1.2 Image Dimensionality Reduction

[0069] This embodiment adjusts the image through dimensionality reduction and generates a visual summary map that only displays features between superpixels. Combining average pooling techniques from deep learning, this embodiment inputs the DCNN processing model (i.e., the average grayscale value of all pixels within each superpixel), and then redistributes the grayscale values ​​to generate a schematic diagram. The preprocessed abdominal schematic diagram clearly presents the separation features of organs and tissues, providing effective support for pancreas segmentation.

[0070] In this embodiment, after the magnetic resonance image data is input into the system, it is first converted in format and cropped to ensure that the image data meets the input requirements of the DCNN processing model. Subsequently, superpixel segmentation can be used to process the image to enhance the contrast of pancreatic tissue boundaries. Superpixel segmentation technology generates a visual summary image by clustering pixels with similar color, brightness, and texture features, thereby significantly improving the contrast of different tissue boundaries and greatly improving the accuracy of pancreatic identification.

[0071] 1.3 DCNN Processing Model Establishment and Performance Evaluation

[0072] This embodiment fine-tunes the existing U-Net model and, based on a large amount of manually labeled MRI water imaging training data, builds a novel single-branch DCNN architecture from scratch to achieve convergence. To improve the accuracy of the DCNN processing model, the feature kernels learned from the lower layers of the VGG-16 network are transferred to the domain of this embodiment. This network exhibits extremely strong discriminative ability while maintaining stable training convergence. The U-Net model has achieved breakthrough results in medical image segmentation by integrating low-level and high-level information, improving accuracy and complex feature extraction capabilities, respectively.

[0073] The final layer of the DCNN processing model uses a 1×1 convolutional classification mechanism, which significantly slows down the inference process. Therefore, in practical applications, directly performing convolutional classification on the entire image may lead to a waste of computational resources, especially when there are many non-target regions in the image, and may introduce unnecessary noise, affecting the accuracy of segmentation. To address this, this embodiment further proposes... Figure 1 The acceleration method shown accelerates the DCNN processing model by assigning superpixel centers. The DCNN processing model only performs convolutional classification on the labeled superpixels at the center positions, which represent pixel clusters with the same features. This method ensures both network speed and classification performance.

[0074] The DCNN processing model is configured to perform convolutional classification only on the central superpixel in the visual summarization image, rather than classifying all pixels within a superpixel. Specifically, after superpixel segmentation, the system identifies the center point of each superpixel and extracts a local image patch centered on that center point, or directly uses the central pixel as the classification object. This aims to improve processing efficiency and segmentation accuracy by focusing on key regions of the image.

[0075] This embodiment significantly reduces the amount of data to be processed by limiting the convolutional classification operation of the DCNN processing model to the central superpixel. Since superpixel segmentation divides the image into regions with similar features, the central superpixel of each superpixel effectively represents the feature information of that region. Therefore, classifying only the central superpixel avoids redundant calculations of pixels within the superpixel, thereby significantly reducing the computational burden on the deep convolutional neural network model and improving processing efficiency. Simultaneously, this focus on key information points helps reduce the interference of background noise and irrelevant regions on the classification results, enabling the DCNN processing model to more accurately identify the pancreatic region and the adipose tissue within the pancreas.

[0076] This training was completed on a workstation equipped with an Intel i9-11900K processor and an NVIDIA 4080 NVIDIA Ti graphics card, using the Python 3.6.0 programming language and the deep learning frameworks TensorFlow and Keras. The DCNN processing model training pipeline consists of successive iterations of forward and backward propagation. Forward propagation generates a pixel-by-pixel prediction set of the pancreatic region from the input image. The segmentation accuracy of each sample is evaluated using the Dice similarity coefficient (DSC), which measures the degree of overlap between the manually annotated pancreatic region and the prediction masks (labeled X and Y, respectively) of the DCNN processing model.

[0077] (10)

[0078] In this embodiment, the core of the DCNN processing model is a deep convolutional neural network model. This model learns from a large amount of labeled data, enabling it to automatically identify and segment the pancreatic region and the adipose tissue within it. For example, during training, the DCNN processing model receives a large number of magnetic resonance images and their corresponding expert-annotated pancreatic and adipose regions. By continuously adjusting the network parameters, it strives to make its output segmentation results as consistent as possible with the expert annotations. Once the model is trained, it can automatically segment new magnetic resonance images.

[0079] To verify the reliability of the DCNN processing model, this embodiment applied it to 10 newly diagnosed prediabetic patients recruited in a recent clinical study. Two independent and experienced researchers segmented the pancreas volume and fat percentage, and then compared their results with those generated by the machine. The processing time and the correlation between the segmentation results were observed.

[0080] 2. Measurement principles of pancreatic volume and pancreatic fat deposition

[0081] This embodiment, after obtaining the segmentation results of the pancreatic region and adipose tissue, can calculate the total volume of the pancreas and the volume of fat within the pancreas. The pancreatic volume can be calculated by multiplying the area of ​​each segmented pancreatic slice by a preset slice thickness and then summing the results. The volume of fat within the pancreas can be calculated based on the number of identified adipose tissue pixels. These calculation results can be presented to the user in the form of intuitive charts or detailed reports to assist doctors in diagnosis.

[0082] 2.1 Phantom Research

[0083] This embodiment converts MRI pixel information into actual three-dimensional space using a series of phantom emulsion scans and calibrates the machine's detection ratio for adipose tissue. To assess the accuracy of each quantitative result, this embodiment uses vegetable oil (soybean oil) and distilled water to prepare a series of homogenized emulsions, a method fully described by Bernard. 0-100% fat volume fraction and lecithin (1% by weight, Sigma-Aldrich) were added to 100 mL bottles. To stabilize the emulsion, 3% by weight of agar gel and sodium dioctyl sulfosuccinate were added to the system. The emulsion was slowly prepared on a heated stirring plate and then cooled to room temperature to ensure uniform suspension distribution. The bottles were then placed in containers of solid agar and scanned under the same MRI operating conditions as the subjects. Monomeric nuclear magnetic resonance spectroscopy (MRS) was performed on 10 emulsions with fat volume fractions ranging from 10-100%, with scan parameters set to a repetition time of 4 seconds, an echo time of 23 milliseconds, and a bandwidth of 2.5 kHz. Figure 2 A shows the simulated phantom results under the same MRI conditions. Among them, (a) is an aqueous phase MRI scan; (b) is a fat phase MRI scan; (c) is a fat gradient heatmap generated by MRS application; and (d) is a machine learning digitally transformed fat fraction gradient generated by the Python Pillow algorithm.

[0084] 2.2 Obtaining Pancreatic Volume

[0085] In this embodiment, aqueous MR images are input into a fully trained DCNN processing model, and the segmented regions are calculated using a newly trained framework to further accumulate the total volume. According to Cavalieri's principle, the segmented volume is calculated as the product of the area of ​​each pancreatic slice (typically 6-9 pancreatic MRI images per subject) and its thickness (3 mm), as shown in the following formula: Two-dimensional pancreatic images are superimposed and reconstructed to ultimately build a three-dimensional model, such as... Figure 2 The pancreatic volume acquisition method and its three-dimensional reconstruction results are shown in Figure B.

[0086]

[0087] This embodiment calculates the pancreatic volume by multiplying the area of ​​each segmented pancreatic slice by a preset slice thickness and then summing the results. The preset slice thickness refers to the fixed thickness of each slice layer during magnetic resonance imaging (MRI). Specifically, this thickness is a known, pre-set parameter that can be obtained from the metadata of the MRI image data. By multiplying the area of ​​each segmented pancreatic slice by the preset slice thickness, the volume of that pancreatic slice layer can be obtained. Furthermore, summing the volumes of all slices yields a more accurate total pancreatic volume.

[0088] This embodiment, by taking into account the thickness of the slices, can more accurately calculate pancreatic volume, thus avoiding calculation errors caused by neglecting slice thickness. This method is particularly suitable for cases with irregular pancreatic shapes or thicker slices, providing more reliable pancreatic volume measurements and offering more accurate references for clinical diagnosis and treatment.

[0089] 2.3 Obtaining the volume of pancreatic fat

[0090] In the fat images, researchers quantified the area of ​​pancreatic fat pixels manually. Pixel percentages were automatically assessed and recorded using ImageJ software (National Institutes of Health, Bethesda, Maryland). To reduce inter- and intra-observer variability, pixels with a fat percentage less than 1% or greater than 20% were manually removed; these pixels corresponded to histologically validated blood vessels, ducts, or visceral fat.

[0091] This embodiment enables the DCNN processing model to perform the same functions as manual measurement. During the process of the DCNN processing model identifying adipose tissue, due to the inherent characteristics of deep learning models, its segmentation accuracy in boundary regions may be affected. This can lead to the identified adipose tissue boundary pixels containing noise or uncertainty, thus affecting the accuracy of the final fat volume calculation. To further improve the accuracy of pancreatic fat volume calculation, this embodiment uses an erosion algorithm to remove the boundary pixels of the adipose tissue identified by the DCNN processing model, and calculates the pancreatic fat volume based on the adipose tissue after removing the boundary pixels.

[0092] In this application, the erosion algorithm is used to remove boundary pixels of adipose tissue identified by the DCNN processing model. This means that extra pixels that may have been generated due to inaccurate segmentation will be removed in the edge areas of adipose tissue, making the contour of the adipose tissue more compact and accurate.

[0093] The newly developed DCNN system in this embodiment can acquire MRI fat images with redundant boundaries. To avoid redundant boundary contamination and improve computational performance, this embodiment uses the erosion algorithm in OpenCV to remove redundant boundary pixels in the segmented image. This algorithm achieves denoising by calculating the local minimum of a given kernel value. In this embodiment, the kernel value is continuously adjusted in the DCNN until the ideal segmentation effect is obtained. The optimal performance stability is achieved when the kernel value is adjusted to 5. Figure 2 The erosion kernel value shown in C is optimized. The phantom study matches the brightness of each pixel in the MRI image with the corresponding actual fat percentage. For example... Figure 2 The process of obtaining pancreatic fat volume from MRI fat images to contour pixels, shown in Figure D, demonstrates the performance of the erosion algorithm and machine pixel annotation in MRI fat images.

[0094] This embodiment, after the DCNN processing model identifies adipose tissue, introduces an erosion algorithm to process the boundary pixels of the adipose tissue, effectively eliminating errors or noise that may exist in the boundary segmentation of the deep convolutional neural network model. It is precisely because the erosion algorithm can accurately shrink the boundaries of adipose tissue and remove uncertain or potentially misidentified edge pixels that more accurate and reliable results can be obtained when calculating the volume of pancreatic fat based on the adipose tissue after removing boundary pixels.

[0095] 2.4 Statistics

[0096] This study employed the SciPy algorithm in Python for statistical analysis. Independent samples t-tests were used to compare the differences in pancreatic fat fraction detection between diabetic patients and healthy controls with normal blood glucose levels, as well as differential scanning calorimetry (DSC) scores under different training conditions. A 95% confidence interval for the critical p-value was considered statistically significant when it was less than 0.05. Linear regression analysis was used to verify the differences in segmentation performance between expert and newly developed DCNN methods.

[0097] 3. Results

[0098] 3.1 Evaluation of the correlation between pancreatic fat ratio and type 2 diabetes

[0099] In this embodiment, manually measured pancreatic fat percentage data were grouped according to the total number of participants and type 2 diabetes (T2D) status. Student's t-test was used to analyze differences between groups, and the results are as follows: Figure 3 As shown, there was a significant difference in pancreatic fat fraction between the normoglycemic control group and diabetic patients, with a t-value of -13.0799 and p<0.01.

[0100] 3.2 Performance of Superpixel Segmentation

[0101] 3.2.1 Comparison results between the improved LSC and SL-IC0

[0102] This embodiment optimizes the LSC algorithm using a novel distance measurement method. For example... Figure 4 Figure A shows a comparison of the improved LSC algorithm and the SL-IC0 algorithm in terms of superpixel segmentation metrics, where a, b, and c correspond to the comparison of UE, BR, and ASA metrics, respectively. The performance of the improved LSC algorithm is compared with that of the SL-IC0 algorithm, and the results show that LSC outperforms SL-IC0 in UE, BR, and ASA metrics. Experimental data indicate that the LSC method is more in line with the actual needs of medical image superpixel segmentation.

[0103] 3.2.2 Requesting the value of r using the improved LSC method

[0104] Medical image superpixel segmentation results typically exhibit high boundary response (BR) and contrast response (CP) values. To achieve this, this embodiment adjusts c using an improved LSC method. s With c c The ratio of values. Based on the experimental results of the simulated annealing algorithm, this embodiment sets the range of r values ​​to the interval [0.05, 1], and calculates the corresponding I value accordingly. Figure 4 B shows the correlation between different superpixel segmentation r values ​​and I values. When r=0.25, the CP and BR values ​​reach the best balance, and the I value reaches the optimal minimum value.

[0105] 3.2.3 Optimal number of superpixels

[0106] This embodiment preprocesses medical images and obtains corresponding superpixel diagrams to achieve optimal pancreas segmentation. The processed images are then input into a DCNN model for training and testing until the model optimizes its DSC value. The number of superpixels obtained is a key factor determining segmentation quality. After 20 training epochs of DCNN, the number of superpixels in the preprocessed images was adjusted to 500, 1000, 1500, 2000, 2500, 3000, and 3500, respectively. Figure 4 As shown in Figure C, the results indicate that the DSC value tends to be optimal when the number of superpixels is 2500.

[0107] 3.3 Evaluation of pretreatment at different stages

[0108] This example compares the average Dice score of the DCNN processing model with and without pre-processed images under different training epochs. The results show that the DCNN processing model using pre-processed image input achieves a higher average Dice score. Figure 5 Figure A shows the DSC comparison results of the improved DCNN processing models with and without superpixel preprocessing at different training stages. Statistical analysis shows that the p-values ​​at training stages 1, 10, and 20 are 0.05, 0.01, and 0.01, respectively, indicating significant differences in results at each stage. The accuracy curve based on Dice scores shows that the weighted network trained with preprocessed images tends to stabilize after 20 training epochs, as shown in Figure A. Figure 5 As shown in B, even without preprocessing images, the model performance stabilizes after 20 training epochs. Considering time costs, this embodiment ultimately determines the optimal training conditions for the DCNN processing model to be 20 training epochs and the number of superpixels to be 2500.

[0109] 3.4 Segmentation Integrity of DCNN Processing Model

[0110] Figure 6 A andFigure 6 B illustrates the performance comparison between centralized and decentralized superpixel processing, including... Figure 6 A comparison of the superpixel centering effects of the DCNN processing model shown in Figure A. Figure 6 Figure B shows a performance comparison of pancreatic image superpixel segmentation methods (including centered and non-centered processing). Experimental results show that superpixel centering significantly improves the DSC value of the DCNN processing model, reaching a maximum DSC value of 91.2% after 20 training epochs. Overall, the DSC value of the DCNN processing model fully meets practical requirements, and the integrity of the pancreatic organ in the test image is as shown. Figure 6 As shown in C, the pancreas segmentation results are compared between the DCNN processing model and the expert manual operation. p<0.05 indicates a significant difference.

[0111] 3.5 Independent Validation of the DCNN Processing Model

[0112] The performance of the DCNN processing model was validated through professional manual segmentation in newly recruited patients. Regression analysis plots show the correlation between the two segmentation methods on pancreatic volume and pancreatic fat ratio, as follows: Figure 7 As shown. Performance validation was performed, measuring the differences in pancreatic and pancreatic fat volume between expert manual operation and the DCNN processing model. The R² values ​​for pancreatic volume and fat ratio were 0.9764 and 0.9675, respectively. Preliminary conclusions can be drawn that the segmentation results of the newly generated DCNN processing model are more reliable than those obtained manually. Furthermore, the DCNN processing model significantly reduced the time required, decreasing the average measurement time per patient from 1.5 hours for manual operation to 5 seconds for the DCNN processing model.

[0113] Example 2:

[0114] This embodiment proposes an automated pancreatic fat measurement system based on a deep convolutional neural network, including a data input module, an image preprocessing module, a DCNN processing module, a result analysis and output module, and a user interface. Wherein:

[0115] The system includes a data input module for receiving and managing magnetic resonance imaging data from different clinical protocols; an image preprocessing module for standardizing and enhancing the input images, including superpixel segmentation; a DCNN processing module equipped with a trained deep convolutional neural network model for automatically segmenting the pancreas and pancreatic fat; a results analysis and output module for calculating pancreatic volume, pancreatic fat volume, and fat fraction, and outputting the measurement results in a visual or report format; and a user interface for users to upload images, view results, and set parameters.

[0116] In practical applications, the input magnetic resonance imaging data is first processed by an image preprocessing module. This module enhances the contrast of pancreatic tissue boundaries through steps such as image format conversion, size cropping, and superpixel segmentation, providing high-quality input for subsequent DCNN processing models. For example, superpixel segmentation can cluster similar pixels in the image to form superpixel regions with clearer boundaries, thereby effectively reducing image noise and highlighting the features of the pancreatic region.

[0117] Subsequently, the preprocessed image data is fed into the DCNN processing module. This module is equipped with a pre-trained deep convolutional neural network model, capable of automatically identifying and segmenting the pancreatic region and adipose tissue within the pancreas in the image. Through its multi-layer convolution and pooling operations, the DCNN processing module can extract complex features from the image and classify each pixel based on these features, determining whether it belongs to the pancreas or adipose tissue. For example, the DCNN processing module can learn the pancreatic-specific texture, shape, and grayscale distribution patterns, thereby accurately identifying the pancreatic region in complex backgrounds.

[0118] Finally, the automatic segmentation results output by the DCNN processing module are passed to the results analysis and output module. Based on the segmentation results, this module accurately calculates the pancreatic volume and the volume of pancreatic fat. For example, by performing 3D reconstruction of the segmented pancreatic region and combining it with slice thickness, the total volume of the pancreas can be calculated. Simultaneously, similar calculations are performed on the segmented adipose tissue region to obtain the volume of pancreatic fat. These calculation results can be presented to users in the form of intuitive charts, reports, or 3D visualization models, providing clinicians with comprehensive and accurate diagnostic information. The entire system's workflow automates and enables high-precision processing from raw image data to final measurement results, significantly improving the efficiency and reliability of pancreatic fat measurement.

[0119] Those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Moreover, this application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0120] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. Since the above embodiments are substantially similar to the method embodiments, their descriptions are relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0121] The above embodiments provide a detailed description of the present invention. Specific examples have been used to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. An automatic method for measuring pancreatic fat based on a deep convolutional neural network, characterized in that, The method includes the following steps: S1. Acquire clinical multi-protocol magnetic resonance image data and perform preprocessing to generate a visual summary image. The preprocessing includes image format conversion, size cropping, and using superpixel segmentation to enhance tissue boundary contrast. S2. A DCNN processing model based on a deep convolutional neural network model is constructed to learn and imitate the expert manual pancreatic segmentation process in order to identify and segment pancreatic regions and pancreatic adipose tissue. S3. The trained DCNN processing model is used to automatically segment the new magnetic resonance image and identify the pancreatic region and its internal adipose tissue to generate automatic segmentation results. S4. Based on the automatic segmentation results, automatically calculate the pancreatic volume and pancreatic fat volume, as well as the pancreatic fat fraction and fat deposition distribution.

2. The automatic pancreatic fat measurement method according to claim 1, characterized in that, The superpixel segmentation method uses linear spectral clustering to perform superpixel segmentation on the magnetic resonance image data to enhance the contrast of pancreatic tissue boundaries; Then, pixels with similar color, brightness, and texture features are clustered to generate a visual summary image.

3. The automatic pancreatic fat measurement method according to claim 2, characterized in that, The ratio of spatial proximity to color similarity in the linear spectral clustering method is determined using the simulated annealing algorithm.

4. The automatic pancreatic fat measurement method according to claim 1, characterized in that, The visual summary image is used to perform dimensionality reduction processing on the magnetic resonance image data, and the pixel values ​​in the visual summary image are determined by the average grayscale value of the pixels within the corresponding superpixel.

5. The automatic pancreatic fat measurement method according to claim 1, characterized in that, The deep convolutional neural network model is a U-Net model, which integrates the feature kernels of the VGG-16 network.

6. The automatic pancreatic fat measurement method according to claim 1, characterized in that, The DCNN processing model is configured to perform convolutional classification only on the central superpixels in the visual summary image.

7. The automatic pancreatic fat measurement method according to claim 1, characterized in that, An erosion algorithm is used to remove the boundary pixels of the adipose tissue identified by the DCNN processing model, and the volume of fat in the pancreas is calculated based on the adipose tissue after removing the boundary pixels.

8. The automatic pancreatic fat measurement method according to claim 1, characterized in that, The pancreatic volume is calculated by multiplying the area of ​​each segmented pancreatic slice by a preset slice thickness and summing the results.

9. A system for implementing the automatic pancreatic fat measurement method according to any one of claims 1-8, characterized in that, include: The data input module is used to receive and manage magnetic resonance image data from different clinical protocols; The image preprocessing module is used to perform normalization and enhancement on the input image, including superpixel segmentation; The DCNN processing module, equipped with a pre-trained deep convolutional neural network model, is used to perform automatic segmentation tasks of the pancreas and pancreatic fat. The results analysis and output module is used to calculate pancreatic volume, pancreatic fat volume, and fat fraction, and output the measurement results in the form of visualization or reports. The user interface is used by users to upload images, view results, and set parameters.

Images