A method for constructing a benign and malignant prediction model for pancreatic tumors based on CT images

Abstract

This invention discloses a method for constructing a benign / malignant prediction model based on CT images of pancreatic tumors. The method includes: acquiring preoperative pancreatic CT images through an image storage and transmission system; identifying and preprocessing a region of interest (ROI) based on the pancreatic cancer tumor region in the CT images; extracting radiomics features from the processed ROI; filtering the extracted features using a maximum correlation minimum redundancy algorithm to select features related to the benign / malignant nature of the pancreatic tumor; constructing a benign / malignant prediction model based on the selected features using a support vector machine model; and outputting the final prediction result. The method described in this invention enables rapid and accurate classification of benign / malignant pancreatic tumors, providing auxiliary information for clinical diagnosis and assisting physicians in developing personalized treatment plans.

Description

Technical Field

[0001] This invention relates to the field of medical image analysis, and in particular to the construction and application of a pancreatic tumor diagnostic model based on CT images, which is a combined application of radiomics and artificial intelligence-assisted diagnostic technologies. Background Technology

[0002] Pancreatic cancer is one of the most aggressive malignant tumors and the seventh leading cause of cancer-related death worldwide. It is characterized by its insidious onset, high malignancy, and difficulty in treatment. Due to the deep anatomical location of the pancreas, symptoms are often not obvious in the early stages, making early diagnosis of pancreatic cancer a major clinical challenge. While CT imaging is widely used for the diagnosis of pancreatic cancer, the identification and interpretation of imaging features require a high level of experience and skill from physicians and is subject to subjective factors. Currently, surgical resection combined with postoperative chemotherapy, immunotherapy, and gene-targeted therapy are the main treatment methods for pancreatic cancer. Due to the unique digestive enzyme secretion function of the pancreas, the risk of postoperative complications after pancreatic surgery is far greater than that after surgery on other organs, and in severe cases, it can be life-threatening. Therefore, developing a model that can automatically extract CT image features and use them for predicting the benign or malignant nature of pancreatic tumors has significant clinical significance and application value. Early and accurate diagnosis of pancreatic tumors can avoid postoperative complications that may result from radical pancreatic cancer surgery in patients with benign tumors, contributing to safe, minimally invasive, and precise treatment of pancreatic tumors.

[0003] In recent years, radiomics technology has been increasingly widely used in medical image analysis, especially in tumor diagnosis and treatment. Radiomics enables high-throughput mining of quantitative features from standard medical images, allowing these data to be effectively used in clinical decision support systems, thereby improving the accuracy of diagnosis, prognosis, and treatment response assessment. The extraction of radiomic features, through quantitative analysis of different texture, morphology, and density characteristics in images, can help reveal the underlying biological characteristics of tumors. Based on these features, more personalized treatment plans can be developed for patients.

[0004] Artificial intelligence (AI) technologies, particularly machine learning and deep learning-based models, have made significant progress in medical image processing in recent years. Deep learning models can automatically learn complex feature representations from large amounts of labeled data, demonstrating superior performance in tasks such as medical image classification, detection, and segmentation. Especially in the diagnosis and prognostic prediction of pancreatic cancer, the combination of radiomics and AI technologies has shown remarkable results. For example, deep learning models based on convolutional neural networks (CNNs) have achieved high sensitivity and specificity in the detection of pancreatic tumors. However, research on predicting benign and malignant tumors based on pancreatic CT images is relatively limited. The imaging features of pancreatic cancer are complex, and the anatomical location of the pancreas and the influence of surrounding structures make the identification of its imaging features highly challenging. Therefore, there is a need to develop an efficient and universally applicable automated model to assist clinicians in diagnosis. Summary of the Invention

[0005] This invention aims to improve the accuracy of early diagnosis of pancreatic cancer and reduce the risk of patients undergoing unnecessary surgery due to misdiagnosis by establishing a CT image-based model for predicting the benign and malignant nature of pancreatic tumors.

[0006] This invention combines radiomics feature extraction with an artificial intelligence classification model. Through in-depth analysis of tumor regions in CT images, it extracts various types of radiomics features, including first-order statistical features, texture features, and morphological features. To improve the model's generalization ability, the invention employs the maximum relevance minimum redundancy (mRMR) algorithm to filter features, thereby retaining features related to tumor benignity and malignancy while reducing redundant information. Finally, a support vector machine (SVM) model is used to construct a predictive model. The model, trained and validated on patient data, demonstrates high diagnostic performance and stability.

[0007] Therefore, this invention not only optimizes image data processing and feature extraction but also incorporates advanced machine learning techniques for predicting the benign or malignant nature of pancreatic tumors, providing a reliable auxiliary tool for clinical practice. Future research and applications based on this model may combine it with other multimodal imaging data (such as MRI and PET-CT) and patient clinical information to further enhance diagnostic performance and clinical application value.

[0008] This invention discloses a method for constructing a benign / malignant prediction model based on CT images of pancreatic tumors, comprising the following steps: acquiring preoperative pancreatic CT images through an image storage and transmission system; obtaining a region of interest (ROI) based on the pancreatic cancer tumor region in the CT images and performing preprocessing; extracting radiomics features based on the processed ROI; using a maximum correlation minimum redundancy algorithm to screen the extracted features, selecting features related to the benign / malignant nature of pancreatic tumors; constructing a benign / malignant prediction model based on the screened features using a support vector machine model, and outputting the final prediction result. This model combines radiomics feature extraction, feature screening, and machine learning methods, aiming to provide clinicians with an auxiliary diagnostic tool and improve the diagnostic accuracy and early detection capability of pancreatic cancer.

[0009] The main steps of this invention include:

[0010] Data Acquisition and Image Processing: Preoperative pancreatic CT images are acquired through a Picture Archiving and Communication System (PACS), including enhanced thin-slice CT data containing arterial, portal venous, delayed, and portal venous phase slices. Regions of interest (ROIs) are derived based on the pancreatic cancer tumor regions within the CT images. For example, 3D Slicer software is used to delineate the tumor regions in the images to obtain the ROI.

[0011] Preprocessing and radiomics feature extraction: All CT images were resampled and standardized at a resolution of 3mm to ensure consistency across different image sources. Next, all images were normalized, and outliers were removed using a 1.5x interquartile range to improve model accuracy. For example, the Pyradiomics library was used to extract first-order statistical features, texture features, shape features, and other features from the CT images. High-throughput mining of these features effectively reflects the internal heterogeneity and microstructural characteristics of pancreatic tumors.

[0012] Feature Selection and Modeling: Due to the large number of extracted features, to avoid excessive model complexity and improve the model's generalization ability, this invention employs the Maximum Relevance Minimum Redundancy (mRMR) algorithm to select features related to the benign or malignant nature of pancreatic tumors. Subsequently, a classification prediction model based on the selected features, i.e., a benign / malignant prediction model, is constructed using a Support Vector Machine (SVM) model. SVM, as a classifier, has good generalization ability and stability, and is particularly suitable for classifying small-sample, high-dimensional data.

[0013] Model Validation and Evaluation: To validate the model's performance, patient data was randomly divided into a training set and a validation set, used for model training and performance evaluation, respectively. The model was evaluated using metrics such as the receiver operating characteristic curve (ROC) and the area under the curve (AUC). The results showed that the SVM model had high AUC values ​​in both the training and validation sets, indicating that the model has good accuracy and stability in predicting the benign and malignant nature of pancreatic tumors.

[0014] The innovation of this invention lies in constructing an automated tool for predicting the benign or malignant nature of pancreatic tumors through multi-phase image fusion and high-throughput feature extraction, combined with machine learning technology. It reduces subjectivity in imaging assessment, improves diagnostic accuracy, and helps clinicians better identify and treat pancreatic cancer patients at an early stage. Furthermore, this invention can be further combined with imaging data from other modalities (such as MRI and PET-CT) and patient clinical information to enhance diagnostic performance and clinical application value. Attached Figure Description

[0015] Figure 1 This is a flowchart of the method for constructing a benign or malignant prediction model based on CT images of pancreatic tumors according to the present invention.

[0016] Figure 2 This is a partial schematic diagram of the method. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will now be described in further detail with reference to the accompanying drawings and embodiments. It should be noted that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the invention.

[0018] Figure 1 This is a flowchart of the method for constructing a benign / malignant prediction model for pancreatic tumors using CT images of pancreatic tumors, as per the present invention. Specifically, it includes the following steps:

[0019] 1. Acquire preoperative pancreatic CT images using a Picture Archiving and Communication System (PACS);

[0020] 2. Based on the pancreatic cancer tumor region in the CT images, obtain the region of interest (ROI) of the pancreatic tumor lesion and perform preprocessing;

[0021] 3. Radiomics features were extracted based on the processed regions of interest (ROIs) of pancreatic tumor lesions;

[0022] 4. The maximum relevance minimum redundancy (mRMR) algorithm was used to screen the extracted features and select those that were related to the benign or malignant nature of pancreatic tumors;

[0023] 5. Construct a benign / malignant prediction model based on screening features using a Support Vector Machine (SVM) model, and output the prediction results. The prediction results could be, for example, the probability of benign status.

[0024] Figure 2 This is a partial schematic diagram of the method. P1 represents image processing and tumor segmentation; P2 represents feature extraction, where (a) represents first-order features, (b) represents texture features, and (c) represents morphological features; P3 represents feature selection; and P4 represents building a model to output predicted classification results, where OC1 represents the receiver operating characteristic curve of the training set, and OC2 represents the receiver operating characteristic curve of the validation set.

[0025] Specifically, in step 1 above, in order to improve the accuracy of model prediction, preoperative pancreatic CT images are acquired through a Picture Storage and Transmission System (PACS), such as enhanced thin-slice CT data containing thin slices of the arterial phase, portal venous phase, delayed phase, and portal venous phase.

[0026] In step 2, to ensure accurate delineation, for example, two professional radiologists with more than five years of experience in imaging diagnosis can manually delineate the tumor region in the image using 3D Slicer software to obtain the region of interest (ROI). If inconsistencies arise in the assessment, they can be further corrected and confirmed by senior physicians.

[0027] For each patient's CT image data, the images are first resampled to ensure a uniform resolution of 3mm across all images. This step is to eliminate potential differences caused by different imaging devices and scanning parameters, ensuring consistency in feature extraction.

[0028] After resampling, the image is normalized to adjust all pixel values ​​to a uniform range. Normalization can be performed as follows:

[0029]

[0030] Where X: the image pixel value in the original data;

[0031] X min The minimum number of image pixels in the original data;

[0032] X max The maximum value of the image pixels in the original data;

[0033] X norm : Normalized image pixel values.

[0034] In addition, outliers can be removed using a method of 1.5 times interquartile range, thereby reducing the impact of noise on model performance.

[0035] In step 3 above, for example, the Pyradiomics library can be used to extract multi-dimensional features for each ROI region, including first-order statistical features (such as mean and variance), texture features (such as gray-level co-occurrence matrix and gray-level run-length matrix), and morphological features. These features fully capture the morphology and internal texture of the tumor, reflecting its complex biological behavior.

[0036] In step 4 above, this invention employs the Maximum Relevance Minimum Redundancy (mRMR) algorithm to filter features, selecting those related to the benign or malignant nature of pancreatic tumors. This aims to avoid excessive model complexity and improve the model's generalization ability. Selecting features related to the benign or malignant nature of pancreatic tumors can specifically involve selecting five features related to their benign or malignant nature, for example... Figure 2 Five features were selected as shown, where feature 1 could be...

[0037] log_sigma_5_0_mm_3D_glrlm_LongRunLowGrayLevelEmphasis_CA, corresponding to the Chinese name "arterial phase image feature: 5.0 mm standard deviation logarithmic filtering of the three-dimensional grayscale run length matrix for long-range low grayscale value emphasis"; feature 2 could be...

[0038] log_sigma_2_0_mm_3D_glrlm_LongRunLowGrayLevelEmphasis_CD, corresponding to the Chinese name "Long RunLowGrayLevelEmphasis_CD," refers to the long-range low grayscale value emphasis of a 2.0 mm standard deviation logarithmic filtering three-dimensional grayscale run length matrix for delayed image features. Feature 3 could be...

[0039] log_sigma_5_0_mm_3D_glcm_MaximumProbability_CD, corresponding to the Chinese name of the maximum probability of the 5.0 mm standard deviation log-filtered three-dimensional gray-level co-occurrence matrix of delayed image features; feature 4 can be...

[0040] wavelet_LLH_gldm_DependenceNonUniformityNormalized_CV, corresponding to the Chinese name "Dependence Non-Uniformity Normalized_CV" for wavelet LLH filtering of venous phase image features; feature 5 can be...

[0041] log_sigma_2_0_mm_3D_glcm_ClusterShade_CV, corresponding to the Chinese name "clustered shadow of a three-dimensional gray-level co-occurrence matrix with 2.0 mm standard deviation logarithmic filtering for venous phase image features".

[0042] In step 5 above, this invention utilizes a Support Vector Machine (SVM) model to construct a classification prediction model based on screening features, namely a benign / malignant prediction model. During model construction, training set data is first used to learn and optimize model parameters. After training, independent validation set data is used to evaluate the model's performance. Evaluation metrics include the Receiver Operating Characteristic (ROC) curve and its area under the curve (AUC), used to measure the model's classification accuracy and robustness. Figure 2 The final results show that the prediction model achieved AUCs of 0.97 and 0.87 on the training and validation sets, respectively, indicating that the model has high accuracy and good generalization ability in the task of predicting benign and malignant cases.

[0043] All or part of the steps described in this invention can be implemented using Python program instructions and R language code.

[0044] This invention provides a tumor benignity / malignancy prediction model based on pancreatic tumor CT images, which has the following advantages:

[0045] Using the method of this invention, clinicians can perform rapid, non-invasive preoperative assessment of the benign or malignant nature of pancreatic tumors, thereby assisting in the development of more precise treatment plans. For patients suspected of having pancreatic cancer, this model can reduce overtreatment of benign tumors and provide more timely intervention for patients with malignant tumors.

[0046] Furthermore, this model can be further combined with other imaging modalities (such as MRI and PET-CT) and the patient's clinical information (such as tumor marker levels and the patient's medical history) to improve the accuracy and comprehensiveness of diagnosis, thereby playing a more important role in multidisciplinary treatment teams.

[0047] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and do not constitute a limitation thereof. Those skilled in the art should understand that any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for constructing a benign / malignant prediction model based on CT images of pancreatic tumors, comprising the following steps: Preoperative pancreatic CT images are acquired through an image storage and transmission system; based on the pancreatic cancer tumor region in the CT images, a region of interest is obtained and preprocessed; radiomics features are extracted based on the processed region of interest; the extracted features are screened using the maximum correlation minimum redundancy algorithm to select features related to the benign or malignant nature of pancreatic tumors; a benign or malignant prediction model based on the screened features is constructed using a support vector machine model, and the final prediction result is output.

2. The method for constructing a benign / malignant prediction model based on pancreatic tumor CT images according to claim 1, characterized in that, The CT images include enhanced thin-slice CT data of the arterial phase, portal venous phase, delayed phase, and portal venous phase.

3. The method for constructing a benign / malignant prediction model based on pancreatic tumor CT images according to claim 1, characterized in that, The process of obtaining the region of interest involves delineating the pancreatic cancer tumor region in the CT image to obtain the region of interest.

4. The method for constructing a benign / malignant prediction model based on pancreatic tumor CT images according to claim 1, characterized in that, The preprocessing involved resampling all CT images to achieve a uniform resolution, then normalizing all CT images and removing outliers using a 1.5 interquartile range.

5. The method for constructing a benign / malignant prediction model based on pancreatic tumor CT images according to claim 1, characterized in that, The extraction of radiomics features based on the processed region of interest includes extracting the gray-level co-occurrence matrix, the gray-level run length matrix, and morphological features.

6. The method for constructing a benign / malignant prediction model based on pancreatic tumor CT images according to claim 1, characterized in that, The selection of features related to the benign or malignant nature of pancreatic tumors includes selecting five imaging features that are associated with the benign or malignant classification of pancreatic tumors.

Images