Construction method and system of intestinal lymphoma and Crohn disease identification model

Abstract

The invention discloses an intestinal lymphoma and Crohn disease identification model construction method and system, and the method comprises the following steps: S1, selecting endoscopic images of an intestinal lymphoma patient and a Crohn disease patient from a hospital, and respectively applying the endoscopic images to an internal verification experiment and an external verification experiment; s2, preprocessing the image through an image mask algorithm and a data enhancement technology; s3, constructing and training an InceptionV3 convolutional neural network model, selecting an optimal hyper-parameter combination through internal verification, and then performing external verification to determine the classification efficiency of the model; and S4, performing statistical analysis on the diagnosis result of the clinician. The InceptionV3 model constructed by adopting the method not only has efficient diagnosis performance which is obviously superior to a primary endoscopic physician and equivalent to a high-grade endoscopic expert, but also remarkably improves the endoscopic diagnosis rate of the primary clinical physician, and provides powerful support for primary medical institutions and medical centers lacking experience.

Description

Technical Field

[0001] The present invention relates to the field of image recognition, and in particular to a method and system for constructing an intestinal lymphoma and Crohn's disease identification model. Background Art

[0002] Intestinal lymphoma (IL) is a rare digestive tract malignancy, accounting for about 0.2–20% of all intestinal malignancies. The clinical manifestations of IL mainly include abdominal pain, diarrhea, bloody stools, weight loss, fever, etc. IL lesions can affect any part of the intestine, especially the small intestine and ileocecal region. According to the source of the lesion cells, it can be divided into B-cell lymphoma and NK / T-cell lymphoma. Pathological examination is the gold standard for the diagnosis of IL. However, due to the lack of specificity of the clinical manifestations of IL, clinicians have limited knowledge of the disease and lack of vigilance. In addition, because the IL lesions originate from the submucosal layer and the lamina propria, the endoscopic biopsy is not deep enough, resulting in less and superficial biopsy specimens, which causes IL to be often misdiagnosed and missed in clinical practice.

[0003] Crohn's disease (CD) is a chronic inflammatory granulomatous disease with recurrent and persistent symptoms. The entire digestive tract can be affected, and it is common in the terminal ileum and proximal colon. The main clinical manifestations of CD include abdominal pain, diarrhea, bloody stools, weight loss, fever, etc. CD lacks a gold standard for diagnosis, and it needs to be judged based on the patient's clinical manifestations, laboratory tests, imaging examinations, endoscopic manifestations, and pathological histological examinations. The typical endoscopic manifestations of CD are segmental longitudinal ulcers, cobblestone signs, aphthous ulcers, and intestinal stenosis. IL and CD are often confused in clinical manifestations and endoscopic characteristics, so many domestic and foreign literatures have reported cases of B-cell or T-cell IL misdiagnosed as CD. Since IL is a hematological malignancy that usually progresses rapidly and has a poor prognosis, misdiagnosis of IL as CD may result in patients being unable to receive correct treatment, ultimately resulting in poor prognosis or even death. Therefore, it is of great clinical significance to effectively improve the ability to distinguish IL from CD.

[0004] Previously, retrospective studies have constructed differential diagnosis models for IL and CD based on patients' clinical symptoms, laboratory indicators, imaging features, and endoscopic manifestations. However, these studies require clinicians to accurately identify imaging or endoscopic features, which places high demands on the experience of clinicians. With the rapid development of artificial intelligence technology, deep learning has been widely used in the diagnosis and localization of gastrointestinal lesions, among which convolutional neural networks are the most commonly used artificial neural networks and have shown good performance in static endoscopic image recognition. Therefore, the present invention uses deep learning technology to develop an endoscopic differential diagnosis model for IL and CD, in order to improve the differential diagnosis ability and endoscopic diagnosis rate of intestinal lymphoma and Crohn's disease of inexperienced clinicians under endoscopy. Summary of the invention

[0005] The purpose of the present invention is to provide a method and system for constructing an intestinal lymphoma and Crohn's disease identification model to solve the problems raised in the above background technology.

[0006] To achieve the above-mentioned object of the invention, one aspect of the present invention provides a method for constructing a model for distinguishing intestinal lymphoma and Crohn's disease, comprising the following steps:

[0007] Step S1, selecting endoscopic images of intestinal lymphoma patients and Crohn's disease patients from a hospital, and using the endoscopic images for an internal validation experiment and an external validation experiment respectively;

[0008] Step S2, preprocessing the image by image mask algorithm and data enhancement technology;

[0009] Step S3, construct and train the InceptionV3 convolutional neural network model, select the best hyperparameter combination through internal verification, and then perform external verification to determine the classification efficiency of the model;

[0010] Step S4: performing statistical analysis on the clinical physician's diagnosis results.

[0011] Furthermore, the conditions that the endoscopic image in step S1 needs to meet include:

[0012] Endoscopic images show lesions, including ulcers, polyps, masses, and diffuse infiltrative lesions;

[0013] Endoscopic images were classified and annotated by a junior endoscopist;

[0014] The annotated images were reviewed and corrected by an endoscopist with more than 15 years of clinical experience and an endoscopic diagnosis rate of more than 90%.

[0015] Further, endoscopic exclusion criteria included:

[0016] Out-of-focus, blurry endoscopic images;

[0017] Endoscopic image of inadequate bowel preparation with feces obscuring lesions.

[0018] Furthermore, the image mask algorithm comprises the following steps:

[0019] Step S201, separating the red channel grayscale image of the input image;

[0020] Step S202, performing threshold binarization processing on the red channel grayscale image, and performing morphological operations to remove the text area in the image;

[0021] Step S203, finding the largest area in the image, filling holes, and removing small connected areas to obtain a mask image of the region of interest of the endoscopic image;

[0022] Step S204, performing calculations on the original image and the mask image to obtain an endoscopic image with invalid information removed;

[0023] Step S205 , detecting the boundary positions of the black pixels in the image and cropping the image to obtain an endoscopic image with the black edges and text removed.

[0024] Further, step S3 includes the following steps:

[0025] Step S301, conduct an internal validation experiment, use a five-fold cross validation method to train the model, and determine the optimal hyperparameter combination, including learning rate, batch size, and number of training rounds, through a preliminary evaluation of the model;

[0026] Step S302, using the best hyperparameter combination determined by the internal validation experiment, obtain the classification results of the InceptionV3 model on the external validation set, and evaluate the classification performance and generalization ability of the model;

[0027] Step S303, using the gradient weighted class activation mapping algorithm to visualize the InceptionV3 model to achieve model interpretability;

[0028] Step S304, plotting a receiver operating characteristic curve according to the experimental results, and using binary classification model evaluation indicators to perform statistical analysis on the model performance, wherein the binary classification model evaluation indicators include accuracy, sensitivity, specificity, precision, F1 value, and area under the receiver operating characteristic curve.

[0029] Furthermore, the performance of the InceptionV3 model under different hyperparameter combinations in the internal validation experiment was compared, and the optimal hyperparameter combination of the model was obtained as follows: the learning rate was 1×10 -5 , the batch size is 32, and the number of training rounds is 200 epochs.

[0030] Further, step S4 includes the following steps:

[0031] Step S401, the clinician classifies the endoscopic images in the external validation data set before and after the assistance of the InceptionV3 model, and obtains two rounds of results respectively. The two rounds of classification results of the clinician are compared with the actual clinical diagnosis to obtain the classification performance of the clinician on the external validation set;

[0032] Step S402, using accuracy, sensitivity, specificity, and Youden index indicators to evaluate the auxiliary effectiveness of the InceptionV3 model in clinical physician classification.

[0033] Furthermore, in step S402, for the results of independent diagnosis by InceptionV3 and clinical physicians, if the overall evaluation index conforms to the normal distribution, a single sample t-test is used to evaluate whether the model is better than the classification performance of clinical physicians, otherwise the Mann-Whitney U rank sum test is used for analysis; for the diagnosis results of clinical physicians before and after InceptionV3 assistance, if the overall difference in evaluation indicators conforms to the normal distribution, a paired sample t-test is used to measure whether AI-assisted diagnosis effectively improves the diagnosis rate of clinical physicians, otherwise the Wilcoxon rank sum test is used for evaluation.

[0034] Furthermore, in step S402, the Kappa test is applied to evaluate the diagnostic consistency between the CNN model and the clinician. The Kappa value ranges from [-1, 1], and its size indicates the degree of consistency. The specific definition is as follows:

[0035] (1) κ = -1: completely inconsistent;

[0036] (2) -1<κ<0: diagnostic agreement is less than chance agreement;

[0037] (3) κ = 0: accidental agreement;

[0038] (4) 0<κ<0.2: very low consistency;

[0039] (5) 0.21<κ<0.4: fair consistency;

[0040] (6) 0.41<κ<0.6: moderate consistency;

[0041] (7) 0.61<κ<0.8: very high consistency;

[0042] (8) 0.81<κ<1: almost perfect agreement;

[0043] (9)κ=1: Completely consistent.

[0044] Another aspect of the present invention provides a system for constructing a model for distinguishing intestinal lymphoma and Crohn's disease, comprising an image screening module, an image processing module, a model construction module, and a classification statistics module, wherein:

[0045] The image screening module is used to select endoscopic images of intestinal lymphoma patients and Crohn's disease patients from the hospital, and use the endoscopic images for internal validation experiments and external validation experiments respectively;

[0046] The image processing module pre-processes the image through image mask algorithm and data enhancement technology;

[0047] The model building module is used to build and train the InceptionV3 convolutional neural network model, select the best hyperparameter combination through internal verification, and then perform external verification to determine the classification efficiency of the model;

[0048] The classification statistics module is used to perform statistical analysis on the diagnosis results of clinicians.

[0049] Compared with the prior art, the present system and method have the following advantages:

[0050] (1) The InceptionV3 model constructed in the present invention incorporates multicenter data, and the number of patients with intestinal lymphoma is larger than the number of patients in previous related studies, which reduces the bias brought by the research data.

[0051] (2) The present invention constructed and fully evaluated the InceptionV3 diagnostic model through internal five-fold cross-validation and external independent center validation. The model not only has good differential diagnosis performance, but also has good robustness and generalization ability, and has the value and potential for clinical application promotion.

[0052] (3) This invention is the first to apply convolutional neural networks to the differential diagnosis of intestinal lymphoma and Crohn's disease. The constructed InceptionV3 model not only has efficient diagnostic performance, which is significantly better than that of junior endoscopists and comparable to that of senior endoscopists, but also significantly improves the endoscopic diagnosis rate of junior clinical physicians, providing strong support for primary medical institutions and inexperienced medical centers. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 Flowchart of the method for constructing the identification model for intestinal lymphoma and Crohn's disease.

[0054] Figure 2 Schematic diagram of the method for constructing a model for the identification of intestinal lymphoma and Crohn's disease.

[0055] Figure 3The diagram shows the division of the data sets for internal and external validation. In the diagram, Ruijin Hospital is the Ruijin Hospital affiliated to Shanghai Jiao Tong University School of Medicine; Soochow University Affiliated Hospital No. 2 is the Second Affiliated Hospital of Soochow University; Air Force General Hospital is the Specialty Medical Center of the People's Liberation Army Air Force; Anhui Medical University Affiliated Hospital No. 1 is the First Affiliated Hospital of Anhui Medical University; Changhai is the First Affiliated Hospital of Naval Medical University; Xiamen University Affiliated Hospital No. 1 is the First Affiliated Hospital of Xiamen University.

[0056] Figure 4 The heat maps generated by the Grad-CAM algorithm, where (A) is the region of interest of InceptionV3 in internal validation, which is consistent with the IL lesion location; (B) is the region of interest of InceptionV3 in internal validation, which is consistent with the CD lesion location; (C) is the region of interest of InceptionV3 in external validation, which is consistent with the IL lesion location; (D) is the region of interest of InceptionV3 in external validation, which is consistent with the CD lesion location. DETAILED DESCRIPTION

[0057] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0058] Deep learning is a type of artificial intelligence algorithm based on artificial neural networks. It learns complex high-dimensional features in input data through multi-layer cascaded neural networks, thereby performing a series of decision-making and prediction processes on the input data. Multi-layer artificial neural networks are deep neural networks. Deep neural networks include models with many different architectures, such as convolutional neural networks (CNN), recurrent neural networks, and generative adversarial networks. CNN is a deep feedforward neural network that includes convolution calculations. Its basic structure consists of convolution layers, pooling layers, and fully connected layers. Among them, the convolution layer is used for feature extraction. Through multiple convolution layers, a high-dimensional feature map of the input data can be obtained. The pooling layer is used to downsample the feature map and reduce the dimension of the feature map, thereby improving the computational efficiency of the model and facilitating the generalization of the model. The fully connected layer is a classifier that converts the feature map obtained by the previous network into a vector and obtains the network output through linear transformation. It is also called a linear layer. CNN is particularly suitable for processing image data, and is therefore widely used in the classification, detection, and segmentation of medical images such as ultrasound, X-ray, CT, MRI, endoscopy, and pathology, thereby assisting clinicians in abnormal lesion detection and disease diagnosis. In the field of digestive endoscopy, CNN can be applied to many studies such as detection of precancerous lesions of the digestive tract, diagnosis of digestive tract malignant tumors and prediction of infiltration depth, detection of small intestinal bleeding, and prediction of prognosis of inflammatory bowel disease. The present invention uses InceptionV3 in CNN to build an endoscopic diagnosis model. InceptionV3 has designed a new module that extracts image features of different scales by connecting multiple convolution kernels of different sizes (1x1, 3x3, 5x5) and the maximum pooling layer in parallel, so that InceptionV3 shows superior performance in image classification.

[0059] like Figure 1 , 2 are respectively a flow chart and a schematic diagram of the method of the present invention. The present invention provides an AI-assisted diagnosis method for intestinal lymphoma and Crohn's disease based on endoscopic images, which assists in distinguishing intestinal lymphoma and Crohn's disease by extracting feature information of white light endoscopic images. The specific steps are:

[0060] Step S1, obtaining endoscopic images, the endoscopic images are from 271 intestinal lymphoma patients and 232 Crohn's disease patients who underwent gastrointestinal endoscopy examinations in six intestinal disease centers including Ruijin Hospital Affiliated to Shanghai Jiaotong University School of Medicine, the First Affiliated Hospital of Naval Medical University, the First Affiliated Hospital of Xiamen University, the Special Medical Center of the Chinese People's Liberation Army Air Force, the Second Affiliated Hospital of Soochow University, and the First Affiliated Hospital of Anhui Medical University from January 2007 to July 2023. White light endoscopic images of patients diagnosed with intestinal lymphoma and Crohn's disease according to pathology or diagnosis and treatment guidelines are used to build an artificial intelligence diagnostic model. Figure 3As shown in Figure 1, the data sets for internal and external validation of the above endoscopic images were divided. 2945 endoscopic images from five intestinal disease centers, including Ruijin Hospital Affiliated to Shanghai Jiao Tong University School of Medicine, the First Affiliated Hospital of Naval Medical University, the Special Medical Center of the Chinese People's Liberation Army Air Force, the Second Affiliated Hospital of Soochow University, and the First Affiliated Hospital of Anhui Medical University, were used for internal validation experiments, including 1517 intestinal lymphoma images and 1428 Crohn's disease images. 443 endoscopic images from the First Affiliated Hospital of Xiamen University were used for independent external validation experiments, including 222 intestinal lymphoma images and 221 Crohn's disease images.

[0061] Inclusion criteria for endoscopic images included:

[0062] 1. Select endoscopic images with lesions, such as ulcers, polyps, masses, diffuse infiltrative lesions, etc.

[0063] 2. A junior endoscopist classifies and annotates the patient's endoscopic images;

[0064] 3. An endoscopic expert with more than 15 years of clinical experience and an endoscopic diagnosis rate of more than 90% will review the annotated images;

[0065] 4. If the annotation results are inconsistent, a face-to-face discussion will be held, and the annotation results of the endoscopist will prevail. Endoscopic images of intestinal lymphoma are annotated according to the results of endoscopic biopsy or postoperative pathology, and the pathological diagnosis refers to the 2016 WHO classification of lymphocytic tumors. Endoscopic images of Crohn's disease are annotated according to the endoscopic manifestations of Crohn's disease described in the "Guidelines for the Diagnosis and Treatment of Crohn's Disease in China (Guangzhou, 2023)".

[0066] Exclusion criteria for endoscopic images were:

[0067] Out-of-focus, blurry endoscopic images;

[0068] Endoscopic image of inadequate bowel preparation with feces obscuring lesions.

[0069] Step S2, preprocessing the endoscopic image. The endoscopic image contains invalid information such as black edges, patient information, and examination time around the endoscopic image, so all images need to be preprocessed before entering the deep learning model. The present invention designs an image mask algorithm, first separates the red channel grayscale image of the input image, then performs threshold binarization on the grayscale image, and performs morphological operations to remove the text area in the image, then finds the largest area in the image, fills the holes, and removes small connected areas to obtain a mask image of the area of ​​interest of the endoscopic image, and then operates on the original image and the mask image to obtain an endoscopic image with invalid information such as text removed, and finally detects the boundary position of the black pixels in the image and cuts it to obtain an endoscopic image with black edges and text removed. The algorithm can detect and cut black edges, and mask out the remaining invalid information, thereby improving the computational efficiency of the model. In order to make the image size meet the requirements of the model input, while considering the computational efficiency of the model and the effectiveness of the image information, the image processed by the mask algorithm is resized to 299×299 pixels before entering the model. In order to avoid model overfitting, data augmentation technology is used to expand the training data set, specifically random vertical flipping, random horizontal flipping, rotation, and color jittering. Data augmentation is a commonly used data augmentation technology in the field of deep learning. It generates new training images by flipping, rotating, cropping, adding noise, and color jittering existing training images, thereby expanding the training data set with limited sample size and increasing its diversity, effectively avoiding model overfitting and improving the robustness and generalization ability of the model.

[0070] Step S3, construct a convolutional neural network InceptionV3 model based on the endoscopic image. The InceptionV3 model is pre-trained using the ImageNet public dataset, and the pre-trained model parameters are used as initial parameters, and then the model parameters are fine-tuned during the training process based on the endoscopic image data of the present invention. The model training and validation experiments are divided into two parts, including internal validation experiments and external validation experiments. The internal validation experiment uses data from five intestinal disease centers for model training and hyperparameter selection, and the external validation experiment uses data from the remaining intestinal disease center for model performance evaluation. All deep learning models are implemented using Python 3.6.13 and Pytorch 1.10.1, and the experimental environment is NVIDIA A40 GPU.

[0071] Step S3 is further divided into the following steps:

[0072] Step S301, network training and internal validation. The internal validation experiment uses a five-fold cross-validation method to train the model, and determines the optimal hyperparameter combination through preliminary evaluation of the model, including learning rate, batch size, and number of training epochs. Figure 3As shown in the figure, 239 patients with intestinal lymphoma and 191 patients with Crohn's disease from five intestinal disease centers formed the internal validation data set and were randomly stratified into five sub-datasets. The random seed was fixed, so the ratio of intestinal lymphoma to Crohn's disease patients in each sub-dataset was relatively stable, ensuring the randomness of data allocation and the repeatability of the experiment. Data allocation was performed at the patient level, and there was no patient overlap between different sub-datasets, ensuring that endoscopic images of the same patient would not appear in both the training set and the validation set, avoiding model overfitting caused by data leakage. The loss function is a binary cross entropy loss function, which is used to calculate the training loss, and the optimization algorithm is the adaptive moment estimation (Adam) algorithm, which is used to optimize the training loss to minimize it. By comparing the performance of the InceptionV3 model in the internal validation experiment under different hyperparameter combinations, the optimal hyperparameter combination of the model can be obtained: the learning rate is 1×10 -5 , the batch size is 32, and the number of training rounds is 200 epochs.

[0073] In step S302, 32 intestinal lymphoma patients and 41 Crohn's disease patients from the remaining intestinal disease center formed an independent external validation data set. Using the optimal hyperparameter combination determined by the internal validation experiment, the prediction results of the InceptionV3 model on the external validation set can be obtained, and the classification performance and generalization ability of the model can be evaluated.

[0074] Step S303, the InceptionV3 model is visualized using the gradient weighted class activation mapping (Grad-CAM) algorithm. Grad-CAM is a commonly used model interpretability algorithm, and its basic principle is as follows: given an input image and a specific category, the feature map of the last convolutional layer is obtained through forward propagation, and then the prediction result of the specific category is back-propagated to obtain the gradient information of the feature map, the gradient information of the last convolutional layer is globally averaged and then weighted summed, and finally the gradient weighted class activation map is generated through the ReLU activation function. Grad-CAM can locate the region of interest during model classification, distinguish and display the contribution of different image regions to the model classification results, and does not require modification of the original structure of the model. Four representative endoscopic images were selected from the internal validation dataset and the external validation dataset, including two images of intestinal lymphoma and two images of Crohn's disease. Figure 4 As shown in the figure, the heat map generated by the Grad-CAM algorithm shows that the region of interest of the InceptionV3 model is consistent with the lesion area in the endoscopic image, which illustrates the reliability of the model.

[0075] Step S304, draw the receiver operating characteristic curve (ROC) according to the experimental results, and evaluate the model performance using six commonly used binary classification model evaluation indicators, including accuracy, sensitivity (recall), specificity, precision, F1 value, and area under the receiver operating characteristic curve (AUC). The F1 value is the weighted harmonic mean of recall and precision, and is a comprehensive indicator that considers both recall and precision, while the AUC is a comprehensive indicator that considers both sensitivity and specificity. Since the accuracy is greatly affected by the imbalance of sample categories, the above indicators need to be considered comprehensively when measuring the model performance. All statistical analyses were implemented using Python 3.6.13 and Python package scikit-learn 0.24.2.

[0076] Step S4: Statistical analysis of the clinician's diagnosis results. In order to further evaluate the diagnostic ability of the InceptionV3 model, the classification performance of InceptionV3 and clinicians was compared on an external validation dataset. The following steps are included:

[0077] Step S401, recruit clinical physicians. The clinical physicians participating in the comparative experiment are divided into two groups, including 2 senior endoscopists with more than 10 years of clinical experience and 4 junior endoscopists with 3 to 5 years of clinical experience. All endoscopic images in the external validation data set are anonymous and randomly arranged. When the patient's medical history, clinical information, and true diagnosis are hidden, the clinicians diagnose the endoscopic images in the external validation data set one by one to obtain the first round of results, and then re-diagnose with the assistance of the InceptionV3 model to obtain the second round of results. Clinicians do not interfere with each other and make independent diagnoses. Finally, the first and second round classification results of the clinicians are compared with the actual clinical diagnosis to obtain the classification performance of the clinicians on the external validation set.

[0078] Step S402, statistically analyze the above classification results. The four indicators of accuracy, sensitivity, specificity, and Youden Index are used to evaluate the diagnostic efficacy of clinicians. The Youden Index is an indicator that comprehensively considers sensitivity and specificity, and its calculation method is as follows: Youden Index = Sensitivity + Specificity - 100%. For the results of independent diagnosis by InceptionV3 and clinicians, if the overall evaluation index conforms to the normal distribution, a single sample t-test is used to evaluate whether the model is better than the classification performance of clinicians, otherwise the Mann-Whitney U rank sum test is used for analysis. For the diagnosis results of clinicians before and after InceptionV3 assistance, if the overall difference in the evaluation index conforms to the normal distribution, a paired sample t-test is used to measure whether AI-assisted diagnosis effectively improves the diagnosis rate of clinicians, otherwise the Wilcoxon rank sum test is used for evaluation. A P value of less than 0.05 indicates that the difference in the diagnostic results is statistically significant. The Kappa test was used to evaluate the diagnostic consistency between the CNN model and clinicians. The Kappa value (κ) ranges from -1 to 1, and its size indicates the degree of consistency: (1) κ = -1: complete inconsistency; (2) -1 < κ < 0: diagnostic consistency is less than chance consistency; (3) κ = 0: chance consistency; (4) 0 < κ < 0.2: extremely low consistency; (5) 0.21 < κ < 0.4: fair consistency; (6) 0.41 < κ < 0.6: moderate consistency; (7) 0.61 < κ < 0.8: extremely high consistency; (8) 0.81 < κ < 1: almost perfect consistency; (9) κ = 1: perfect consistency. All statistical analyses were performed using R 4.0.5.

[0079] The diagnostic performance of the InceptionV3 model constructed by the present invention in internal validation and external validation is shown in Table 1. The diagnostic performance of clinicians on the external validation set before and after AI assistance is shown in Table 2. The results of the Kappa consistency test between the InceptionV3 model and the clinician's diagnosis are shown in Table 3. In the external validation set, our InceptionV3 model has a significantly better diagnostic efficacy than junior endoscopists for endoscopic images of intestinal lymphoma and Crohn's disease, which is comparable to that of senior endoscopists, and has moderate diagnostic consistency with senior endoscopists, which is higher than the diagnostic consistency between senior endoscopists and junior endoscopists, and between junior endoscopists. With the assistance of the InceptionV3 model, the diagnostic performance of junior endoscopists for endoscopic images of intestinal lymphoma and Crohn's disease has been significantly improved.

[0080] Table 1: Diagnostic performance of InceptionV3 on internal validation and external validation sets.

[0081] CNN Model Accuracy (%) AUC Sensitivity (%) Specificity (%) Accuracy (%) F1 value (%) Internal validation set 83.1 0.910 83.4 82.7 85.0 84.2 External validation set 81.9 0.895 83.8 80.1 80.9 82.3

[0082] Table 2: Diagnostic performance of clinicians on the external validation set before and after AI assistance.

[0083] Accuracy (%) Sensitivity (%) Specificity (%) Youden Index (%) Before AI assistance Senior Expert 1 77.7 77.9 77.4 55.3 Senior Expert 2 84.4 98.2 70.6 68.8 Junior Doctor 1 64.1 97.8 30.3 28.1 Junior Doctor 2 47.9 30.2 65.6 -4.2 Junior Doctor 3 70.2 59.0 81.5 40.5 Junior Doctor 4 51.5 33.8 69.2 3.0 After AI assistance Senior Expert 1 85.6 95.5 75.6 71.1 Senior Expert 2 90.3 94.6 86.0 80.6 Junior Doctor 1 80.1 91.0 69.2 60.2 Junior Doctor 2 78.1 77.9 78.3 56.2 Junior Doctor 3 88.9 95.5 82.4 77.9 Junior Doctor 4 77.0 86.9 67.0 53.9

[0084] Table 3: Kappa consistency test results between the InceptionV3 model and clinical physicians.

[0085] InceptionV3 Expert 1 Expert 2 Primary 1 Primary 2 Primary 3 Primary 4 InceptionV3 0.422 0.48 0.196 -0. 08 0.243 0. 045 Expert 1 0.539 0.202 -0.108 0.292 -0. 009 Expert 2 0.378 -0. 019 0.316 0. 014 Primary 1 0. 002 0. 002 0.141 Primary 2 -0. 035 -0. 053 Primary 3 -0. 064 Primary 4

[0086] This paper applies convolutional neural networks to the differential diagnosis of intestinal lymphoma and Crohn's disease for the first time. The InceptionV3 model we constructed not only has efficient diagnostic performance, which is significantly better than that of junior endoscopists and comparable to that of senior endoscopists, but also significantly improves the endoscopic diagnosis rate of junior clinical physicians, providing strong support for primary medical institutions and inexperienced medical centers. In addition, our InceptionV3 model has good generalization ability and interpretability, and has value and potential for clinical application.

[0087] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for constructing a model for distinguishing intestinal lymphoma and Crohn's disease, characterized in that: The following steps are involved: Step S1, selecting endoscopic images of intestinal lymphoma patients and Crohn's disease patients from a hospital, and using the endoscopic images for an internal validation experiment and an external validation experiment respectively; Step S2, preprocessing the image by image mask algorithm and data enhancement technology; Step S3, construct and train the InceptionV3 convolutional neural network model, select the best hyperparameter combination through internal verification, and then perform external verification to determine the classification efficiency of the model; Step S4, performing statistical analysis on the clinical physician's diagnosis results.

2. The method for constructing a model for distinguishing intestinal lymphoma and Crohn's disease according to claim 1, characterized in that: The conditions that the endoscopic image needs to meet in step S1 include: Endoscopic images show lesions, including ulcers, polyps, masses, and diffuse infiltrative lesions; Endoscopic images were classified and annotated by a junior endoscopist; The annotated images were reviewed and corrected by an endoscopist with more than 15 years of clinical experience and an endoscopic diagnosis rate of more than 90%.

3. The method for constructing a model for distinguishing intestinal lymphoma and Crohn's disease according to claim 1, characterized in that: The exclusion criteria for endoscopic images were: Out-of-focus, blurry endoscopic images; Endoscopic image of inadequate bowel preparation with feces obscuring lesions.

4. The method for constructing a model for distinguishing intestinal lymphoma and Crohn's disease according to claim 1, characterized in that: The image mask algorithm comprises the following steps: Step S201, separating the red channel grayscale image of the input image; Step S202, performing threshold binarization processing on the red channel grayscale image, and performing morphological operations to remove the text area in the image; Step S203, finding the largest area in the image, filling holes, and removing small connected areas to obtain a mask image of the region of interest of the endoscopic image; Step S204, performing calculations on the original image and the mask image to obtain an endoscopic image with invalid information removed; Step S205 , detecting the boundary positions of the black pixels in the image and cropping the image to obtain an endoscopic image with the black edges and text removed.

5. The method for constructing a model for distinguishing intestinal lymphoma and Crohn's disease according to claim 1, characterized in that: Step S3 includes the following steps: Step S301, conduct an internal validation experiment, use a five-fold cross validation method to train the model, and determine the optimal hyperparameter combination, including learning rate, batch size, and number of training rounds, through a preliminary evaluation of the model; Step S302, using the best hyperparameter combination determined by the internal validation experiment, obtain the classification results of the InceptionV3 model on the external validation set, and evaluate the classification performance and generalization ability of the model; Step S303, using the gradient weighted class activation mapping algorithm to visualize the InceptionV3 model to achieve model interpretability; Step S304, draw a receiver operating characteristic curve according to the experimental results, and use binary classification model evaluation indicators to perform statistical analysis on model performance. The binary classification model evaluation indicators include accuracy, sensitivity, specificity, precision, F1 value, and area under the receiver operating characteristic curve.

6. The method for constructing a model for distinguishing intestinal lymphoma and Crohn's disease according to claim 4, characterized in that: The performance of the InceptionV3 model under different hyperparameter combinations in the internal validation experiment was compared, and the best hyperparameter combination of the model was obtained: the learning rate was 1×10 -5 , the batch size is 32, and the number of training rounds is 200 epochs.

7. The method for constructing a model for distinguishing intestinal lymphoma and Crohn's disease according to claim 1, characterized in that: Step S4 includes the following steps: Step S401, the clinician classifies the endoscopic images in the external validation data set before and after the assistance of the InceptionV3 model, and obtains two rounds of results respectively. The two rounds of classification results of the clinician are compared with the actual clinical diagnosis to obtain the classification performance of the clinician on the external validation set; Step S402, using accuracy, sensitivity, specificity, and Youden index indicators to evaluate the auxiliary effectiveness of the InceptionV3 model in clinical physician classification.

8. The method for constructing a model for distinguishing intestinal lymphoma and Crohn's disease according to claim 7, characterized in that: In step S402, for the results of independent diagnosis by InceptionV3 and clinical physicians, if the overall evaluation index conforms to the normal distribution, a single sample t-test is used to evaluate whether the model is better than the classification performance of clinical physicians, otherwise the Mann-WhitneyU rank sum test is used for analysis; for the diagnosis results of clinical physicians before and after InceptionV3 assistance, if the overall difference in evaluation indicators conforms to the normal distribution, a paired sample t-test is used to measure whether AI-assisted diagnosis effectively improves the diagnosis rate of clinical physicians, otherwise the Wilcoxon rank sum test is used for evaluation.

9. The method for constructing a model for distinguishing intestinal lymphoma and Crohn's disease according to claim 7, characterized in that: In step S402, the Kappa test is applied to evaluate the diagnostic consistency between the CNN model and the clinician. The Kappa value ranges from [-1, 1], and its size indicates the degree of consistency. The specific definition is as follows: (1) κ = -1: completely inconsistent; (2) -1<κ<0: diagnostic agreement is less than chance agreement; (3) κ = 0: accidental agreement; (4) 0<κ<0.2: very low consistency; (5) 0.21<κ<0.4: fair consistency; (6) 0.41<κ<0.6: moderate consistency; (7) 0.61<κ<0.8: very high consistency; (8) 0.81<κ<1: almost perfect agreement; (9)κ=1: Completely consistent.

10. A system for constructing a model for distinguishing intestinal lymphoma and Crohn's disease, characterized in that: It includes image screening module, image processing module, model building module and classification statistics module, among which: The image screening module is used to select endoscopic images of intestinal lymphoma patients and Crohn's disease patients from the hospital, and use the endoscopic images for internal validation experiments and external validation experiments respectively; The image processing module pre-processes the image through image mask algorithm and data enhancement technology; The model building module is used to build and train the InceptionV3 convolutional neural network model, select the best hyperparameter combination through internal verification, and then perform external verification to determine the classification efficiency of the model; The classification statistics module is used to perform statistical analysis on the diagnosis results of clinicians.

Images