Search

CN-122025092-A - Ultrasonic image deep learning and clinical feature fusion-based gallbladder polypoid lesion benign and malignant prediction method and device and storage medium thereof

CN122025092ACN 122025092 ACN122025092 ACN 122025092ACN-122025092-A

Abstract

A method for predicting benign and malignant gallbladder polyp lesions based on deep learning of ultrasonic images and clinical feature fusion comprises the steps of deep learning of ultrasonic images through a deep learning network, outputting malignant polyp risk scores, inputting the malignant polyp risk scores and clinical features into a trained machine learning model for learning, and obtaining final probability of benign and malignant gallbladder polyp lesions, wherein the clinical features comprise CA19-9, CEA and maximum polyp diameter, and the machine learning model is selected from one of Logistic regression, lightGBM, random Forest and XGBoost, extraTrees, naiveBayes, SVM.

Inventors

  • QU CHAO
  • JIANG YUMIN
  • HU WEIYU
  • ZENG JIANGE
  • Nie Silin
  • SUN RUITAO
  • JIANG DANNI
  • TAN BIN
  • CAO JINGYU

Assignees

  • 青岛大学附属医院

Dates

Publication Date
20260512
Application Date
20260128

Claims (9)

  1. 1. A method for predicting benign and malignant gallbladder polypoid lesions based on ultrasonic image deep learning and clinical feature fusion, comprising: performing deep learning on an ultrasonic image of a subject through a trained deep learning network, and outputting a malignant polyp risk score based on the deep learning of the ultrasonic image, wherein the ultrasonic image comprises a gallbladder gray-scale image containing a transverse position and a longitudinal position of a complete polypoid lesion region, and the deep learning network is selected from one of VGG, resNet, denseNet, googLeNet and Vision Transformer; And then inputting the malignant polyp risk score and clinical characteristics of the subject into a trained machine learning model for learning to obtain the final probability of benign and malignant gallbladder polypoid lesions, wherein the clinical characteristics of the subject comprise CA19-9, CEA and maximum polyp diameter, and the machine learning model is one selected from Logistic regression, lightGBM, random Forest and XGBoost, extraTrees, naiveBayes, SVM.
  2. 2. The method for predicting benign and malignant gallbladder polypoid lesions based on deep learning of ultrasonic images and clinical feature fusion according to claim 1, wherein the method comprises the following steps of: The deep learning network is a VGG-13BN network in the VGG network, and the machine learning model is Logistic regression.
  3. 3. The method for predicting benign and malignant gallbladder polypoid lesions based on deep learning of ultrasonic images and clinical feature fusion according to claim 1, wherein the method comprises the following steps of: the ultrasound pattern is pre-processed before being input into the trained VGG-13BN network.
  4. 4. The method for predicting benign and malignant gallbladder polypoid lesions based on deep learning of ultrasonic images and clinical feature fusion according to claim 1, wherein the method comprises the following steps of: And loading a pre-training model weight in a model training stage for the trained VGG-13BN network, modifying a full-connection layer at the end of the network to output a classification result to realize prediction of benign and malignant gallbladder polypoid lesions, adopting a cross entropy loss function as an optimization target in the training process, and updating model parameters by using a random gradient descent (SGD) optimizer, wherein the super-parameters of model training are set to be 32 as batch size, 0.01 as initial learning rate and 64 as training epoch number.
  5. 5. A device for predicting benign and malignant gallbladder polypoid lesions based on ultrasonic image deep learning and clinical feature fusion comprises a deep learning unit and a machine learning unit; The deep learning unit comprises a deep learning network, wherein the deep learning network performs deep learning on an ultrasonic image of a subject and outputs a malignant polyp risk score based on the deep learning of the ultrasonic image, the ultrasonic image comprises a gallbladder gray-scale image containing a transverse position and a longitudinal position of a complete polypoid lesion region, and the deep learning network is selected from one of VGG, resNet, denseNet, googLeNet and Vision Transformer; The machine learning unit comprises a trained machine learning model, wherein the machine learning model takes the malignant polyp risk score and clinical characteristics of a subject as input to obtain final benign and malignant probability of the gallbladder polypoid lesions, the clinical characteristics of the subject comprise CA19-9, CEA and the maximum polyp diameter, and the machine learning model is selected from one of Logistic regression, lightGBM, random Forest and XGBoost, extraTrees, naiveBayes, SVM.
  6. 6. The device for predicting malignancy of gallbladder polypoid lesions based on ultrasound image deep learning and clinical feature fusion according to claim 5, wherein the deep learning network is a VGG-13BN network in a VGG network, and the machine learning model is Logistic regression.
  7. 7. The ultrasonic image deep learning and clinical feature fusion-based gallbladder polypoid lesion benign and malignant prediction device according to claim 5, further comprising an image preprocessing unit; The image preprocessing unit preprocesses the ultrasonic image and inputs the preprocessed ultrasonic image into the trained VGG-13BN network.
  8. 8. The ultrasonic image deep learning and clinical feature fusion-based gallbladder polypoid lesion benign and malignant prediction device according to claim 5, wherein: And loading a pre-training model weight in a model training stage for the trained VGG-13BN network, modifying a full-connection layer at the end of the network to output a classification result to realize prediction of benign and malignant gallbladder polypoid lesions, adopting a cross entropy loss function as an optimization target in the training process, and updating model parameters by using a random gradient descent (SGD) optimizer, wherein the super-parameters of model training are set to be 32 as batch size, 0.01 as initial learning rate and 64 as training epoch number.
  9. 9. A computer-readable storage medium having stored thereon a computer program which, when executed, implements the method for predicting benign and malignant gallbladder polypoid lesions based on ultrasound image deep learning fused with clinical features as claimed in any one of claims 1-4.

Description

Ultrasonic image deep learning and clinical feature fusion-based gallbladder polypoid lesion benign and malignant prediction method and device and storage medium thereof Technical Field The invention relates to the technical field of medical image artificial intelligence, in particular to a method and a device for predicting malignancy of a gallbladder polypoid lesion based on two-dimensional ultrasonic deep learning and clinical feature fusion, and a storage medium thereof, belonging to the technical field of intersection of Computer-Aided Diagnosis (CAD), medical ultrasonic intelligent analysis and tumor risk prediction. Background Gallbladder polypoid lesions (Gallbladder Polypoid Lesions, GPLs) are clinically common gallbladder occupancy lesions, and ultrasonic examination is the most prominent and most common primary screening method. Although most polyps are benign, a small number of patients may be at risk of gall bladder cancer, and neglecting treatment may lead to rapid progression of the malignant condition, severely threatening the health. However, blindly following the current treatment recommendations for ultrasound-prompted gallbladder polyps >1cm in diameter, i.e., surgical cholecystectomy, can result in a large number of benign polyp patients undergoing cholecystectomy treatment, which can not only increase patient pain and potential impact on the digestive system after cholecystectomy. Accurate pre-operative risk assessment is therefore of great importance for the formulation of rational surgical strategies. The preoperative evaluation method aiming at GPL malignant risk at the present stage mainly depends on the following approaches: (1) Traditional two-dimensional ultrasound inspection As a first-line imaging examination mode, the two-dimensional ultrasound has the advantages of noninvasive, convenience and economy. However, the judgment of benign and malignant polyps mainly depends on the experience and subjective judgment of operators, and has great difference in description of indexes such as focus morphology, internal echo, substrate width and the like, and has limited diagnosis accuracy. The sensitivity of the benign and malignant distinction is only 47% -67% as reported by the prior researches, and misdiagnosis, missed diagnosis and improper operation strategies are easy to cause. (2) Prediction model based on clinical variables Some researches try to combine clinical factors (such as age, tumor markers, inflammation indexes and the like) with ultrasonic characteristics to establish a predictive scoring system, but the stability and generalization capability of a model are limited due to the dependence on manual extraction of the characteristics, and meanwhile, the ultrasonic subjective characteristics have high observer dependence, so that the clinical popularization of the model is limited. (3) Radiohistology method Although radiology can extract a large number of quantitative features from medical images, it relies on manual segmentation, preprocessing and complex feature engineering, does not have real-time properties, and its application to ultrasound images is affected by noise, artifacts and image quality variability, and its clinical transformation is still limited. (4) Deep learning artificial intelligence method Deep learning can realize end-to-end automatic extraction of high-dimensional features of images and reduce the influence of subjective factors, so that the method has wide potential in oncology. However, the existing deep learning research on malignancy of gallbladder polyps is mostly limited to single-center and small-scale samples, the model generalization is not strong, and the effective fusion of clinical variables is lacking, the explanatory analysis is lacking, and the application of the model in clinic is limited. In summary, the following problems to be solved still exist in the prior art: 1. The two-dimensional ultrasonic image has the problems of strong subjective dependence and poor diagnosis accuracy. 2. The existing model is lack of deep learning automatic feature extraction, and ultrasonic implicit information is not fully utilized. 3. Prediction methods based on clinical variables are susceptible to data characteristics and have insufficient generalization capability. 4. Most of the existing researches are single-center small sample researches, and are difficult to be suitable for multi-center clinical actual scenes. 5. There is a lack of interpretable, deployable, and versatile preoperative risk assessment tools. Therefore, an intelligent prediction method capable of automatically extracting two-dimensional ultrasonic image features and effectively fusing the two-dimensional ultrasonic image features with key clinical variables is urgently needed to realize accurate, interpretable and generalized preoperative prediction of malignant risk of gallbladder polypoid lesions. Disclosure of Invention In view of the above, there is a strong need for an int