Disease risk prediction system based on multi-region abdominal fat radiomics characteristics

Abstract

The invention discloses a disease risk prediction system based on multi-region abdominal fat imageomics characteristics. The disease risk prediction system comprises a fat segmentation module used for performing fat segmentation on five types of fat regions including subcutaneous fat, retroperitoneal fat, abdominal anterior fat, mesenteric fat and pelvic medial fat on a preprocessed image by using a neural network model; the feature extraction module is used for extracting image features from the segmented fat region; the ratio calculation module is used for calculating a key ratio index between fat regions; and the risk prediction module is used for constructing a risk prediction model, establishing independent dichotomous sub-models for diseases, and performing disease risk prediction by using the dichotomous sub-models. According to the method, full-process analysis from image preprocessing, region segmentation, feature extraction to risk scoring can be automatically completed under the condition that manual intervention is not needed, and the structural utilization efficiency of the abdomen image in clinic is greatly improved.

Description

Technical Field

[0001] This invention relates to the field of image recognition, and more particularly to a disease risk prediction system based on multi-regional abdominal fat radiomics features. Background Technology

[0002] With the continuous rise in the incidence of metabolic diseases (such as metabolic syndrome and diabetes) and chronic inflammatory diseases (such as inflammatory bowel disease and Crohn's disease), the medical community is increasingly emphasizing the role of body fat distribution in the occurrence, development, and risk assessment of these diseases. Traditional methods of fat assessment rely heavily on superficial indicators such as body mass index (BMI), waist circumference, and waist-to-hip ratio. While these indicators have some clinical reference value, they are insufficient to accurately reflect the distribution characteristics and pathological significance of fat in different anatomical regions of the abdomen.

[0003] Numerous studies have shown that adipose tissue in different anatomical locations differs significantly in structure, morphology, metabolic activity, and involvement in disease mechanisms. Subcutaneous fat primarily serves a cushioning and heat-preserving function, while visceral adipose tissues such as retroperitoneal fat, abdominal mesorectal fat, and medial pelvic fat exhibit higher metabolic activity and are closely associated with various diseases. For example:

[0004] Increased retroperitoneal fat is considered a potential marker of insulin resistance and metabolic syndrome;

[0005] Abnormal proliferation of abdominal mesorectal fat is prominent in inflammatory bowel disease (especially Crohn's disease) and is known as "creeping fat," with its volume positively correlated with disease activity.

[0006] The distribution characteristics of fat in the inner pelvis are closely related to the complexity of pelvic surgery and postoperative recovery, and also have potential evaluation value in patients with prostate cancer, rectal cancer, etc.

[0007] Although some studies have attempted to quantitatively analyze the volume or area of ​​abdominal fat using medical imaging (such as CT and MRI), most methods are limited to classifying fat into two broad categories: "visceral fat" and "subcutaneous fat," lacking refined differentiation of specific anatomical regions such as the retroperitoneum, mesentery, and medial pelvis. Furthermore, current clinical assessments largely rely on manual image interpretation, which is highly subjective, inefficient, and struggles to accurately capture high-dimensional microstructural information such as morphology, density, and texture of fat regions. A systematic analysis of "radiomics characteristics" is also lacking.

[0008] Furthermore, existing medical imaging AI systems mostly focus on organ identification or tumor detection, and have not yet established a comprehensive modeling method based on the path of "abdominal fat structure identification - feature extraction - disease risk modeling". In particular, research on the diagnostic value of ratios between fat regions (such as retroperitoneal fat / total abdominal fat, mesangial fat / subcutaneous fat, etc.) is scarce, and standardized analysis strategies or predictive models have not yet been formed. For example, "A method for analyzing abdominal fat components" (Chinese Patent No.: CN110415246A) and "A method for segmenting subcutaneous fat and visceral fat regions based on abdominal CT images" (Chinese Patent No.: CN114359302A).

[0009] Therefore, there is a need for a technical solution that can automatically segment multiple abdominal fat regions, extract their radiomics features, and combine them with inter-regional ratio information to construct an intelligent, multi-task disease risk prediction model, so as to achieve higher accuracy, stronger interpretability, and greater clinical value in fat imaging assessment. Summary of the Invention

[0010] The purpose of this invention is to provide a disease risk prediction system based on multi-regional abdominal fat radiomics features to solve the problems mentioned in the background art.

[0011] To achieve the above-mentioned objectives, this invention provides a disease risk prediction system based on multi-regional abdominal fat radiomics features, comprising a fat segmentation module, a feature extraction module, a ratio calculation module, and a risk prediction module, wherein:

[0012] The fat segmentation module is used to segment five types of fat regions—subcutaneous fat, retroperitoneal fat, anterior abdominal fat, mesorectal fat, and medial pelvic fat—using a neural network model on preprocessed images.

[0013] The feature extraction module is used to extract image features from the segmented fat regions;

[0014] The ratio calculation module is used to calculate key ratio indicators between fat regions;

[0015] The risk prediction module is used to build risk prediction models, establish independent binary sub-models for each disease, and use the binary sub-models to predict disease risk.

[0016] Furthermore, the fat segmentation module normalizes the CT image using an image processing tool before segmentation.

[0017] Furthermore, the neural network model is a 2D U-net neural network model, which takes a preprocessed image as input and outputs a predicted fat segmentation result.

[0018] Furthermore, the 2D U-net neural network model is trained by selecting effective slices containing at least one fat region from sagittal 2D slices manually annotated by professional radiologists, and the training data and test data are divided in an 8:2 ratio. The training data is used to train the 2D U-net neural network model.

[0019] Furthermore, the model input is a 3-channel image composed of three consecutive layers of adjacent slices, and the training parameters are set to batch_size=8 and learning rate 5e. -4 We used Dice Loss + Focal Loss as the loss function and achieved automatic and accurate segmentation of fat regions through end-to-end training.

[0020] Furthermore, the feature extraction module uses pyradiomics to automatically extract various image omics features (including volume, density distribution, texture features, morphological parameters, etc.). The extracted image features include: original shape features, first-order histogram features, gray-level co-occurrence matrix features, gray-level dependency matrix features, gray-level run-length matrix features, gray-level region size matrix features, and neighborhood gray-level difference matrix features.

[0021] Furthermore, in the ratio calculation module, the key ratio indicators include: retroperitoneal fat / abdominal mesorectal fat, abdominal mesorectal fat / subcutaneous fat, medial pelvic fat / retroperitoneal fat, retroperitoneal fat / abdominal fat, abdominal mesorectal fat / abdominal fat, medial pelvic fat / abdominal fat, and abdominal fat / subcutaneous fat.

[0022] Furthermore, the risk prediction module uses LightGBM to construct disease risk prediction models, establishing binary sub-models for metabolic syndrome, type 2 diabetes, and inflammatory bowel disease. The inputs of the binary sub-models are radiomics features and volume ratio features, and the outputs the risk probability of the corresponding disease are compared and evaluated with the real labels.

[0023] Furthermore, the binary sub-model is trained using binary labels for the corresponding diseases. All labels are annotated by professionals, and the disease risk probability is output through the sigmoid function.

[0024] Compared with existing technologies, this system and method have the following advantages:

[0025] 1. For the first time, five metabolically specific fat regions (abdominal fat, retroperitoneal fat, subcutaneous fat, abdominal mesorectal fat, and medial pelvic fat) were defined and segmented in CT images, overcoming the limitation of existing technologies that can only segment 2-3 general regions.

[0026] 2. Innovatively combining fat region volume ratio (e.g., retroperitoneal fat / subcutaneous fat) with radiomics features (e.g., morphology, texture, grayscale co-occurrence matrix) to solve the problem of insufficient predictive efficacy of single indicators.

[0027] 3. It can automatically complete the entire process of analysis, from image preprocessing, region segmentation, feature extraction to risk scoring, without manual intervention, greatly improving the efficiency of structured utilization of abdominal images in clinical practice. On the one hand, it provides objective reference for less experienced doctors, assisting in clinical interpretation; on the other hand, it can also help experienced experts improve efficiency and enhance consistency in high-intensity image reading or screening scenarios, possessing good clinical promotion value and industrial application prospects. Attached Figure Description

[0028] Figure 1 This is a flowchart illustrating the structure of a disease risk prediction system based on multi-regional abdominal fat imaging features.

[0029] Figure 2 This is a schematic diagram of the fat segmentation results. Detailed Implementation

[0030] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0031] like Figure 1 The diagram shown is a flowchart of the system structure of this invention. This invention provides a disease risk prediction system based on multi-regional abdominal fat imaging features. The system includes:

[0032] The fat segmentation module is used to segment five types of fat regions—subcutaneous fat, retroperitoneal fat, anterior abdominal fat, mesorectal fat, and medial pelvic fat—using a neural network model on preprocessed images.

[0033] The feature extraction module is used to extract image features from the segmented fat regions;

[0034] The ratio calculation module is used to calculate key ratio indicators between fat regions;

[0035] The risk prediction module is used to build risk prediction models, establish independent binary sub-models for each disease, and use the binary sub-models to predict disease risk.

[0036] The fat segmentation module includes a self-trained fat segmentation model that can preprocess CT images and normalize the input raw images using a window width of 200 and a window level of -50.

[0037] For fat segmentation, a 2D U-net neural network is used. The preprocessed image is input, and the predicted fat segmentation results are output, including five types of fat: subcutaneous fat, retroperitoneal fat, anterior abdominal fat, mesorectal fat, and medial pelvic fat. The sagittal segmentation results are shown below. Figure 2 As shown in the diagram, the red area represents abdominal fat, the green area represents retroperitoneal fat, the blue area represents subcutaneous fat, the yellow area represents mesorectal fat, and the cyan area represents pelvic fat.

[0038] Specifically, the five fat regions (subcutaneous fat, retroperitoneal fat, anterior abdominal fat, mesorectal fat, and medial pelvic fat) in the 40 cases were manually labeled by professional radiologists. The model was trained using sagittal 2D slices, and 16,728 valid slices containing at least one fat region were selected and divided into training data and test data in an 8:2 ratio.

[0039] The model takes a 3-channel image composed of three consecutive adjacent slices as input and outputs five segmentation tasks, specifically for segmenting the following five fat regions: subcutaneous fat, retroperitoneal fat, anterior abdominal fat, mesorectal fat, and medial pelvic fat. Training parameters are set to batch_size=8 and learning rate 5e. -4 We used Dice Loss + Focal Loss as the loss function and achieved automatic and accurate segmentation of fat regions through end-to-end training.

[0040] The feature extraction module is used for radiomics feature extraction. The open-source radiomics extraction tool Pyradiomics was used to extract image features from five types of fat regions: subcutaneous fat, retroperitoneal fat, anterior abdominal fat, mesorectal fat, and medial pelvic fat. Each type yielded 107 image features, including: 14 original shape features, 18 first-order histogram features, 24 gray-level co-occurrence matrix features, 14 gray-level dependency matrix features, 16 gray-level run-length matrix features, 16 gray-level region size matrix features, and 5 neighborhood gray-level difference matrix features.

[0041] The ratio calculation module uses the volume of the above five types of fat regions to calculate key ratio indicators between fat regions, including the following seven: retroperitoneal fat / abdominal mesorectal fat, abdominal mesorectal fat / subcutaneous fat, medial pelvic fat / retroperitoneal fat, retroperitoneal fat / abdominal fat, abdominal mesorectal fat / abdominal fat, medial pelvic fat / abdominal fat, and abdominal fat / subcutaneous fat.

[0042] The risk prediction module uses LightGBM to build disease risk prediction models, establishing three independent binary sub-models for metabolic syndrome, type 2 diabetes, and inflammatory bowel disease. Each binary sub-model is trained only on the binary labels of the corresponding disease.

[0043] The input to each binary classification sub-model is a fused feature vector, including:

[0044] (1) Radiomic features from 5 fat regions (a total of 535 dimensions, summarized from each region);

[0045] (2) Characteristics of volume ratios in 7 regions.

[0046] During model training, grid search is used to fine-tune hyperparameters such as learning rate, maximum depth, number of leaf nodes, and L2 regularization strength, with the average AUC of cross-validation used as the evaluation metric. Features are standardized using the mean and standard deviation of the training set before training, and missing values ​​are imputed using the mean. An early stopping mechanism is employed during training; if the log loss on the validation set does not improve within 50 consecutive epochs, training is terminated early.

[0047] Each binary sub-model is trained using binary labels (positive / negative) for the corresponding disease, with all labels annotated by professionals. The model outputs the disease risk probability using a sigmoid function, and the loss function is binary cross-entropy.

[0048] This experiment was based on abdominal imaging data from 150 cases, which were randomly divided into a training set (80%, 120 cases) and a test set (20%, 30 cases) at an 8:2 ratio. In the training set, there were 52 patients with metabolic syndrome, 45 with type 2 diabetes, and 38 with inflammatory bowel disease (some patients had two or three diseases). In the test set, there were 13 patients with metabolic syndrome, 11 with type 2 diabetes, and 10 with inflammatory bowel disease.

[0049] Three binary risk prediction sub-models were constructed using the LightGBM algorithm. During the testing phase, the models output the risk probability of corresponding diseases based on the individual's input radiomics features, and the results were compared with the ground truth labels. The results showed that the metabolic syndrome sub-model achieved an accuracy of 84% on the test set, the type 2 diabetes sub-model 81%, and the inflammatory bowel disease sub-model 79%. Overall, the experimental results demonstrate that the risk prediction model can effectively predict the risk of these three types of diseases, validating its application value in early disease screening and risk management.

[0050] The fat region segmentation method used in this invention employs the UNet-Like deep neural network method. Other deep learning and machine learning methods can be used to obtain fat segmentation results, which are not substantially different from this invention.

[0051] The feature extraction method used in this invention is the open-source toolkit Pyradiomics. Other methods can be used for feature extraction, and they are not substantially different from those in this invention.

[0052] The risk model used in this invention is based on the LightGBM algorithm. Using other methods such as Random Forest and XGBoost to construct disease risk models is not substantially different from this invention.

[0053] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A disease risk prediction system based on multi-regional abdominal fat radiomics features, characterized in that, It includes a fat segmentation module, a feature extraction module, a ratio calculation module, and a risk prediction module, among which: The fat segmentation module is used to segment five types of fat regions—subcutaneous fat, retroperitoneal fat, anterior abdominal fat, mesorectal fat, and medial pelvic fat—using a neural network model on preprocessed images. The feature extraction module is used to extract image features from the segmented fat regions; The ratio calculation module is used to calculate key ratio indicators between fat regions; The risk prediction module is used to build risk prediction models, establish independent binary sub-models for each disease, and use the binary sub-models to predict disease risk.

2. The disease risk prediction system based on multi-regional abdominal fat radiomics features according to claim 1, characterized in that, The fat segmentation module normalizes the CT images using image processing tools before segmentation.

3. The disease risk prediction system based on multi-regional abdominal fat radiomics features according to claim 1, characterized in that, The neural network model is a 2D U-net neural network model, which takes a preprocessed image as input and outputs the predicted fat segmentation result.

4. The disease risk prediction system based on multi-regional abdominal fat radiomics features according to claim 3, characterized in that, The 2D U-net neural network model is trained using sagittal 2D slices manually annotated by professional radiologists. Valid slices containing at least one fat region are selected, and training and testing data are divided in an 8:2 ratio. The training data is used to train the 2D U-net neural network model.

5. The disease risk prediction system based on multi-regional abdominal fat radiomics features according to claim 3 or 4, characterized in that, The model input is a 3-channel image composed of three consecutive layers of adjacent slices. The training parameters are set to batch_size=8 and learning rate 5e. -4 We used Dice Loss + Focal Loss as the loss function and achieved automatic and accurate segmentation of fat regions through end-to-end training.

6. The disease risk prediction system based on multi-regional abdominal fat radiomics features according to claim 1, characterized in that, The image features extracted by the feature extraction module include: original shape features, first-order histogram features, gray-level co-occurrence matrix features, gray-level dependency matrix features, gray-level run-length matrix features, gray-level region size matrix features, and neighborhood gray-level difference matrix features.

7. The disease risk prediction system based on multi-regional abdominal fat radiomics features according to claim 1, characterized in that, In the ratio calculation module, the key ratio indicators include: retroperitoneal fat / abdominal mesorectal fat, abdominal mesorectal fat / subcutaneous fat, medial pelvic fat / retroperitoneal fat, retroperitoneal fat / abdominal fat, abdominal mesorectal fat / abdominal fat, medial pelvic fat / abdominal fat, and abdominal fat / subcutaneous fat.

8. The disease risk prediction system based on multi-regional abdominal fat radiomics features according to claim 1, characterized in that, The risk prediction module uses LightGBM to build disease risk prediction models, establishing binary sub-models for metabolic syndrome, type 2 diabetes, and inflammatory bowel disease. The inputs of the binary sub-models are radiomics features and volume ratio features, and the outputs the risk probability of the corresponding disease are compared and evaluated with the real labels.

9. The disease risk prediction system based on multi-regional abdominal fat radiomics features according to claim 1 or 8, characterized in that, The binary sub-model is trained using binary labels for the corresponding diseases. All labels are annotated by professionals, and the disease risk probability is output through the sigmoid function.

Images